[Devel] Re: [RFC][PATCH 0/4][user-cr]: First try at integrating LXC and USER-CR

[Sukadev Bhattiprolu]

Oren Laadan [orenl at cs.columbia.edu] wrote:
> Suka,
>
> This is good stuff: convert restart to a library and intergate
> with LXC is excellent target.
>
> We need to give a thought to the API design and the sort of
> functionality that the library will offer. I'd prefer to get
> a better view of the whole plan before committing to any
> incremental step.
>
> For example:
>
> * what prefix will the library calls have ?  e.g. cr_checkpoint(),
> cr_restart() etc.  Could also be checkpoint_... or ckpt_... In
> any case, must be consistent and unique, and also agree with the
> library name: libcheckpoint.a and usercr.h doesn't make sense to
> me.

Agree. I have been worried about publishing an interface and getting
stuck with a bad interface. But I have also been trying to prototype
interface with LIBLXC available so did not yet spend lot of time
on the interface.

I have been meaning to add a disclaimer to usercr.h that the structures
will change without regard to backward compatibility

libcheckpoint.a looked like a reasonable name but for the header,
we already have a checkpoint.h from the kernel :-) Open to new/better
names/prefixes for the library calls.

Anyway, these names are just for the prototype and to start a discussion,
My guess is that we are not ready to publish a formal interface at
this point, but we can start moving towards that.

My current pass has been to eliminate all global references in
app_restart() and app_checkpoint() (like perror(), stderr, stdout etc).
while preserving the current behavior. Once we have an interface
that LIBLXC can use, maybe we could focus more on stabilizing
the interface ? It seemed to hard to do that with all the
globals in the code.

>
> * what arguments are really necessary in struct restart_args ? for
> example, a ->logfd is necessary, but ->logfile (pathname) is not
> (the caller should already open the file)

I have several small patches to clean this up. Was hoping to post
later today. They remove ->logfile, ->input, ->output fromt he
structures and just leave the relevant fds.

>
> * the library should provide an api to initialize the default args
> (e.g. the logfd should be -1 by default), e.g. cr_init_restart_args()
>
> * similarly, only constants and macros relevant to a caller
> should be exposed, not internal data structures or macros.
>
> * verbosity, debugging, warn and fail: currently that can be
> configurable to some extent; but we should never impose on the
> caller a perror() - instead the caller should pass FILE * for
> stdout and stderr (if NULL - the library will keep silent).

Yes, my current patchset uses a restart_args->ulogfd for the
user-space error messages. The restart-main.c sets this field
to fileno(stderr).

>
> * probably makes sense to add interface to freeze a process tree
> identified by its root pid, and thaw a process tree (or a cgroup).

True. For the near-term, LIBLXC has a way to freeze/thaw the containers

>
> * does the mnt-ns manipulation belong to cr_restart() ?  perhaps
> make it a separate api in the library ?

Good point. We need to think about that.

>
> * we need to change the way restart detects errors and cleans up
> from relying on signals to, e.g. on pipes. I've been wanting to
> do this for the longest time, and it will eliminate most (if not
> all) of the global variables.

Ok.
_______________________________________________
Containers mailing list
Containers at lists.linux-foundation.org
https://lists.linux-foundation.org/mailman/listinfo/containers