Re: [RFC][PATCH 01/20] pid: Intoduce the concept of a wid (wait id)

[ebiederm@xmission.com (Eric W. Biederman)]

Jeff Dike <jdike@addtoit.com> writes:

> First of all, for an RFC, this is very thorough.

Thank you.

> Second, I've been thinking along these lines for UML.  The motivation
> is to get UML out of the system call tracing business as much as
> possible, and to do so by having the host set up such that it can run
> system calls itself and they do the same thing as the UML system call
> would.
>
> For example, for a UML process chrooted into a UML filesystem, the
> file operations on normal files will do the same thing as they would
> in UML, so they could be left to run on the host.
>
> Similarly, something like virtualized processes could be made to do
> the same thing with the process operations.  Trivially, getpid() will
> return the right value if left to run on the host, so UML wouldn't
> need to intercept it.  If there is a process tree inside a container
> that mirrors the UML process tree, then lots of other system calls
> also work, and don't need to be intercepted.
>
> Ideally, I'd like namespaces on the host for all the resources under
> UML control, and for a container to group those namespaces.  However,
> something which stops short of that is still usable - UML just gets
> less benefit from it.

Having all of the namespaces is certainly on my TODO list.

I'm not at all certain if there is a need for a kernel container
concept.

> As far as processes go, ideally I'd like a containerized process to be
> an empty shell which can be completely filled from userspace.  The
> motivation for this is that when you have a UP UML with 100 processes,
> it's wasteful to have 100 virtualized processes on the host.  What I
> would want is one virtualized process which can be completely refilled
> with new attributes on a context switch.
>
> What I want to do is related to process migration, where you want to
> move a process but have it not be able to tell.  I'm describing
> migrating a process from the UML to the host such that the host
> performs as many system calls itself, but those which can't get
> intercepted and executed within the UML.  For migration between
> physical machines, this would be the same as redirecting a system call
> from the new host back to its original home.  You want to do that as
> infrequently as possible, so you want the container to provide as much
> context from the home host as possible.

Currently redirecting a system call from the new host back to it's
original home is not something I had planned on.  Most of the reasons
I want to migrate relate to avoiding the hardware I am migrating from.
Either to reduce it's load or to leave before the hardware dies.

That said the idea of a user space monitor that can handle
the strange virtualization things that don't fit well into the
kernel is appealing.

Note all of the migration I am looking is not process migration but
container migration.  So I want a container per application.

Eric