On Thu, May 31, 2012 at 03:42:38AM -0400, KOSAKI Motohiro wrote:

(5/31/12 3:35 AM), David Rientjes wrote:

quoted

On Thu, 31 May 2012, KOSAKI Motohiro wrote:

quoted

quoted

As I said, LXC and namespace isolation is a tangent to the discussion of
faking the /proc/meminfo for the memcg context of a thread.

Because of, /proc/meminfo affect a lot of libraries behavior. So, it's not only
application issue. If you can't rewrite _all_ of userland assets, fake meminfo
can't be escaped. Again see alternative container implementation.

It's a tangent because it isn't a complete psuedo /proc/meminfo for all
threads attached to a memcg regardless of any namespace isolation; the LXC
solution has existed for a couple of years by its procfs patchset that
overlaps procfs with fuse and can suppress or modify any output in the
context of a memory controller using things like
memory.{limit,usage}_in_bytes.  I'm sure all other fields could be
modified if outputted in some structured way via memcg; it looks like
memory.stat would need to be extended to provide that.  If that's mounted
prior to executing the application, then your isolation is achieved and
all libraries should see the new output that you've defined in LXC.

However, this seems like a seperate topic than the patch at hand which
does this directly to /proc/meminfo based on a thread's memcg context,
that's the part that I'm nacking.

Then, I NAKed current patch too. Yeah, current one is ugly. It assume _all_
user need namespace isolation and it clearly is not.

Actually, it only chooses the memcg version for tasks that are not in
the init pid namespace.  Tying this to the pid namespace is a bit
ugly, but would probably end up doing the right thing most of the
time.  A separate namespace would be better.