On Thu, 31 May 2012, Kirill A. Shutemov wrote:
> > Why? Because the information exported by /proc/meminfo is considered by
> > applications to be static whereas the limit of a memcg may change without
> > any knowledge of the application.
>
> Memory hotplug does the same, right?
>
Memory hotplug is a seperate topic, it changes the amount of physical
memory that is available to the kernel, not any limitation of memory
available to a set of tasks. For memory hot-add, this does not
automatically increase the memory.limit_in_bytes of any non-root memcg,
the memory usage is still constrained as it was before the hotplug event.
Thus, applications would want to depend on memory.{limit,usage}_in_bytes
specifically to determine the amount of available memory even with
CONFIG_MEMORY_HOTPLUG.
Also, under certain cirucmstances such as when a thread is oom killed, it
may allocate memory in excess of its memcg limitation and this wouldn't be
visible as available with this patch via /proc/meminfo. Cpusets allows
softwall allocations even when a thread is simply exiting on all nodes
(and for GFP_ATOMIC allocations) and this also wouldn't be visible in
/proc/meminfo.
_______________________________________________
Containers mailing list
Containers@xxxxxxxxxxxxxxxxxxxxxxxxxx
https://lists.linuxfoundation.org/mailman/listinfo/containers
