Well, the per-conainer nr_running and nr_uninterruptibe
accounting is the only way to calculate loadavg and idle
time for *container*.
Look, the idle time for container (for one CPU) is the time
when no tasks within this container were ready to run on
this CPU. That's the definition implicitly used in global
idle time accounting. Current per-rq counters can't solve
this problem and neither can the update_process_times()
method.
The same with loadavg. To calculate it per-container we need
to have some nr_active_in_container() function, but it must
work faster than walking the tasks within the container.
Creating something similar would be a bit problematic.
Some stats are stored on task_struct some are pointed
from a signal_struct, some are reported via /proc files,
some via netlink sockets, some statistics can be per-task
some cannot.
On the other hand containers provide a generic way to group
the tasks and report the statistics for it. So we can keep
the stats in one place and report in a similar way.
Since this patch uses Paul's containers to define the term
of a group it uses the provided facilities for reporting the
results :)
Not reimplement, but collect it in two or three stages: per-task (if
needed), per-container and overall.
There are two ways of doing so:
1. collect it on-demand by walking tasks in container;
2. collect it on-the-fly reporting the values accumulated.
The first way is less intrusive, while the second one is
probably faster. Moreover - the first way is inapplicable
to some statistics, e.g. loadavg. The same stays true for
overall statistics - we may report nr_running by walking
tasklist each time, but it is not used. We suggest to make
it the same way in per-container accounting as well.
-
