Hello,
cgroup currently disassociates a task from its cgroups on exit and
reassigns it to the root cgroup. This behavior turns out to be
problematic for several reasons.
* Resources can't be tracked for zombies. This breaks pids controller
as zombies escape resource restriction. A cgroup can easily go way
above its limits by creating a bunch of zombies.
* It's difficult to tell where zombies came from. /proc/PID/cgroup
gets reset to / on exit so given a zombie it's difficult to tell
from which cgroup the zombie came from.
* It creates an extra work for controllers for no reason. cpu and
perf_events controllers implement exit callbacks to switch the
exiting task's membership to root when just leaving it as-is is
enough.
Unfortunately, fixing this involves opening a few cans of worms.
* Decoupling tasks being on a css_set from its reference counting so
that css_set can be pinned w/o tasks being on it and decoupling
css_set existence from whether a cgroup is populated so that pinning
a css_set doesn't confuse populated state tracking and populated
state can be used to decide whether certain operations are allowed.
* Making css task iteration drop css_set_rwsem between iteration steps
so that internal locking is not exposed to iterator users and
css_set_rwsem can be converted to a spinlock which can be grabbed
from task free path.
After this patchset, besides pids controller being fixed, the visible
behavior isn't changed on traditional hierarchies but on the default
hierarchy a zombie reports its cgroup at the time of exit in
/proc/PID/cgroup. If the cgroup gets removed before the task is
reaped, " (deleted)" is appended to the reported path.
This patchset contains the following 14 patches.
0001-cgroup-remove-an-unused-parameter-from-cgroup_task_m.patch
0002-cgroup-make-cgroup-nr_populated-count-the-number-of-.patch
0003-cgroup-replace-cgroup_has_tasks-with-cgroup_is_popul.patch
0004-cgroup-move-check_for_release-invocation.patch
0005-cgroup-relocate-cgroup_-try-get-put.patch
0006-cgroup-make-css_sets-pin-the-associated-cgroups.patch
0007-cgroup-make-cgroup_destroy_locked-test-cgroup_is_pop.patch
0008-cgroup-keep-css_set-and-task-lists-in-chronological-.patch
0009-cgroup-factor-out-css_set_move_task.patch
0010-cgroup-reorganize-css_task_iter-functions.patch
0011-cgroup-don-t-hold-css_set_rwsem-across-css-task-iter.patch
0012-cgroup-make-css_set_rwsem-a-spinlock-and-rename-it-t.patch
0013-cgroup-keep-zombies-associated-with-their-original-c.patch
0014-cgroup-add-cgroup_subsys-free-method-and-use-it-to-f.patch
0001-0007 decouple populated state tracking from css_set existence and
allows css_sets to be pinned without tasks on them.
0008-0012 update css_set task iterator to not hold lock across
iteration steps and replace css_set_rwsem with a spinlock.
0013 makes zombies keep their cgroup associations. 0014 introduces
->exit() method and fixes pids controller.
The patchset is pretty lightly tested and I need to verify that the
corner cases behave as expected.
This patchset is on top of cgroup/for-4.4 a3e72739b7a7 ("cgroup: fix
too early usage of static_branch_disable()") and available in the
following git branch.
git://git.kernel.org/pub/scm/linux/kernel/git/tj/cgroup.git review-zombies
diffstat follows. Thanks.
Documentation/cgroups/cgroups.txt | 4
Documentation/cgroups/unified-hierarchy.txt | 4
include/linux/cgroup-defs.h | 16
include/linux/cgroup.h | 14
kernel/cgroup.c | 522 +++++++++++++++++-----------
kernel/cgroup_pids.c | 8
kernel/cpuset.c | 2
kernel/events/core.c | 16
kernel/fork.c | 1
kernel/sched/core.c | 16
mm/memcontrol.c | 2
11 files changed, 354 insertions(+), 251 deletions(-)
--
tejun
--
To unsubscribe from this list: send the line "unsubscribe cgroups" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
