cgroup: implement dynamic subtree controller enable/disable on the default hierarchy

cgroup is switching away from multiple hierarchies and will use one
unified default hierarchy where controllers can be dynamically enabled
and disabled per subtree.  The default hierarchy will serve as the
unified hierarchy to which all controllers are attached and a css on
the default hierarchy would need to also serve the tasks of descendant
cgroups which don't have the controller enabled - ie. the tree may be
collapsed from leaf towards root when viewed from specific
controllers.  This has been implemented through effective css in the
previous patches.

This patch finally implements dynamic subtree controller
enable/disable on the default hierarchy via a new knob -
"cgroup.subtree_control" which controls which controllers are enabled
on the child cgroups.  Let's assume a hierarchy like the following.

  root - A - B - C
               \ D

root's "cgroup.subtree_control" determines which controllers are
enabled on A.  A's on B.  B's on C and D.  This coincides with the
fact that controllers on the immediate sub-level are used to
distribute the resources of the parent.  In fact, it's natural to
assume that resource control knobs of a child belong to its parent.
Enabling a controller in "cgroup.subtree_control" declares that
distribution of the respective resources of the cgroup will be
controlled.  Note that this means that controller enable states are
shared among siblings.

The default hierarchy has an extra restriction - only cgroups which
don't contain any task may have controllers enabled in
"cgroup.subtree_control".  Combined with the other properties of the
default hierarchy, this guarantees that, from the view point of
controllers, tasks are only on the leaf cgroups.  In other words, only
leaf csses may contain tasks.  This rules out situations where child
cgroups compete against internal tasks of the parent, which is a
competition between two different types of entities without any clear
way to determine resource distribution between the two.  Different
controllers handle it differently and all the implemented behaviors
are ambiguous, ad-hoc, cumbersome and/or just wrong.  Having this
structural constraints imposed from cgroup core removes the burden
from controller implementations and enables showing one consistent
behavior across all controllers.

When a controller is enabled or disabled, css associations for the
controller in the subtrees of each child should be updated.  After
enabling, the whole subtree of a child should point to the new css of
the child.  After disabling, the whole subtree of a child should point
to the cgroup's css.  This is implemented by first updating cgroup
states such that cgroup_e_css() result points to the appropriate css
and then invoking cgroup_update_dfl_csses() which migrates all tasks
in the affected subtrees to the self cgroup on the default hierarchy.

* When read, "cgroup.subtree_control" lists all the currently enabled
  controllers on the children of the cgroup.

* White-space separated list of controller names prefixed with either
  '+' or '-' can be written to "cgroup.subtree_control".  The ones
  prefixed with '+' are enabled on the controller and '-' disabled.

* A controller can be enabled iff the parent's
  "cgroup.subtree_control" enables it and disabled iff no child's
  "cgroup.subtree_control" has it enabled.

* If a cgroup has tasks, no controller can be enabled via
  "cgroup.subtree_control".  Likewise, if "cgroup.subtree_control" has
  some controllers enabled, tasks can't be migrated into the cgroup.

* All controllers which aren't bound on other hierarchies are
  automatically associated with the root cgroup of the default
  hierarchy.  All the controllers which are bound to the default
  hierarchy are listed in the read-only file "cgroup.controllers" in
  the root directory.

* "cgroup.controllers" in all non-root cgroups is read-only file whose
  content is equal to that of "cgroup.subtree_control" of the parent.
  This indicates which controllers can be used in the cgroup's
  "cgroup.subtree_control".

This is still experimental and there are some holes, one of which is
that ->can_attach() failure during cgroup_update_dfl_csses() may leave
the cgroups in an undefined state.  The issues will be addressed by
future patches.

v2: Non-root cgroups now also have "cgroup.controllers".
Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Li Zefan <lizefan@huawei.com>

include/linux/cgroup.h

@@ -21,6 +21,7 @@
 #include <linux/percpu-refcount.h>
 #include <linux/seq_file.h>
 #include <linux/kernfs.h>
+#include <linux/wait.h>
 
 #ifdef CONFIG_CGROUPS
 
@@ -164,6 +165,7 @@ struct cgroup {
 
 	struct cgroup *parent;		/* my parent */
 	struct kernfs_node *kn;		/* cgroup kernfs entry */
+	struct kernfs_node *control_kn;	/* kn for "cgroup.subtree_control" */
 
 	/*
 	 * Monotonically increasing unique serial number which defines a
@@ -216,6 +218,9 @@ struct cgroup {
 	/* For css percpu_ref killing and RCU-protected deletion */
 	struct rcu_head rcu_head;
 	struct work_struct destroy_work;
+
+	/* used to wait for offlining of csses */
+	wait_queue_head_t offline_waitq;
 };
 
 #define MAX_CGROUP_ROOT_NAMELEN 64