]> git.proxmox.com Git - mirror_ubuntu-artful-kernel.git/commitdiff
cgroup: convert cgroup_has_live_children() into css_has_online_children()
authorTejun Heo <tj@kernel.org>
Fri, 16 May 2014 17:22:52 +0000 (13:22 -0400)
committerTejun Heo <tj@kernel.org>
Fri, 16 May 2014 17:22:52 +0000 (13:22 -0400)
Now that cgroup liveliness and css onliness are the same state,
convert cgroup_has_live_children() into css_has_online_children() so
that it can be used for actual csses too.  The function now uses
css_for_each_child() for iteration and is published.

Signed-off-by: Tejun Heo <tj@kernel.org>
Acked-by: Li Zefan <lizefan@huawei.com>
include/linux/cgroup.h
kernel/cgroup.c

index 51a339c99eb6e201316deaf50e02ccdb9efa8cd1..b76999954beb5ffbb835cd66dc917a051fb194e2 100644 (file)
@@ -873,6 +873,8 @@ css_next_descendant_post(struct cgroup_subsys_state *pos,
        for ((pos) = css_next_descendant_post(NULL, (css)); (pos);      \
             (pos) = css_next_descendant_post((pos), (css)))
 
+bool css_has_online_children(struct cgroup_subsys_state *css);
+
 /* A css_task_iter should be treated as an opaque object */
 struct css_task_iter {
        struct cgroup_subsys            *ss;
index 004004fd0dedf06a59ae5c36b7e352c959072678..082bb842b11a94a6163f38cf52ecbd56e9d1f881 100644 (file)
@@ -175,7 +175,6 @@ static int need_forkexit_callback __read_mostly;
 static struct cftype cgroup_base_files[];
 
 static void cgroup_put(struct cgroup *cgrp);
-static bool cgroup_has_live_children(struct cgroup *cgrp);
 static int rebind_subsystems(struct cgroup_root *dst_root,
                             unsigned int ss_mask);
 static int cgroup_destroy_locked(struct cgroup *cgrp);
@@ -1769,7 +1768,7 @@ static void cgroup_kill_sb(struct super_block *sb)
         * This prevents new mounts by disabling percpu_ref_tryget_live().
         * cgroup_mount() may wait for @root's release.
         */
-       if (cgroup_has_live_children(&root->cgrp))
+       if (css_has_online_children(&root->cgrp.self))
                cgroup_put(&root->cgrp);
        else
                percpu_ref_kill(&root->cgrp.self.refcnt);
@@ -3291,19 +3290,28 @@ css_next_descendant_post(struct cgroup_subsys_state *pos,
        return pos->parent;
 }
 
-static bool cgroup_has_live_children(struct cgroup *cgrp)
+/**
+ * css_has_online_children - does a css have online children
+ * @css: the target css
+ *
+ * Returns %true if @css has any online children; otherwise, %false.  This
+ * function can be called from any context but the caller is responsible
+ * for synchronizing against on/offlining as necessary.
+ */
+bool css_has_online_children(struct cgroup_subsys_state *css)
 {
-       struct cgroup *child;
+       struct cgroup_subsys_state *child;
+       bool ret = false;
 
        rcu_read_lock();
-       list_for_each_entry_rcu(child, &cgrp->self.children, self.sibling) {
-               if (!cgroup_is_dead(child)) {
-                       rcu_read_unlock();
-                       return true;
+       css_for_each_child(child, css) {
+               if (css->flags & CSS_ONLINE) {
+                       ret = true;
+                       break;
                }
        }
        rcu_read_unlock();
-       return false;
+       return ret;
 }
 
 /**
@@ -4535,7 +4543,7 @@ static int cgroup_destroy_locked(struct cgroup *cgrp)
         * ->self.children as dead children linger on it while being
         * drained; otherwise, "rmdir parent/child parent" may fail.
         */
-       if (cgroup_has_live_children(cgrp))
+       if (css_has_online_children(&cgrp->self))
                return -EBUSY;
 
        /*
@@ -5014,8 +5022,8 @@ void cgroup_exit(struct task_struct *tsk)
 
 static void check_for_release(struct cgroup *cgrp)
 {
-       if (cgroup_is_releasable(cgrp) &&
-           list_empty(&cgrp->cset_links) && !cgroup_has_live_children(cgrp)) {
+       if (cgroup_is_releasable(cgrp) && list_empty(&cgrp->cset_links) &&
+           !css_has_online_children(&cgrp->self)) {
                /*
                 * Control Group is currently removeable. If it's not
                 * already queued for a userspace notification, queue