cgroup: always use RCU accessors for protected accesses
authorTejun Heo <tj@kernel.org>
Fri, 21 Jun 2013 22:52:33 +0000 (15:52 -0700)
committerTejun Heo <tj@kernel.org>
Wed, 26 Jun 2013 17:48:48 +0000 (10:48 -0700)
kernel/cgroup.c still has places where a RCU pointer is set and
accessed directly without going through RCU_INIT_POINTER() or
rcu_dereference_protected().  They're all properly protected accesses
so nothing is broken but it leads to spurious sparse RCU address space
warnings.

Substitute direct accesses with RCU_INIT_POINTER() and
rcu_dereference_protected().  Note that %true is specified as the
extra condition for all derference updates.  This isn't ideal as all
it does is suppressing warning without actually policing
synchronization rules; however, most are scheduled to be removed
pretty soon along with css_id itself, so no reason to be more
elaborate.

Combined with the previous changes, this removes all RCU related
sparse warnings from cgroup.

Signed-off-by: Tejun Heo <tj@kernel.org>
Reported-by: Fengguang Wu <fengguang.wu@intel.com>
Acked-by; Li Zefan <lizefan@huawei.com>

kernel/cgroup.c

index ee9f0c1c8bff0e11b46a9dddcbb2d28ffdb896e9..4ed86773fff77a89e261e581ed62973389552e46 100644 (file)
@@ -1427,7 +1427,7 @@ static void init_cgroup_root(struct cgroupfs_root *root)
        INIT_LIST_HEAD(&root->root_list);
        root->number_of_cgroups = 1;
        cgrp->root = root;
-       cgrp->name = &root_cgroup_name;
+       RCU_INIT_POINTER(cgrp->name, &root_cgroup_name);
        init_cgroup_housekeeping(cgrp);
 }
 
@@ -2558,7 +2558,7 @@ static int cgroup_rename(struct inode *old_dir, struct dentry *old_dentry,
                return ret;
        }
 
-       old_name = cgrp->name;
+       old_name = rcu_dereference_protected(cgrp->name, true);
        rcu_assign_pointer(cgrp->name, name);
 
        kfree_rcu(old_name, rcu_head);
@@ -4177,13 +4177,15 @@ static int cgroup_populate_dir(struct cgroup *cgrp, bool base_files,
        /* This cgroup is ready now */
        for_each_root_subsys(cgrp->root, ss) {
                struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
+               struct css_id *id = rcu_dereference_protected(css->id, true);
+
                /*
                 * Update id->css pointer and make this css visible from
                 * CSS ID functions. This pointer will be dereferened
                 * from RCU-read-side without locks.
                 */
-               if (css->id)
-                       rcu_assign_pointer(css->id->css, css);
+               if (id)
+                       rcu_assign_pointer(id->css, css);
        }
 
        return 0;
@@ -4863,7 +4865,7 @@ int __init cgroup_init_early(void)
        css_set_count = 1;
        init_cgroup_root(&cgroup_dummy_root);
        cgroup_root_count = 1;
-       init_task.cgroups = &init_css_set;
+       RCU_INIT_POINTER(init_task.cgroups, &init_css_set);
 
        init_cgrp_cset_link.cset = &init_css_set;
        init_cgrp_cset_link.cgrp = cgroup_dummy_top;
@@ -5380,7 +5382,8 @@ bool css_is_ancestor(struct cgroup_subsys_state *child,
 
 void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
 {
-       struct css_id *id = css->id;
+       struct css_id *id = rcu_dereference_protected(css->id, true);
+
        /* When this is called before css_id initialization, id can be NULL */
        if (!id)
                return;
@@ -5446,8 +5449,8 @@ static int __init_or_module cgroup_init_idr(struct cgroup_subsys *ss,
                return PTR_ERR(newid);
 
        newid->stack[0] = newid->id;
-       newid->css = rootcss;
-       rootcss->id = newid;
+       RCU_INIT_POINTER(newid->css, rootcss);
+       RCU_INIT_POINTER(rootcss->id, newid);
        return 0;
 }
 
@@ -5461,7 +5464,7 @@ static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
        subsys_id = ss->subsys_id;
        parent_css = parent->subsys[subsys_id];
        child_css = child->subsys[subsys_id];
-       parent_id = parent_css->id;
+       parent_id = rcu_dereference_protected(parent_css->id, true);
        depth = parent_id->depth + 1;
 
        child_id = get_new_cssid(ss, depth);