Convert obvious places to deactivate_locked_super()
[linux-2.6.git] / kernel / cgroup.c
index a391ab3..a7267bf 100644 (file)
@@ -84,7 +84,7 @@ struct cgroupfs_root {
        /* Tracks how many cgroups are currently defined in hierarchy.*/
        int number_of_cgroups;
 
-       /* A list running through the mounted hierarchies */
+       /* A list running through the active hierarchies */
        struct list_head root_list;
 
        /* Hierarchy-specific flags */
@@ -94,7 +94,6 @@ struct cgroupfs_root {
        char release_agent_path[PATH_MAX];
 };
 
-
 /*
  * The "rootnode" hierarchy is the "dummy hierarchy", reserved for the
  * subsystems that are otherwise unattached - it never has more than a
@@ -102,6 +101,39 @@ struct cgroupfs_root {
  */
 static struct cgroupfs_root rootnode;
 
+/*
+ * CSS ID -- ID per subsys's Cgroup Subsys State(CSS). used only when
+ * cgroup_subsys->use_id != 0.
+ */
+#define CSS_ID_MAX     (65535)
+struct css_id {
+       /*
+        * The css to which this ID points. This pointer is set to valid value
+        * after cgroup is populated. If cgroup is removed, this will be NULL.
+        * This pointer is expected to be RCU-safe because destroy()
+        * is called after synchronize_rcu(). But for safe use, css_is_removed()
+        * css_tryget() should be used for avoiding race.
+        */
+       struct cgroup_subsys_state *css;
+       /*
+        * ID of this css.
+        */
+       unsigned short id;
+       /*
+        * Depth in hierarchy which this ID belongs to.
+        */
+       unsigned short depth;
+       /*
+        * ID is freed by RCU. (and lookup routine is RCU safe.)
+        */
+       struct rcu_head rcu_head;
+       /*
+        * Hierarchy of CSS ID belongs to.
+        */
+       unsigned short stack[0]; /* Array of Length (depth+1) */
+};
+
+
 /* The list of hierarchy roots */
 
 static LIST_HEAD(roots);
@@ -148,8 +180,8 @@ static int notify_on_release(const struct cgroup *cgrp)
 #define for_each_subsys(_root, _ss) \
 list_for_each_entry(_ss, &_root->subsys_list, sibling)
 
-/* for_each_root() allows you to iterate across the active hierarchies */
-#define for_each_root(_root) \
+/* for_each_active_root() allows you to iterate across the active hierarchies */
+#define for_each_active_root(_root) \
 list_for_each_entry(_root, &roots, root_list)
 
 /* the list of cgroups eligible for automatic release. Protected by
@@ -185,6 +217,8 @@ struct cg_cgroup_link {
 static struct css_set init_css_set;
 static struct cg_cgroup_link init_css_set_link;
 
+static int cgroup_subsys_init_idr(struct cgroup_subsys *ss);
+
 /* css_set_lock protects the list of css_set objects, and the
  * chain of tasks off each css_set.  Nests outside task->alloc_lock
  * due to cgroup_iter_start() */
@@ -271,7 +305,7 @@ static void __put_css_set(struct css_set *cg, int taskexit)
 
        rcu_read_lock();
        for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
-               struct cgroup *cgrp = cg->subsys[i]->cgroup;
+               struct cgroup *cgrp = rcu_dereference(cg->subsys[i]->cgroup);
                if (atomic_dec_and_test(&cgrp->count) &&
                    notify_on_release(cgrp)) {
                        if (taskexit)
@@ -384,6 +418,25 @@ static int allocate_cg_links(int count, struct list_head *tmp)
        return 0;
 }
 
+/**
+ * link_css_set - a helper function to link a css_set to a cgroup
+ * @tmp_cg_links: cg_cgroup_link objects allocated by allocate_cg_links()
+ * @cg: the css_set to be linked
+ * @cgrp: the destination cgroup
+ */
+static void link_css_set(struct list_head *tmp_cg_links,
+                        struct css_set *cg, struct cgroup *cgrp)
+{
+       struct cg_cgroup_link *link;
+
+       BUG_ON(list_empty(tmp_cg_links));
+       link = list_first_entry(tmp_cg_links, struct cg_cgroup_link,
+                               cgrp_link_list);
+       link->cg = cg;
+       list_move(&link->cgrp_link_list, &cgrp->css_sets);
+       list_add(&link->cg_link_list, &cg->cg_links);
+}
+
 /*
  * find_css_set() takes an existing cgroup group and a
  * cgroup object, and returns a css_set object that's
@@ -399,7 +452,6 @@ static struct css_set *find_css_set(
        int i;
 
        struct list_head tmp_cg_links;
-       struct cg_cgroup_link *link;
 
        struct hlist_head *hhead;
 
@@ -444,26 +496,11 @@ static struct css_set *find_css_set(
                 * only do it for the first subsystem in each
                 * hierarchy
                 */
-               if (ss->root->subsys_list.next == &ss->sibling) {
-                       BUG_ON(list_empty(&tmp_cg_links));
-                       link = list_entry(tmp_cg_links.next,
-                                         struct cg_cgroup_link,
-                                         cgrp_link_list);
-                       list_del(&link->cgrp_link_list);
-                       list_add(&link->cgrp_link_list, &cgrp->css_sets);
-                       link->cg = res;
-                       list_add(&link->cg_link_list, &res->cg_links);
-               }
-       }
-       if (list_empty(&rootnode.subsys_list)) {
-               link = list_entry(tmp_cg_links.next,
-                                 struct cg_cgroup_link,
-                                 cgrp_link_list);
-               list_del(&link->cgrp_link_list);
-               list_add(&link->cgrp_link_list, &dummytop->css_sets);
-               link->cg = res;
-               list_add(&link->cg_link_list, &res->cg_links);
+               if (ss->root->subsys_list.next == &ss->sibling)
+                       link_css_set(&tmp_cg_links, res, cgrp);
        }
+       if (list_empty(&rootnode.subsys_list))
+               link_css_set(&tmp_cg_links, res, dummytop);
 
        BUG_ON(!list_empty(&tmp_cg_links));
 
@@ -564,6 +601,9 @@ static struct backing_dev_info cgroup_backing_dev_info = {
        .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
 };
 
+static int alloc_css_id(struct cgroup_subsys *ss,
+                       struct cgroup *parent, struct cgroup *child);
+
 static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
 {
        struct inode *inode = new_inode(sb);
@@ -582,13 +622,25 @@ static struct inode *cgroup_new_inode(mode_t mode, struct super_block *sb)
  * Call subsys's pre_destroy handler.
  * This is called before css refcnt check.
  */
-static void cgroup_call_pre_destroy(struct cgroup *cgrp)
+static int cgroup_call_pre_destroy(struct cgroup *cgrp)
 {
        struct cgroup_subsys *ss;
+       int ret = 0;
+
        for_each_subsys(cgrp->root, ss)
-               if (ss->pre_destroy)
-                       ss->pre_destroy(ss, cgrp);
-       return;
+               if (ss->pre_destroy) {
+                       ret = ss->pre_destroy(ss, cgrp);
+                       if (ret)
+                               break;
+               }
+       return ret;
+}
+
+static void free_cgroup_rcu(struct rcu_head *obj)
+{
+       struct cgroup *cgrp = container_of(obj, struct cgroup, rcu_head);
+
+       kfree(cgrp);
 }
 
 static void cgroup_diput(struct dentry *dentry, struct inode *inode)
@@ -616,11 +668,13 @@ static void cgroup_diput(struct dentry *dentry, struct inode *inode)
                cgrp->root->number_of_cgroups--;
                mutex_unlock(&cgroup_mutex);
 
-               /* Drop the active superblock reference that we took when we
-                * created the cgroup */
+               /*
+                * Drop the active superblock reference that we took when we
+                * created the cgroup
+                */
                deactivate_super(cgrp->root->sb);
 
-               kfree(cgrp);
+               call_rcu(&cgrp->rcu_head, free_cgroup_rcu);
        }
        iput(inode);
 }
@@ -673,6 +727,22 @@ static void cgroup_d_remove_dir(struct dentry *dentry)
        remove_dir(dentry);
 }
 
+/*
+ * A queue for waiters to do rmdir() cgroup. A tasks will sleep when
+ * cgroup->count == 0 && list_empty(&cgroup->children) && subsys has some
+ * reference to css->refcnt. In general, this refcnt is expected to goes down
+ * to zero, soon.
+ *
+ * CGRP_WAIT_ON_RMDIR flag is modified under cgroup's inode->i_mutex;
+ */
+DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
+
+static void cgroup_wakeup_rmdir_waiters(const struct cgroup *cgrp)
+{
+       if (unlikely(test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
+               wake_up_all(&cgroup_rmdir_waitq);
+}
+
 static int rebind_subsystems(struct cgroupfs_root *root,
                              unsigned long final_bits)
 {
@@ -710,23 +780,26 @@ static int rebind_subsystems(struct cgroupfs_root *root,
                        BUG_ON(cgrp->subsys[i]);
                        BUG_ON(!dummytop->subsys[i]);
                        BUG_ON(dummytop->subsys[i]->cgroup != dummytop);
+                       mutex_lock(&ss->hierarchy_mutex);
                        cgrp->subsys[i] = dummytop->subsys[i];
                        cgrp->subsys[i]->cgroup = cgrp;
-                       list_add(&ss->sibling, &root->subsys_list);
-                       rcu_assign_pointer(ss->root, root);
+                       list_move(&ss->sibling, &root->subsys_list);
+                       ss->root = root;
                        if (ss->bind)
                                ss->bind(ss, cgrp);
-
+                       mutex_unlock(&ss->hierarchy_mutex);
                } else if (bit & removed_bits) {
                        /* We're removing this subsystem */
                        BUG_ON(cgrp->subsys[i] != dummytop->subsys[i]);
                        BUG_ON(cgrp->subsys[i]->cgroup != cgrp);
+                       mutex_lock(&ss->hierarchy_mutex);
                        if (ss->bind)
                                ss->bind(ss, dummytop);
                        dummytop->subsys[i]->cgroup = dummytop;
                        cgrp->subsys[i] = NULL;
-                       rcu_assign_pointer(subsys[i]->root, &rootnode);
-                       list_del(&ss->sibling);
+                       subsys[i]->root = &rootnode;
+                       list_move(&ss->sibling, &rootnode.subsys_list);
+                       mutex_unlock(&ss->hierarchy_mutex);
                } else if (bit & final_bits) {
                        /* Subsystem state should already exist */
                        BUG_ON(!cgrp->subsys[i]);
@@ -842,16 +915,16 @@ static int cgroup_remount(struct super_block *sb, int *flags, char *data)
        }
 
        ret = rebind_subsystems(root, opts.subsys_bits);
+       if (ret)
+               goto out_unlock;
 
        /* (re)populate subsystem files */
-       if (!ret)
-               cgroup_populate_dir(cgrp);
+       cgroup_populate_dir(cgrp);
 
        if (opts.release_agent)
                strcpy(root->release_agent_path, opts.release_agent);
  out_unlock:
-       if (opts.release_agent)
-               kfree(opts.release_agent);
+       kfree(opts.release_agent);
        mutex_unlock(&cgroup_mutex);
        mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
        return ret;
@@ -954,15 +1027,13 @@ static int cgroup_get_sb(struct file_system_type *fs_type,
        /* First find the desired set of subsystems */
        ret = parse_cgroupfs_options(data, &opts);
        if (ret) {
-               if (opts.release_agent)
-                       kfree(opts.release_agent);
+               kfree(opts.release_agent);
                return ret;
        }
 
        root = kzalloc(sizeof(*root), GFP_KERNEL);
        if (!root) {
-               if (opts.release_agent)
-                       kfree(opts.release_agent);
+               kfree(opts.release_agent);
                return -ENOMEM;
        }
 
@@ -988,7 +1059,7 @@ static int cgroup_get_sb(struct file_system_type *fs_type,
                root = NULL;
        } else {
                /* New superblock */
-               struct cgroup *cgrp = &root->top_cgroup;
+               struct cgroup *root_cgrp = &root->top_cgroup;
                struct inode *inode;
                int i;
 
@@ -1029,7 +1100,7 @@ static int cgroup_get_sb(struct file_system_type *fs_type,
                list_add(&root->root_list, &roots);
                root_count++;
 
-               sb->s_root->d_fsdata = &root->top_cgroup;
+               sb->s_root->d_fsdata = root_cgrp;
                root->top_cgroup.dentry = sb->s_root;
 
                /* Link the top cgroup in this hierarchy into all
@@ -1040,40 +1111,29 @@ static int cgroup_get_sb(struct file_system_type *fs_type,
                        struct hlist_node *node;
                        struct css_set *cg;
 
-                       hlist_for_each_entry(cg, node, hhead, hlist) {
-                               struct cg_cgroup_link *link;
-
-                               BUG_ON(list_empty(&tmp_cg_links));
-                               link = list_entry(tmp_cg_links.next,
-                                                 struct cg_cgroup_link,
-                                                 cgrp_link_list);
-                               list_del(&link->cgrp_link_list);
-                               link->cg = cg;
-                               list_add(&link->cgrp_link_list,
-                                        &root->top_cgroup.css_sets);
-                               list_add(&link->cg_link_list, &cg->cg_links);
-                       }
+                       hlist_for_each_entry(cg, node, hhead, hlist)
+                               link_css_set(&tmp_cg_links, cg, root_cgrp);
                }
                write_unlock(&css_set_lock);
 
                free_cg_links(&tmp_cg_links);
 
-               BUG_ON(!list_empty(&cgrp->sibling));
-               BUG_ON(!list_empty(&cgrp->children));
+               BUG_ON(!list_empty(&root_cgrp->sibling));
+               BUG_ON(!list_empty(&root_cgrp->children));
                BUG_ON(root->number_of_cgroups != 1);
 
-               cgroup_populate_dir(cgrp);
+               cgroup_populate_dir(root_cgrp);
                mutex_unlock(&inode->i_mutex);
                mutex_unlock(&cgroup_mutex);
        }
 
-       return simple_set_mnt(mnt, sb);
+       simple_set_mnt(mnt, sb);
+       return 0;
 
  free_cg_links:
        free_cg_links(&tmp_cg_links);
  drop_new_super:
-       up_write(&sb->s_umount);
-       deactivate_super(sb);
+       deactivate_locked_super(sb);
        return ret;
 }
 
@@ -1115,10 +1175,11 @@ static void cgroup_kill_sb(struct super_block *sb) {
                list_del(&root->root_list);
                root_count--;
        }
+
        mutex_unlock(&cgroup_mutex);
 
-       kfree(root);
        kill_litter_super(sb);
+       kfree(root);
 }
 
 static struct file_system_type cgroup_fs_type = {
@@ -1143,14 +1204,16 @@ static inline struct cftype *__d_cft(struct dentry *dentry)
  * @buf: the buffer to write the path into
  * @buflen: the length of the buffer
  *
- * Called with cgroup_mutex held. Writes path of cgroup into buf.
- * Returns 0 on success, -errno on error.
+ * Called with cgroup_mutex held or else with an RCU-protected cgroup
+ * reference.  Writes path of cgroup into buf.  Returns 0 on success,
+ * -errno on error.
  */
 int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
 {
        char *start;
+       struct dentry *dentry = rcu_dereference(cgrp->dentry);
 
-       if (cgrp == dummytop) {
+       if (!dentry || cgrp == dummytop) {
                /*
                 * Inactive subsystems have no dentry for their root
                 * cgroup
@@ -1163,13 +1226,14 @@ int cgroup_path(const struct cgroup *cgrp, char *buf, int buflen)
 
        *--start = '\0';
        for (;;) {
-               int len = cgrp->dentry->d_name.len;
+               int len = dentry->d_name.len;
                if ((start -= len) < buf)
                        return -ENAMETOOLONG;
                memcpy(start, cgrp->dentry->d_name.name, len);
                cgrp = cgrp->parent;
                if (!cgrp)
                        break;
+               dentry = rcu_dereference(cgrp->dentry);
                if (!cgrp->parent)
                        continue;
                if (--start < buf)
@@ -1214,7 +1278,7 @@ int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
        int retval = 0;
        struct cgroup_subsys *ss;
        struct cgroup *oldcgrp;
-       struct css_set *cg = tsk->cgroups;
+       struct css_set *cg;
        struct css_set *newcg;
        struct cgroupfs_root *root = cgrp->root;
        int subsys_id;
@@ -1234,11 +1298,16 @@ int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
                }
        }
 
+       task_lock(tsk);
+       cg = tsk->cgroups;
+       get_css_set(cg);
+       task_unlock(tsk);
        /*
         * Locate or allocate a new css_set for this task,
         * based on its final set of cgroups
         */
        newcg = find_css_set(cg, cgrp);
+       put_css_set(cg);
        if (!newcg)
                return -ENOMEM;
 
@@ -1266,6 +1335,12 @@ int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
        set_bit(CGRP_RELEASABLE, &oldcgrp->flags);
        synchronize_rcu();
        put_css_set(cg);
+
+       /*
+        * wake up rmdir() waiter. the rmdir should fail since the cgroup
+        * is no longer empty.
+        */
+       cgroup_wakeup_rmdir_waiters(cgrp);
        return 0;
 }
 
@@ -1611,10 +1686,10 @@ static struct inode_operations cgroup_dir_inode_operations = {
        .rename = cgroup_rename,
 };
 
-static int cgroup_create_file(struct dentry *dentry, int mode,
+static int cgroup_create_file(struct dentry *dentry, mode_t mode,
                                struct super_block *sb)
 {
-       static struct dentry_operations cgroup_dops = {
+       static const struct dentry_operations cgroup_dops = {
                .d_iput = cgroup_diput,
        };
 
@@ -1657,7 +1732,7 @@ static int cgroup_create_file(struct dentry *dentry, int mode,
  * @mode: mode to set on new directory.
  */
 static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
-                               int mode)
+                               mode_t mode)
 {
        struct dentry *parent;
        int error = 0;
@@ -1667,7 +1742,7 @@ static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
        if (!error) {
                dentry->d_fsdata = cgrp;
                inc_nlink(parent->d_inode);
-               cgrp->dentry = dentry;
+               rcu_assign_pointer(cgrp->dentry, dentry);
                dget(dentry);
        }
        dput(dentry);
@@ -1675,6 +1750,33 @@ static int cgroup_create_dir(struct cgroup *cgrp, struct dentry *dentry,
        return error;
 }
 
+/**
+ * cgroup_file_mode - deduce file mode of a control file
+ * @cft: the control file in question
+ *
+ * returns cft->mode if ->mode is not 0
+ * returns S_IRUGO|S_IWUSR if it has both a read and a write handler
+ * returns S_IRUGO if it has only a read handler
+ * returns S_IWUSR if it has only a write hander
+ */
+static mode_t cgroup_file_mode(const struct cftype *cft)
+{
+       mode_t mode = 0;
+
+       if (cft->mode)
+               return cft->mode;
+
+       if (cft->read || cft->read_u64 || cft->read_s64 ||
+           cft->read_map || cft->read_seq_string)
+               mode |= S_IRUGO;
+
+       if (cft->write || cft->write_u64 || cft->write_s64 ||
+           cft->write_string || cft->trigger)
+               mode |= S_IWUSR;
+
+       return mode;
+}
+
 int cgroup_add_file(struct cgroup *cgrp,
                       struct cgroup_subsys *subsys,
                       const struct cftype *cft)
@@ -1682,6 +1784,7 @@ int cgroup_add_file(struct cgroup *cgrp,
        struct dentry *dir = cgrp->dentry;
        struct dentry *dentry;
        int error;
+       mode_t mode;
 
        char name[MAX_CGROUP_TYPE_NAMELEN + MAX_CFTYPE_NAME + 2] = { 0 };
        if (subsys && !test_bit(ROOT_NOPREFIX, &cgrp->root->flags)) {
@@ -1692,7 +1795,8 @@ int cgroup_add_file(struct cgroup *cgrp,
        BUG_ON(!mutex_is_locked(&dir->d_inode->i_mutex));
        dentry = lookup_one_len(name, dir, strlen(name));
        if (!IS_ERR(dentry)) {
-               error = cgroup_create_file(dentry, 0644 | S_IFREG,
+               mode = cgroup_file_mode(cft);
+               error = cgroup_create_file(dentry, mode | S_IFREG,
                                                cgrp->root->sb);
                if (!error)
                        dentry->d_fsdata = (void *)cft;
@@ -2011,14 +2115,16 @@ int cgroup_scan_tasks(struct cgroup_scanner *scan)
  */
 static int pid_array_load(pid_t *pidarray, int npids, struct cgroup *cgrp)
 {
-       int n = 0;
+       int n = 0, pid;
        struct cgroup_iter it;
        struct task_struct *tsk;
        cgroup_iter_start(cgrp, &it);
        while ((tsk = cgroup_iter_next(cgrp, &it))) {
                if (unlikely(n == npids))
                        break;
-               pidarray[n++] = task_pid_vnr(tsk);
+               pid = task_pid_vnr(tsk);
+               if (pid > 0)
+                       pidarray[n++] = pid;
        }
        cgroup_iter_end(cgrp, &it);
        return n;
@@ -2050,7 +2156,6 @@ int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
 
        ret = 0;
        cgrp = dentry->d_fsdata;
-       rcu_read_lock();
 
        cgroup_iter_start(cgrp, &it);
        while ((tsk = cgroup_iter_next(cgrp, &it))) {
@@ -2075,7 +2180,6 @@ int cgroupstats_build(struct cgroupstats *stats, struct dentry *dentry)
        }
        cgroup_iter_end(cgrp, &it);
 
-       rcu_read_unlock();
 err:
        return ret;
 }
@@ -2274,6 +2378,7 @@ static struct cftype files[] = {
                .write_u64 = cgroup_tasks_write,
                .release = cgroup_tasks_release,
                .private = FILE_TASKLIST,
+               .mode = S_IRUGO | S_IWUSR,
        },
 
        {
@@ -2313,6 +2418,17 @@ static int cgroup_populate_dir(struct cgroup *cgrp)
                if (ss->populate && (err = ss->populate(ss, cgrp)) < 0)
                        return err;
        }
+       /* This cgroup is ready now */
+       for_each_subsys(cgrp->root, ss) {
+               struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
+               /*
+                * 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);
+       }
 
        return 0;
 }
@@ -2322,14 +2438,38 @@ static void init_cgroup_css(struct cgroup_subsys_state *css,
                               struct cgroup *cgrp)
 {
        css->cgroup = cgrp;
-       atomic_set(&css->refcnt, 0);
+       atomic_set(&css->refcnt, 1);
        css->flags = 0;
+       css->id = NULL;
        if (cgrp == dummytop)
                set_bit(CSS_ROOT, &css->flags);
        BUG_ON(cgrp->subsys[ss->subsys_id]);
        cgrp->subsys[ss->subsys_id] = css;
 }
 
+static void cgroup_lock_hierarchy(struct cgroupfs_root *root)
+{
+       /* We need to take each hierarchy_mutex in a consistent order */
+       int i;
+
+       for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
+               struct cgroup_subsys *ss = subsys[i];
+               if (ss->root == root)
+                       mutex_lock(&ss->hierarchy_mutex);
+       }
+}
+
+static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
+{
+       int i;
+
+       for (i = 0; i < CGROUP_SUBSYS_COUNT; i++) {
+               struct cgroup_subsys *ss = subsys[i];
+               if (ss->root == root)
+                       mutex_unlock(&ss->hierarchy_mutex);
+       }
+}
+
 /*
  * cgroup_create - create a cgroup
  * @parent: cgroup that will be parent of the new cgroup
@@ -2339,7 +2479,7 @@ static void init_cgroup_css(struct cgroup_subsys_state *css,
  * Must be called with the mutex on the parent inode held
  */
 static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
-                            int mode)
+                            mode_t mode)
 {
        struct cgroup *cgrp;
        struct cgroupfs_root *root = parent->root;
@@ -2376,9 +2516,15 @@ static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
                        goto err_destroy;
                }
                init_cgroup_css(css, ss, cgrp);
+               if (ss->use_id)
+                       if (alloc_css_id(ss, parent, cgrp))
+                               goto err_destroy;
+               /* At error, ->destroy() callback has to free assigned ID. */
        }
 
+       cgroup_lock_hierarchy(root);
        list_add(&cgrp->sibling, &cgrp->parent->children);
+       cgroup_unlock_hierarchy(root);
        root->number_of_cgroups++;
 
        err = cgroup_create_dir(cgrp, dentry, mode);
@@ -2398,7 +2544,9 @@ static long cgroup_create(struct cgroup *parent, struct dentry *dentry,
 
  err_remove:
 
+       cgroup_lock_hierarchy(root);
        list_del(&cgrp->sibling);
+       cgroup_unlock_hierarchy(root);
        root->number_of_cgroups--;
 
  err_destroy:
@@ -2429,7 +2577,7 @@ static int cgroup_has_css_refs(struct cgroup *cgrp)
 {
        /* Check the reference count on each subsystem. Since we
         * already established that there are no tasks in the
-        * cgroup, if the css refcount is also 0, then there should
+        * cgroup, if the css refcount is also 1, then there should
         * be no outstanding references, so the subsystem is safe to
         * destroy. We scan across all subsystems rather than using
         * the per-hierarchy linked list of mounted subsystems since
@@ -2450,20 +2598,75 @@ static int cgroup_has_css_refs(struct cgroup *cgrp)
                 * matter, since it can only happen if the cgroup
                 * has been deleted and hence no longer needs the
                 * release agent to be called anyway. */
-               if (css && atomic_read(&css->refcnt))
+               if (css && (atomic_read(&css->refcnt) > 1))
                        return 1;
        }
        return 0;
 }
 
+/*
+ * Atomically mark all (or else none) of the cgroup's CSS objects as
+ * CSS_REMOVED. Return true on success, or false if the cgroup has
+ * busy subsystems. Call with cgroup_mutex held
+ */
+
+static int cgroup_clear_css_refs(struct cgroup *cgrp)
+{
+       struct cgroup_subsys *ss;
+       unsigned long flags;
+       bool failed = false;
+       local_irq_save(flags);
+       for_each_subsys(cgrp->root, ss) {
+               struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
+               int refcnt;
+               while (1) {
+                       /* We can only remove a CSS with a refcnt==1 */
+                       refcnt = atomic_read(&css->refcnt);
+                       if (refcnt > 1) {
+                               failed = true;
+                               goto done;
+                       }
+                       BUG_ON(!refcnt);
+                       /*
+                        * Drop the refcnt to 0 while we check other
+                        * subsystems. This will cause any racing
+                        * css_tryget() to spin until we set the
+                        * CSS_REMOVED bits or abort
+                        */
+                       if (atomic_cmpxchg(&css->refcnt, refcnt, 0) == refcnt)
+                               break;
+                       cpu_relax();
+               }
+       }
+ done:
+       for_each_subsys(cgrp->root, ss) {
+               struct cgroup_subsys_state *css = cgrp->subsys[ss->subsys_id];
+               if (failed) {
+                       /*
+                        * Restore old refcnt if we previously managed
+                        * to clear it from 1 to 0
+                        */
+                       if (!atomic_read(&css->refcnt))
+                               atomic_set(&css->refcnt, 1);
+               } else {
+                       /* Commit the fact that the CSS is removed */
+                       set_bit(CSS_REMOVED, &css->flags);
+               }
+       }
+       local_irq_restore(flags);
+       return !failed;
+}
+
 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
 {
        struct cgroup *cgrp = dentry->d_fsdata;
        struct dentry *d;
        struct cgroup *parent;
+       DEFINE_WAIT(wait);
+       int ret;
 
        /* the vfs holds both inode->i_mutex already */
-
+again:
        mutex_lock(&cgroup_mutex);
        if (atomic_read(&cgrp->count) != 0) {
                mutex_unlock(&cgroup_mutex);
@@ -2479,25 +2682,51 @@ static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry)
         * Call pre_destroy handlers of subsys. Notify subsystems
         * that rmdir() request comes.
         */
-       cgroup_call_pre_destroy(cgrp);
+       ret = cgroup_call_pre_destroy(cgrp);
+       if (ret)
+               return ret;
 
        mutex_lock(&cgroup_mutex);
        parent = cgrp->parent;
-
-       if (atomic_read(&cgrp->count)
-           || !list_empty(&cgrp->children)
-           || cgroup_has_css_refs(cgrp)) {
+       if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
                mutex_unlock(&cgroup_mutex);
                return -EBUSY;
        }
+       /*
+        * css_put/get is provided for subsys to grab refcnt to css. In typical
+        * case, subsystem has no reference after pre_destroy(). But, under
+        * hierarchy management, some *temporal* refcnt can be hold.
+        * To avoid returning -EBUSY to a user, waitqueue is used. If subsys
+        * is really busy, it should return -EBUSY at pre_destroy(). wake_up
+        * is called when css_put() is called and refcnt goes down to 0.
+        */
+       set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
+       prepare_to_wait(&cgroup_rmdir_waitq, &wait, TASK_INTERRUPTIBLE);
+
+       if (!cgroup_clear_css_refs(cgrp)) {
+               mutex_unlock(&cgroup_mutex);
+               schedule();
+               finish_wait(&cgroup_rmdir_waitq, &wait);
+               clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
+               if (signal_pending(current))
+                       return -EINTR;
+               goto again;
+       }
+       /* NO css_tryget() can success after here. */
+       finish_wait(&cgroup_rmdir_waitq, &wait);
+       clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
 
        spin_lock(&release_list_lock);
        set_bit(CGRP_REMOVED, &cgrp->flags);
        if (!list_empty(&cgrp->release_list))
                list_del(&cgrp->release_list);
        spin_unlock(&release_list_lock);
-       /* delete my sibling from parent->children */
+
+       cgroup_lock_hierarchy(cgrp->root);
+       /* delete this cgroup from parent->children */
        list_del(&cgrp->sibling);
+       cgroup_unlock_hierarchy(cgrp->root);
+
        spin_lock(&cgrp->dentry->d_lock);
        d = dget(cgrp->dentry);
        spin_unlock(&d->d_lock);
@@ -2519,6 +2748,7 @@ static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
        printk(KERN_INFO "Initializing cgroup subsys %s\n", ss->name);
 
        /* Create the top cgroup state for this subsystem */
+       list_add(&ss->sibling, &rootnode.subsys_list);
        ss->root = &rootnode;
        css = ss->create(ss, dummytop);
        /* We don't handle early failures gracefully */
@@ -2538,6 +2768,8 @@ static void __init cgroup_init_subsys(struct cgroup_subsys *ss)
         * need to invoke fork callbacks here. */
        BUG_ON(!list_empty(&init_task.tasks));
 
+       mutex_init(&ss->hierarchy_mutex);
+       lockdep_set_class(&ss->hierarchy_mutex, &ss->subsys_key);
        ss->active = 1;
 }
 
@@ -2556,7 +2788,6 @@ int __init cgroup_init_early(void)
        INIT_HLIST_NODE(&init_css_set.hlist);
        css_set_count = 1;
        init_cgroup_root(&rootnode);
-       list_add(&rootnode.root_list, &roots);
        root_count = 1;
        init_task.cgroups = &init_css_set;
 
@@ -2608,6 +2839,8 @@ int __init cgroup_init(void)
                struct cgroup_subsys *ss = subsys[i];
                if (!ss->early_init)
                        cgroup_init_subsys(ss);
+               if (ss->use_id)
+                       cgroup_subsys_init_idr(ss);
        }
 
        /* Add init_css_set to the hash table */
@@ -2663,15 +2896,12 @@ static int proc_cgroup_show(struct seq_file *m, void *v)
 
        mutex_lock(&cgroup_mutex);
 
-       for_each_root(root) {
+       for_each_active_root(root) {
                struct cgroup_subsys *ss;
                struct cgroup *cgrp;
                int subsys_id;
                int count = 0;
 
-               /* Skip this hierarchy if it has no active subsystems */
-               if (!root->actual_subsys_bits)
-                       continue;
                seq_printf(m, "%lu:", root->subsys_bits);
                for_each_subsys(root, ss)
                        seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
@@ -2903,18 +3133,21 @@ int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
                mutex_unlock(&cgroup_mutex);
                return 0;
        }
-       cg = tsk->cgroups;
-       parent = task_cgroup(tsk, subsys->subsys_id);
 
        /* Pin the hierarchy */
-       if (!atomic_inc_not_zero(&parent->root->sb->s_active)) {
+       if (!atomic_inc_not_zero(&root->sb->s_active)) {
                /* We race with the final deactivate_super() */
                mutex_unlock(&cgroup_mutex);
                return 0;
        }
 
        /* Keep the cgroup alive */
+       task_lock(tsk);
+       parent = task_cgroup(tsk, subsys->subsys_id);
+       cg = tsk->cgroups;
        get_css_set(cg);
+       task_unlock(tsk);
+
        mutex_unlock(&cgroup_mutex);
 
        /* Now do the VFS work to create a cgroup */
@@ -2953,7 +3186,7 @@ int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
                mutex_unlock(&inode->i_mutex);
                put_css_set(cg);
 
-               deactivate_super(parent->root->sb);
+               deactivate_super(root->sb);
                /* The cgroup is still accessible in the VFS, but
                 * we're not going to try to rmdir() it at this
                 * point. */
@@ -2979,23 +3212,24 @@ int cgroup_clone(struct task_struct *tsk, struct cgroup_subsys *subsys,
        mutex_lock(&cgroup_mutex);
        put_css_set(cg);
        mutex_unlock(&cgroup_mutex);
-       deactivate_super(parent->root->sb);
+       deactivate_super(root->sb);
        return ret;
 }
 
 /**
- * cgroup_is_descendant - see if @cgrp is a descendant of current task's cgrp
+ * cgroup_is_descendant - see if @cgrp is a descendant of @task's cgrp
  * @cgrp: the cgroup in question
+ * @task: the task in question
  *
- * See if @cgrp is a descendant of the current task's cgroup in
- * the appropriate hierarchy.
+ * See if @cgrp is a descendant of @task's cgroup in the appropriate
+ * hierarchy.
  *
  * If we are sending in dummytop, then presumably we are creating
  * the top cgroup in the subsystem.
  *
  * Called only by the ns (nsproxy) cgroup.
  */
-int cgroup_is_descendant(const struct cgroup *cgrp)
+int cgroup_is_descendant(const struct cgroup *cgrp, struct task_struct *task)
 {
        int ret;
        struct cgroup *target;
@@ -3005,7 +3239,7 @@ int cgroup_is_descendant(const struct cgroup *cgrp)
                return 1;
 
        get_first_subsys(cgrp, NULL, &subsys_id);
-       target = task_cgroup(current, subsys_id);
+       target = task_cgroup(task, subsys_id);
        while (cgrp != target && cgrp!= cgrp->top_cgroup)
                cgrp = cgrp->parent;
        ret = (cgrp == target);
@@ -3038,9 +3272,12 @@ void __css_put(struct cgroup_subsys_state *css)
 {
        struct cgroup *cgrp = css->cgroup;
        rcu_read_lock();
-       if (atomic_dec_and_test(&css->refcnt) && notify_on_release(cgrp)) {
-               set_bit(CGRP_RELEASABLE, &cgrp->flags);
-               check_for_release(cgrp);
+       if (atomic_dec_return(&css->refcnt) == 1) {
+               if (notify_on_release(cgrp)) {
+                       set_bit(CGRP_RELEASABLE, &cgrp->flags);
+                       check_for_release(cgrp);
+               }
+               cgroup_wakeup_rmdir_waiters(cgrp);
        }
        rcu_read_unlock();
 }
@@ -3140,3 +3377,232 @@ static int __init cgroup_disable(char *str)
        return 1;
 }
 __setup("cgroup_disable=", cgroup_disable);
+
+/*
+ * Functons for CSS ID.
+ */
+
+/*
+ *To get ID other than 0, this should be called when !cgroup_is_removed().
+ */
+unsigned short css_id(struct cgroup_subsys_state *css)
+{
+       struct css_id *cssid = rcu_dereference(css->id);
+
+       if (cssid)
+               return cssid->id;
+       return 0;
+}
+
+unsigned short css_depth(struct cgroup_subsys_state *css)
+{
+       struct css_id *cssid = rcu_dereference(css->id);
+
+       if (cssid)
+               return cssid->depth;
+       return 0;
+}
+
+bool css_is_ancestor(struct cgroup_subsys_state *child,
+                   const struct cgroup_subsys_state *root)
+{
+       struct css_id *child_id = rcu_dereference(child->id);
+       struct css_id *root_id = rcu_dereference(root->id);
+
+       if (!child_id || !root_id || (child_id->depth < root_id->depth))
+               return false;
+       return child_id->stack[root_id->depth] == root_id->id;
+}
+
+static void __free_css_id_cb(struct rcu_head *head)
+{
+       struct css_id *id;
+
+       id = container_of(head, struct css_id, rcu_head);
+       kfree(id);
+}
+
+void free_css_id(struct cgroup_subsys *ss, struct cgroup_subsys_state *css)
+{
+       struct css_id *id = css->id;
+       /* When this is called before css_id initialization, id can be NULL */
+       if (!id)
+               return;
+
+       BUG_ON(!ss->use_id);
+
+       rcu_assign_pointer(id->css, NULL);
+       rcu_assign_pointer(css->id, NULL);
+       spin_lock(&ss->id_lock);
+       idr_remove(&ss->idr, id->id);
+       spin_unlock(&ss->id_lock);
+       call_rcu(&id->rcu_head, __free_css_id_cb);
+}
+
+/*
+ * This is called by init or create(). Then, calls to this function are
+ * always serialized (By cgroup_mutex() at create()).
+ */
+
+static struct css_id *get_new_cssid(struct cgroup_subsys *ss, int depth)
+{
+       struct css_id *newid;
+       int myid, error, size;
+
+       BUG_ON(!ss->use_id);
+
+       size = sizeof(*newid) + sizeof(unsigned short) * (depth + 1);
+       newid = kzalloc(size, GFP_KERNEL);
+       if (!newid)
+               return ERR_PTR(-ENOMEM);
+       /* get id */
+       if (unlikely(!idr_pre_get(&ss->idr, GFP_KERNEL))) {
+               error = -ENOMEM;
+               goto err_out;
+       }
+       spin_lock(&ss->id_lock);
+       /* Don't use 0. allocates an ID of 1-65535 */
+       error = idr_get_new_above(&ss->idr, newid, 1, &myid);
+       spin_unlock(&ss->id_lock);
+
+       /* Returns error when there are no free spaces for new ID.*/
+       if (error) {
+               error = -ENOSPC;
+               goto err_out;
+       }
+       if (myid > CSS_ID_MAX)
+               goto remove_idr;
+
+       newid->id = myid;
+       newid->depth = depth;
+       return newid;
+remove_idr:
+       error = -ENOSPC;
+       spin_lock(&ss->id_lock);
+       idr_remove(&ss->idr, myid);
+       spin_unlock(&ss->id_lock);
+err_out:
+       kfree(newid);
+       return ERR_PTR(error);
+
+}
+
+static int __init cgroup_subsys_init_idr(struct cgroup_subsys *ss)
+{
+       struct css_id *newid;
+       struct cgroup_subsys_state *rootcss;
+
+       spin_lock_init(&ss->id_lock);
+       idr_init(&ss->idr);
+
+       rootcss = init_css_set.subsys[ss->subsys_id];
+       newid = get_new_cssid(ss, 0);
+       if (IS_ERR(newid))
+               return PTR_ERR(newid);
+
+       newid->stack[0] = newid->id;
+       newid->css = rootcss;
+       rootcss->id = newid;
+       return 0;
+}
+
+static int alloc_css_id(struct cgroup_subsys *ss, struct cgroup *parent,
+                       struct cgroup *child)
+{
+       int subsys_id, i, depth = 0;
+       struct cgroup_subsys_state *parent_css, *child_css;
+       struct css_id *child_id, *parent_id = NULL;
+
+       subsys_id = ss->subsys_id;
+       parent_css = parent->subsys[subsys_id];
+       child_css = child->subsys[subsys_id];
+       depth = css_depth(parent_css) + 1;
+       parent_id = parent_css->id;
+
+       child_id = get_new_cssid(ss, depth);
+       if (IS_ERR(child_id))
+               return PTR_ERR(child_id);
+
+       for (i = 0; i < depth; i++)
+               child_id->stack[i] = parent_id->stack[i];
+       child_id->stack[depth] = child_id->id;
+       /*
+        * child_id->css pointer will be set after this cgroup is available
+        * see cgroup_populate_dir()
+        */
+       rcu_assign_pointer(child_css->id, child_id);
+
+       return 0;
+}
+
+/**
+ * css_lookup - lookup css by id
+ * @ss: cgroup subsys to be looked into.
+ * @id: the id
+ *
+ * Returns pointer to cgroup_subsys_state if there is valid one with id.
+ * NULL if not. Should be called under rcu_read_lock()
+ */
+struct cgroup_subsys_state *css_lookup(struct cgroup_subsys *ss, int id)
+{
+       struct css_id *cssid = NULL;
+
+       BUG_ON(!ss->use_id);
+       cssid = idr_find(&ss->idr, id);
+
+       if (unlikely(!cssid))
+               return NULL;
+
+       return rcu_dereference(cssid->css);
+}
+
+/**
+ * css_get_next - lookup next cgroup under specified hierarchy.
+ * @ss: pointer to subsystem
+ * @id: current position of iteration.
+ * @root: pointer to css. search tree under this.
+ * @foundid: position of found object.
+ *
+ * Search next css under the specified hierarchy of rootid. Calling under
+ * rcu_read_lock() is necessary. Returns NULL if it reaches the end.
+ */
+struct cgroup_subsys_state *
+css_get_next(struct cgroup_subsys *ss, int id,
+            struct cgroup_subsys_state *root, int *foundid)
+{
+       struct cgroup_subsys_state *ret = NULL;
+       struct css_id *tmp;
+       int tmpid;
+       int rootid = css_id(root);
+       int depth = css_depth(root);
+
+       if (!rootid)
+               return NULL;
+
+       BUG_ON(!ss->use_id);
+       /* fill start point for scan */
+       tmpid = id;
+       while (1) {
+               /*
+                * scan next entry from bitmap(tree), tmpid is updated after
+                * idr_get_next().
+                */
+               spin_lock(&ss->id_lock);
+               tmp = idr_get_next(&ss->idr, &tmpid);
+               spin_unlock(&ss->id_lock);
+
+               if (!tmp)
+                       break;
+               if (tmp->depth >= depth && tmp->stack[depth] == rootid) {
+                       ret = rcu_dereference(tmp->css);
+                       if (ret) {
+                               *foundid = tmpid;
+                               break;
+                       }
+               }
+               /* continue to scan from next id */
+               tmpid = tmpid + 1;
+       }
+       return ret;
+}
+