cgroups: support named cgroups hierarchies
[safe/jmp/linux-2.6] / kernel / cgroup.c
index 9a6c2bf..0320404 100644 (file)
@@ -23,6 +23,7 @@
  */
 
 #include <linux/cgroup.h>
+#include <linux/ctype.h>
 #include <linux/errno.h>
 #include <linux/fs.h>
 #include <linux/kernel.h>
@@ -46,6 +47,8 @@
 #include <linux/cgroupstats.h>
 #include <linux/hash.h>
 #include <linux/namei.h>
+#include <linux/smp_lock.h>
+#include <linux/pid_namespace.h>
 
 #include <asm/atomic.h>
 
@@ -58,6 +61,8 @@ static struct cgroup_subsys *subsys[] = {
 #include <linux/cgroup_subsys.h>
 };
 
+#define MAX_CGROUP_ROOT_NAMELEN 64
+
 /*
  * A cgroupfs_root represents the root of a cgroup hierarchy,
  * and may be associated with a superblock to form an active
@@ -92,6 +97,9 @@ struct cgroupfs_root {
 
        /* The path to use for release notifications. */
        char release_agent_path[PATH_MAX];
+
+       /* The name for this hierarchy - may be empty */
+       char name[MAX_CGROUP_ROOT_NAMELEN];
 };
 
 /*
@@ -594,10 +602,11 @@ void cgroup_unlock(void)
 static int cgroup_mkdir(struct inode *dir, struct dentry *dentry, int mode);
 static int cgroup_rmdir(struct inode *unused_dir, struct dentry *dentry);
 static int cgroup_populate_dir(struct cgroup *cgrp);
-static struct inode_operations cgroup_dir_inode_operations;
+static const struct inode_operations cgroup_dir_inode_operations;
 static struct file_operations proc_cgroupstats_operations;
 
 static struct backing_dev_info cgroup_backing_dev_info = {
+       .name           = "cgroup",
        .capabilities   = BDI_CAP_NO_ACCT_AND_WRITEBACK,
 };
 
@@ -733,16 +742,28 @@ static void cgroup_d_remove_dir(struct dentry *dentry)
  * 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;
+ * CGRP_WAIT_ON_RMDIR flag is set under cgroup's inode->i_mutex;
  */
 DECLARE_WAIT_QUEUE_HEAD(cgroup_rmdir_waitq);
 
-static void cgroup_wakeup_rmdir_waiters(const struct cgroup *cgrp)
+static void cgroup_wakeup_rmdir_waiter(struct cgroup *cgrp)
 {
-       if (unlikely(test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
+       if (unlikely(test_and_clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags)))
                wake_up_all(&cgroup_rmdir_waitq);
 }
 
+void cgroup_exclude_rmdir(struct cgroup_subsys_state *css)
+{
+       css_get(css);
+}
+
+void cgroup_release_and_wakeup_rmdir(struct cgroup_subsys_state *css)
+{
+       cgroup_wakeup_rmdir_waiter(css->cgroup);
+       css_put(css);
+}
+
+
 static int rebind_subsystems(struct cgroupfs_root *root,
                              unsigned long final_bits)
 {
@@ -826,6 +847,8 @@ static int cgroup_show_options(struct seq_file *seq, struct vfsmount *vfs)
                seq_puts(seq, ",noprefix");
        if (strlen(root->release_agent_path))
                seq_printf(seq, ",release_agent=%s", root->release_agent_path);
+       if (strlen(root->name))
+               seq_printf(seq, ",name=%s", root->name);
        mutex_unlock(&cgroup_mutex);
        return 0;
 }
@@ -834,6 +857,9 @@ struct cgroup_sb_opts {
        unsigned long subsys_bits;
        unsigned long flags;
        char *release_agent;
+       char *name;
+
+       struct cgroupfs_root *new_root;
 };
 
 /* Convert a hierarchy specifier into a bitmask of subsystems and
@@ -842,10 +868,13 @@ static int parse_cgroupfs_options(char *data,
                                     struct cgroup_sb_opts *opts)
 {
        char *token, *o = data ?: "all";
+       unsigned long mask = (unsigned long)-1;
 
-       opts->subsys_bits = 0;
-       opts->flags = 0;
-       opts->release_agent = NULL;
+#ifdef CONFIG_CPUSETS
+       mask = ~(1UL << cpuset_subsys_id);
+#endif
+
+       memset(opts, 0, sizeof(*opts));
 
        while ((token = strsep(&o, ",")) != NULL) {
                if (!*token)
@@ -865,11 +894,33 @@ static int parse_cgroupfs_options(char *data,
                        /* Specifying two release agents is forbidden */
                        if (opts->release_agent)
                                return -EINVAL;
-                       opts->release_agent = kzalloc(PATH_MAX, GFP_KERNEL);
+                       opts->release_agent =
+                               kstrndup(token + 14, PATH_MAX, GFP_KERNEL);
                        if (!opts->release_agent)
                                return -ENOMEM;
-                       strncpy(opts->release_agent, token + 14, PATH_MAX - 1);
-                       opts->release_agent[PATH_MAX - 1] = 0;
+               } else if (!strncmp(token, "name=", 5)) {
+                       int i;
+                       const char *name = token + 5;
+                       /* Can't specify an empty name */
+                       if (!strlen(name))
+                               return -EINVAL;
+                       /* Must match [\w.-]+ */
+                       for (i = 0; i < strlen(name); i++) {
+                               char c = name[i];
+                               if (isalnum(c))
+                                       continue;
+                               if ((c == '.') || (c == '-') || (c == '_'))
+                                       continue;
+                               return -EINVAL;
+                       }
+                       /* Specifying two names is forbidden */
+                       if (opts->name)
+                               return -EINVAL;
+                       opts->name = kstrndup(name,
+                                             MAX_CGROUP_ROOT_NAMELEN,
+                                             GFP_KERNEL);
+                       if (!opts->name)
+                               return -ENOMEM;
                } else {
                        struct cgroup_subsys *ss;
                        int i;
@@ -886,8 +937,17 @@ static int parse_cgroupfs_options(char *data,
                }
        }
 
+       /*
+        * Option noprefix was introduced just for backward compatibility
+        * with the old cpuset, so we allow noprefix only if mounting just
+        * the cpuset subsystem.
+        */
+       if (test_bit(ROOT_NOPREFIX, &opts->flags) &&
+           (opts->subsys_bits & mask))
+               return -EINVAL;
+
        /* We can't have an empty hierarchy */
-       if (!opts->subsys_bits)
+       if (!opts->subsys_bits && !opts->name)
                return -EINVAL;
 
        return 0;
@@ -900,6 +960,7 @@ static int cgroup_remount(struct super_block *sb, int *flags, char *data)
        struct cgroup *cgrp = &root->top_cgroup;
        struct cgroup_sb_opts opts;
 
+       lock_kernel();
        mutex_lock(&cgrp->dentry->d_inode->i_mutex);
        mutex_lock(&cgroup_mutex);
 
@@ -914,22 +975,31 @@ static int cgroup_remount(struct super_block *sb, int *flags, char *data)
                goto out_unlock;
        }
 
+       /* Don't allow name to change at remount */
+       if (opts.name && strcmp(opts.name, root->name)) {
+               ret = -EINVAL;
+               goto out_unlock;
+       }
+
        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:
        kfree(opts.release_agent);
+       kfree(opts.name);
        mutex_unlock(&cgroup_mutex);
        mutex_unlock(&cgrp->dentry->d_inode->i_mutex);
+       unlock_kernel();
        return ret;
 }
 
-static struct super_operations cgroup_ops = {
+static const struct super_operations cgroup_ops = {
        .statfs = simple_statfs,
        .drop_inode = generic_delete_inode,
        .show_options = cgroup_show_options,
@@ -942,8 +1012,10 @@ static void init_cgroup_housekeeping(struct cgroup *cgrp)
        INIT_LIST_HEAD(&cgrp->children);
        INIT_LIST_HEAD(&cgrp->css_sets);
        INIT_LIST_HEAD(&cgrp->release_list);
+       INIT_LIST_HEAD(&cgrp->pids_list);
        init_rwsem(&cgrp->pids_mutex);
 }
+
 static void init_cgroup_root(struct cgroupfs_root *root)
 {
        struct cgroup *cgrp = &root->top_cgroup;
@@ -957,31 +1029,59 @@ static void init_cgroup_root(struct cgroupfs_root *root)
 
 static int cgroup_test_super(struct super_block *sb, void *data)
 {
-       struct cgroupfs_root *new = data;
+       struct cgroup_sb_opts *opts = data;
        struct cgroupfs_root *root = sb->s_fs_info;
 
-       /* First check subsystems */
-       if (new->subsys_bits != root->subsys_bits)
-           return 0;
+       /* If we asked for a name then it must match */
+       if (opts->name && strcmp(opts->name, root->name))
+               return 0;
 
-       /* Next check flags */
-       if (new->flags != root->flags)
+       /* If we asked for subsystems then they must match */
+       if (opts->subsys_bits && (opts->subsys_bits != root->subsys_bits))
                return 0;
 
        return 1;
 }
 
+static struct cgroupfs_root *cgroup_root_from_opts(struct cgroup_sb_opts *opts)
+{
+       struct cgroupfs_root *root;
+
+       /* Empty hierarchies aren't supported */
+       if (!opts->subsys_bits)
+               return NULL;
+
+       root = kzalloc(sizeof(*root), GFP_KERNEL);
+       if (!root)
+               return ERR_PTR(-ENOMEM);
+
+       init_cgroup_root(root);
+       root->subsys_bits = opts->subsys_bits;
+       root->flags = opts->flags;
+       if (opts->release_agent)
+               strcpy(root->release_agent_path, opts->release_agent);
+       if (opts->name)
+               strcpy(root->name, opts->name);
+       return root;
+}
+
 static int cgroup_set_super(struct super_block *sb, void *data)
 {
        int ret;
-       struct cgroupfs_root *root = data;
+       struct cgroup_sb_opts *opts = data;
+
+       /* If we don't have a new root, we can't set up a new sb */
+       if (!opts->new_root)
+               return -EINVAL;
+
+       BUG_ON(!opts->subsys_bits);
 
        ret = set_anon_super(sb, NULL);
        if (ret)
                return ret;
 
-       sb->s_fs_info = root;
-       root->sb = sb;
+       sb->s_fs_info = opts->new_root;
+       opts->new_root->sb = sb;
 
        sb->s_blocksize = PAGE_CACHE_SIZE;
        sb->s_blocksize_bits = PAGE_CACHE_SHIFT;
@@ -1018,48 +1118,43 @@ static int cgroup_get_sb(struct file_system_type *fs_type,
                         void *data, struct vfsmount *mnt)
 {
        struct cgroup_sb_opts opts;
+       struct cgroupfs_root *root;
        int ret = 0;
        struct super_block *sb;
-       struct cgroupfs_root *root;
-       struct list_head tmp_cg_links;
+       struct cgroupfs_root *new_root;
 
        /* First find the desired set of subsystems */
        ret = parse_cgroupfs_options(data, &opts);
-       if (ret) {
-               kfree(opts.release_agent);
-               return ret;
-       }
-
-       root = kzalloc(sizeof(*root), GFP_KERNEL);
-       if (!root) {
-               kfree(opts.release_agent);
-               return -ENOMEM;
-       }
+       if (ret)
+               goto out_err;
 
-       init_cgroup_root(root);
-       root->subsys_bits = opts.subsys_bits;
-       root->flags = opts.flags;
-       if (opts.release_agent) {
-               strcpy(root->release_agent_path, opts.release_agent);
-               kfree(opts.release_agent);
+       /*
+        * Allocate a new cgroup root. We may not need it if we're
+        * reusing an existing hierarchy.
+        */
+       new_root = cgroup_root_from_opts(&opts);
+       if (IS_ERR(new_root)) {
+               ret = PTR_ERR(new_root);
+               goto out_err;
        }
+       opts.new_root = new_root;
 
-       sb = sget(fs_type, cgroup_test_super, cgroup_set_super, root);
-
+       /* Locate an existing or new sb for this hierarchy */
+       sb = sget(fs_type, cgroup_test_super, cgroup_set_super, &opts);
        if (IS_ERR(sb)) {
-               kfree(root);
-               return PTR_ERR(sb);
+               ret = PTR_ERR(sb);
+               kfree(opts.new_root);
+               goto out_err;
        }
 
-       if (sb->s_fs_info != root) {
-               /* Reusing an existing superblock */
-               BUG_ON(sb->s_root == NULL);
-               kfree(root);
-               root = NULL;
-       } else {
-               /* New superblock */
+       root = sb->s_fs_info;
+       BUG_ON(!root);
+       if (root == opts.new_root) {
+               /* We used the new root structure, so this is a new hierarchy */
+               struct list_head tmp_cg_links;
                struct cgroup *root_cgrp = &root->top_cgroup;
                struct inode *inode;
+               struct cgroupfs_root *existing_root;
                int i;
 
                BUG_ON(sb->s_root != NULL);
@@ -1072,6 +1167,18 @@ static int cgroup_get_sb(struct file_system_type *fs_type,
                mutex_lock(&inode->i_mutex);
                mutex_lock(&cgroup_mutex);
 
+               if (strlen(root->name)) {
+                       /* Check for name clashes with existing mounts */
+                       for_each_active_root(existing_root) {
+                               if (!strcmp(existing_root->name, root->name)) {
+                                       ret = -EBUSY;
+                                       mutex_unlock(&cgroup_mutex);
+                                       mutex_unlock(&inode->i_mutex);
+                                       goto drop_new_super;
+                               }
+                       }
+               }
+
                /*
                 * We're accessing css_set_count without locking
                 * css_set_lock here, but that's OK - it can only be
@@ -1090,7 +1197,8 @@ static int cgroup_get_sb(struct file_system_type *fs_type,
                if (ret == -EBUSY) {
                        mutex_unlock(&cgroup_mutex);
                        mutex_unlock(&inode->i_mutex);
-                       goto free_cg_links;
+                       free_cg_links(&tmp_cg_links);
+                       goto drop_new_super;
                }
 
                /* EBUSY should be the only error here */
@@ -1122,18 +1230,27 @@ static int cgroup_get_sb(struct file_system_type *fs_type,
                BUG_ON(root->number_of_cgroups != 1);
 
                cgroup_populate_dir(root_cgrp);
-               mutex_unlock(&inode->i_mutex);
                mutex_unlock(&cgroup_mutex);
+               mutex_unlock(&inode->i_mutex);
+       } else {
+               /*
+                * We re-used an existing hierarchy - the new root (if
+                * any) is not needed
+                */
+               kfree(opts.new_root);
        }
 
        simple_set_mnt(mnt, sb);
+       kfree(opts.release_agent);
+       kfree(opts.name);
        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);
+ out_err:
+       kfree(opts.release_agent);
+       kfree(opts.name);
+
        return ret;
 }
 
@@ -1340,7 +1457,7 @@ int cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
         * wake up rmdir() waiter. the rmdir should fail since the cgroup
         * is no longer empty.
         */
-       cgroup_wakeup_rmdir_waiters(cgrp);
+       cgroup_wakeup_rmdir_waiter(cgrp);
        return 0;
 }
 
@@ -1679,14 +1796,14 @@ static struct file_operations cgroup_file_operations = {
        .release = cgroup_file_release,
 };
 
-static struct inode_operations cgroup_dir_inode_operations = {
+static const struct inode_operations cgroup_dir_inode_operations = {
        .lookup = simple_lookup,
        .mkdir = cgroup_mkdir,
        .rmdir = cgroup_rmdir,
        .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 const struct dentry_operations cgroup_dops = {
@@ -1732,7 +1849,7 @@ 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;
@@ -1750,6 +1867,33 @@ 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)
@@ -1757,6 +1901,7 @@ 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)) {
@@ -1767,7 +1912,8 @@ 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;
@@ -2155,12 +2301,30 @@ err:
        return ret;
 }
 
+/*
+ * Cache pids for all threads in the same pid namespace that are
+ * opening the same "tasks" file.
+ */
+struct cgroup_pids {
+       /* The node in cgrp->pids_list */
+       struct list_head list;
+       /* The cgroup those pids belong to */
+       struct cgroup *cgrp;
+       /* The namepsace those pids belong to */
+       struct pid_namespace *ns;
+       /* Array of process ids in the cgroup */
+       pid_t *tasks_pids;
+       /* How many files are using the this tasks_pids array */
+       int use_count;
+       /* Length of the current tasks_pids array */
+       int length;
+};
+
 static int cmppid(const void *a, const void *b)
 {
        return *(pid_t *)a - *(pid_t *)b;
 }
 
-
 /*
  * seq_file methods for the "tasks" file. The seq_file position is the
  * next pid to display; the seq_file iterator is a pointer to the pid
@@ -2175,45 +2339,47 @@ static void *cgroup_tasks_start(struct seq_file *s, loff_t *pos)
         * after a seek to the start). Use a binary-search to find the
         * next pid to display, if any
         */
-       struct cgroup *cgrp = s->private;
+       struct cgroup_pids *cp = s->private;
+       struct cgroup *cgrp = cp->cgrp;
        int index = 0, pid = *pos;
        int *iter;
 
        down_read(&cgrp->pids_mutex);
        if (pid) {
-               int end = cgrp->pids_length;
+               int end = cp->length;
 
                while (index < end) {
                        int mid = (index + end) / 2;
-                       if (cgrp->tasks_pids[mid] == pid) {
+                       if (cp->tasks_pids[mid] == pid) {
                                index = mid;
                                break;
-                       } else if (cgrp->tasks_pids[mid] <= pid)
+                       } else if (cp->tasks_pids[mid] <= pid)
                                index = mid + 1;
                        else
                                end = mid;
                }
        }
        /* If we're off the end of the array, we're done */
-       if (index >= cgrp->pids_length)
+       if (index >= cp->length)
                return NULL;
        /* Update the abstract position to be the actual pid that we found */
-       iter = cgrp->tasks_pids + index;
+       iter = cp->tasks_pids + index;
        *pos = *iter;
        return iter;
 }
 
 static void cgroup_tasks_stop(struct seq_file *s, void *v)
 {
-       struct cgroup *cgrp = s->private;
+       struct cgroup_pids *cp = s->private;
+       struct cgroup *cgrp = cp->cgrp;
        up_read(&cgrp->pids_mutex);
 }
 
 static void *cgroup_tasks_next(struct seq_file *s, void *v, loff_t *pos)
 {
-       struct cgroup *cgrp = s->private;
+       struct cgroup_pids *cp = s->private;
        int *p = v;
-       int *end = cgrp->tasks_pids + cgrp->pids_length;
+       int *end = cp->tasks_pids + cp->length;
 
        /*
         * Advance to the next pid in the array. If this goes off the
@@ -2233,33 +2399,40 @@ static int cgroup_tasks_show(struct seq_file *s, void *v)
        return seq_printf(s, "%d\n", *(int *)v);
 }
 
-static struct seq_operations cgroup_tasks_seq_operations = {
+static const struct seq_operations cgroup_tasks_seq_operations = {
        .start = cgroup_tasks_start,
        .stop = cgroup_tasks_stop,
        .next = cgroup_tasks_next,
        .show = cgroup_tasks_show,
 };
 
-static void release_cgroup_pid_array(struct cgroup *cgrp)
+static void release_cgroup_pid_array(struct cgroup_pids *cp)
 {
+       struct cgroup *cgrp = cp->cgrp;
+
        down_write(&cgrp->pids_mutex);
-       BUG_ON(!cgrp->pids_use_count);
-       if (!--cgrp->pids_use_count) {
-               kfree(cgrp->tasks_pids);
-               cgrp->tasks_pids = NULL;
-               cgrp->pids_length = 0;
+       BUG_ON(!cp->use_count);
+       if (!--cp->use_count) {
+               list_del(&cp->list);
+               put_pid_ns(cp->ns);
+               kfree(cp->tasks_pids);
+               kfree(cp);
        }
        up_write(&cgrp->pids_mutex);
 }
 
 static int cgroup_tasks_release(struct inode *inode, struct file *file)
 {
-       struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
+       struct seq_file *seq;
+       struct cgroup_pids *cp;
 
        if (!(file->f_mode & FMODE_READ))
                return 0;
 
-       release_cgroup_pid_array(cgrp);
+       seq = file->private_data;
+       cp = seq->private;
+
+       release_cgroup_pid_array(cp);
        return seq_release(inode, file);
 }
 
@@ -2278,6 +2451,8 @@ static struct file_operations cgroup_tasks_operations = {
 static int cgroup_tasks_open(struct inode *unused, struct file *file)
 {
        struct cgroup *cgrp = __d_cgrp(file->f_dentry->d_parent);
+       struct pid_namespace *ns = current->nsproxy->pid_ns;
+       struct cgroup_pids *cp;
        pid_t *pidarray;
        int npids;
        int retval;
@@ -2304,20 +2479,37 @@ static int cgroup_tasks_open(struct inode *unused, struct file *file)
         * array if necessary
         */
        down_write(&cgrp->pids_mutex);
-       kfree(cgrp->tasks_pids);
-       cgrp->tasks_pids = pidarray;
-       cgrp->pids_length = npids;
-       cgrp->pids_use_count++;
+
+       list_for_each_entry(cp, &cgrp->pids_list, list) {
+               if (ns == cp->ns)
+                       goto found;
+       }
+
+       cp = kzalloc(sizeof(*cp), GFP_KERNEL);
+       if (!cp) {
+               up_write(&cgrp->pids_mutex);
+               kfree(pidarray);
+               return -ENOMEM;
+       }
+       cp->cgrp = cgrp;
+       cp->ns = ns;
+       get_pid_ns(ns);
+       list_add(&cp->list, &cgrp->pids_list);
+found:
+       kfree(cp->tasks_pids);
+       cp->tasks_pids = pidarray;
+       cp->length = npids;
+       cp->use_count++;
        up_write(&cgrp->pids_mutex);
 
        file->f_op = &cgroup_tasks_operations;
 
        retval = seq_open(file, &cgroup_tasks_seq_operations);
        if (retval) {
-               release_cgroup_pid_array(cgrp);
+               release_cgroup_pid_array(cp);
                return retval;
        }
-       ((struct seq_file *)file->private_data)->private = cgrp;
+       ((struct seq_file *)file->private_data)->private = cp;
        return 0;
 }
 
@@ -2349,6 +2541,7 @@ static struct cftype files[] = {
                .write_u64 = cgroup_tasks_write,
                .release = cgroup_tasks_release,
                .private = FILE_TASKLIST,
+               .mode = S_IRUGO | S_IWUSR,
        },
 
        {
@@ -2449,7 +2642,7 @@ static void cgroup_unlock_hierarchy(struct cgroupfs_root *root)
  * 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;
@@ -2649,33 +2842,42 @@ again:
        mutex_unlock(&cgroup_mutex);
 
        /*
+        * In general, subsystem has no css->refcnt after pre_destroy(). But
+        * in racy cases, subsystem may have to get css->refcnt after
+        * pre_destroy() and it makes rmdir return with -EBUSY. This sometimes
+        * make rmdir return -EBUSY too often. To avoid that, we use waitqueue
+        * for cgroup's rmdir. CGRP_WAIT_ON_RMDIR is for synchronizing rmdir
+        * and subsystem's reference count handling. Please see css_get/put
+        * and css_tryget() and cgroup_wakeup_rmdir_waiter() implementation.
+        */
+       set_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
+
+       /*
         * Call pre_destroy handlers of subsys. Notify subsystems
         * that rmdir() request comes.
         */
        ret = cgroup_call_pre_destroy(cgrp);
-       if (ret)
+       if (ret) {
+               clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
                return ret;
+       }
 
        mutex_lock(&cgroup_mutex);
        parent = cgrp->parent;
        if (atomic_read(&cgrp->count) || !list_empty(&cgrp->children)) {
+               clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
                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();
+               /*
+                * Because someone may call cgroup_wakeup_rmdir_waiter() before
+                * prepare_to_wait(), we need to check this flag.
+                */
+               if (test_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags))
+                       schedule();
                finish_wait(&cgroup_rmdir_waitq, &wait);
                clear_bit(CGRP_WAIT_ON_RMDIR, &cgrp->flags);
                if (signal_pending(current))
@@ -2875,6 +3077,9 @@ static int proc_cgroup_show(struct seq_file *m, void *v)
                seq_printf(m, "%lu:", root->subsys_bits);
                for_each_subsys(root, ss)
                        seq_printf(m, "%s%s", count++ ? "," : "", ss->name);
+               if (strlen(root->name))
+                       seq_printf(m, "%sname=%s", count ? "," : "",
+                                  root->name);
                seq_putc(m, ':');
                get_first_subsys(&root->top_cgroup, NULL, &subsys_id);
                cgrp = task_cgroup(tsk, subsys_id);
@@ -3247,7 +3452,7 @@ void __css_put(struct cgroup_subsys_state *css)
                        set_bit(CGRP_RELEASABLE, &cgrp->flags);
                        check_for_release(cgrp);
                }
-               cgroup_wakeup_rmdir_waiters(cgrp);
+               cgroup_wakeup_rmdir_waiter(cgrp);
        }
        rcu_read_unlock();
 }
@@ -3374,7 +3579,7 @@ unsigned short css_depth(struct cgroup_subsys_state *css)
 }
 
 bool css_is_ancestor(struct cgroup_subsys_state *child,
-                   struct cgroup_subsys_state *root)
+                   const struct cgroup_subsys_state *root)
 {
        struct css_id *child_id = rcu_dereference(child->id);
        struct css_id *root_id = rcu_dereference(root->id);