devcg: propagate local changes down the hierarchy
[linux-3.10.git] / security / device_cgroup.c
index 7bd296c..221967d 100644 (file)
@@ -1,5 +1,5 @@
 /*
- * dev_cgroup.c - device cgroup subsystem
+ * device_cgroup.c - device cgroup subsystem
  *
  * Copyright 2007 IBM Corp
  */
@@ -10,6 +10,9 @@
 #include <linux/list.h>
 #include <linux/uaccess.h>
 #include <linux/seq_file.h>
+#include <linux/slab.h>
+#include <linux/rcupdate.h>
+#include <linux/mutex.h>
 
 #define ACC_MKNOD 1
 #define ACC_READ  2
 #define DEV_CHAR  2
 #define DEV_ALL   4  /* this represents all devices */
 
+static DEFINE_MUTEX(devcgroup_mutex);
+
+enum devcg_behavior {
+       DEVCG_DEFAULT_NONE,
+       DEVCG_DEFAULT_ALLOW,
+       DEVCG_DEFAULT_DENY,
+};
+
 /*
- * whitelist locking rules:
- * cgroup_lock() cannot be taken under dev_cgroup->lock.
- * dev_cgroup->lock can be taken with or without cgroup_lock().
- *
- * modifications always require cgroup_lock
- * modifications to a list which is visible require the
- *   dev_cgroup->lock *and* cgroup_lock()
- * walking the list requires dev_cgroup->lock or cgroup_lock().
- *
- * reasoning: dev_whitelist_copy() needs to kmalloc, so needs
- *   a mutex, which the cgroup_lock() is.  Since modifying
- *   a visible list requires both locks, either lock can be
- *   taken for walking the list.
+ * exception list locking rules:
+ * hold devcgroup_mutex for update/read.
+ * hold rcu_read_lock() for read.
  */
 
-struct dev_whitelist_item {
+struct dev_exception_item {
        u32 major, minor;
        short type;
        short access;
@@ -46,8 +47,10 @@ struct dev_whitelist_item {
 
 struct dev_cgroup {
        struct cgroup_subsys_state css;
-       struct list_head whitelist;
-       spinlock_t lock;
+       struct list_head exceptions;
+       enum devcg_behavior behavior;
+       /* temporary list for pending propagation operations */
+       struct list_head propagate_pending;
 };
 
 static inline struct dev_cgroup *css_to_devcgroup(struct cgroup_subsys_state *s)
@@ -67,168 +70,191 @@ static inline struct dev_cgroup *task_devcgroup(struct task_struct *task)
 
 struct cgroup_subsys devices_subsys;
 
-static int devcgroup_can_attach(struct cgroup_subsys *ss,
-               struct cgroup *new_cgroup, struct task_struct *task)
+static int devcgroup_can_attach(struct cgroup *new_cgrp,
+                               struct cgroup_taskset *set)
 {
-       if (current != task && !capable(CAP_SYS_ADMIN))
-                       return -EPERM;
+       struct task_struct *task = cgroup_taskset_first(set);
 
+       if (current != task && !capable(CAP_SYS_ADMIN))
+               return -EPERM;
        return 0;
 }
 
 /*
- * called under cgroup_lock()
+ * called under devcgroup_mutex
  */
-static int dev_whitelist_copy(struct list_head *dest, struct list_head *orig)
+static int dev_exceptions_copy(struct list_head *dest, struct list_head *orig)
 {
-       struct dev_whitelist_item *wh, *tmp, *new;
+       struct dev_exception_item *ex, *tmp, *new;
+
+       lockdep_assert_held(&devcgroup_mutex);
 
-       list_for_each_entry(wh, orig, list) {
-               new = kmalloc(sizeof(*wh), GFP_KERNEL);
+       list_for_each_entry(ex, orig, list) {
+               new = kmemdup(ex, sizeof(*ex), GFP_KERNEL);
                if (!new)
                        goto free_and_exit;
-               new->major = wh->major;
-               new->minor = wh->minor;
-               new->type = wh->type;
-               new->access = wh->access;
                list_add_tail(&new->list, dest);
        }
 
        return 0;
 
 free_and_exit:
-       list_for_each_entry_safe(wh, tmp, dest, list) {
-               list_del(&wh->list);
-               kfree(wh);
+       list_for_each_entry_safe(ex, tmp, dest, list) {
+               list_del(&ex->list);
+               kfree(ex);
        }
        return -ENOMEM;
 }
 
-/* Stupid prototype - don't bother combining existing entries */
 /*
- * called under cgroup_lock()
- * since the list is visible to other tasks, we need the spinlock also
+ * called under devcgroup_mutex
  */
-static int dev_whitelist_add(struct dev_cgroup *dev_cgroup,
-                       struct dev_whitelist_item *wh)
+static int dev_exception_add(struct dev_cgroup *dev_cgroup,
+                            struct dev_exception_item *ex)
 {
-       struct dev_whitelist_item *whcopy, *walk;
+       struct dev_exception_item *excopy, *walk;
 
-       whcopy = kmalloc(sizeof(*whcopy), GFP_KERNEL);
-       if (!whcopy)
+       lockdep_assert_held(&devcgroup_mutex);
+
+       excopy = kmemdup(ex, sizeof(*ex), GFP_KERNEL);
+       if (!excopy)
                return -ENOMEM;
 
-       memcpy(whcopy, wh, sizeof(*whcopy));
-       spin_lock(&dev_cgroup->lock);
-       list_for_each_entry(walk, &dev_cgroup->whitelist, list) {
-               if (walk->type != wh->type)
+       list_for_each_entry(walk, &dev_cgroup->exceptions, list) {
+               if (walk->type != ex->type)
                        continue;
-               if (walk->major != wh->major)
+               if (walk->major != ex->major)
                        continue;
-               if (walk->minor != wh->minor)
+               if (walk->minor != ex->minor)
                        continue;
 
-               walk->access |= wh->access;
-               kfree(whcopy);
-               whcopy = NULL;
+               walk->access |= ex->access;
+               kfree(excopy);
+               excopy = NULL;
        }
 
-       if (whcopy != NULL)
-               list_add_tail_rcu(&whcopy->list, &dev_cgroup->whitelist);
-       spin_unlock(&dev_cgroup->lock);
+       if (excopy != NULL)
+               list_add_tail_rcu(&excopy->list, &dev_cgroup->exceptions);
        return 0;
 }
 
-static void whitelist_item_free(struct rcu_head *rcu)
-{
-       struct dev_whitelist_item *item;
-
-       item = container_of(rcu, struct dev_whitelist_item, rcu);
-       kfree(item);
-}
-
 /*
- * called under cgroup_lock()
- * since the list is visible to other tasks, we need the spinlock also
+ * called under devcgroup_mutex
  */
-static void dev_whitelist_rm(struct dev_cgroup *dev_cgroup,
-                       struct dev_whitelist_item *wh)
+static void dev_exception_rm(struct dev_cgroup *dev_cgroup,
+                            struct dev_exception_item *ex)
 {
-       struct dev_whitelist_item *walk, *tmp;
+       struct dev_exception_item *walk, *tmp;
 
-       spin_lock(&dev_cgroup->lock);
-       list_for_each_entry_safe(walk, tmp, &dev_cgroup->whitelist, list) {
-               if (walk->type == DEV_ALL)
-                       goto remove;
-               if (walk->type != wh->type)
+       lockdep_assert_held(&devcgroup_mutex);
+
+       list_for_each_entry_safe(walk, tmp, &dev_cgroup->exceptions, list) {
+               if (walk->type != ex->type)
                        continue;
-               if (walk->major != ~0 && walk->major != wh->major)
+               if (walk->major != ex->major)
                        continue;
-               if (walk->minor != ~0 && walk->minor != wh->minor)
+               if (walk->minor != ex->minor)
                        continue;
 
-remove:
-               walk->access &= ~wh->access;
+               walk->access &= ~ex->access;
                if (!walk->access) {
                        list_del_rcu(&walk->list);
-                       call_rcu(&walk->rcu, whitelist_item_free);
+                       kfree_rcu(walk, rcu);
                }
        }
-       spin_unlock(&dev_cgroup->lock);
+}
+
+static void __dev_exception_clean(struct dev_cgroup *dev_cgroup)
+{
+       struct dev_exception_item *ex, *tmp;
+
+       list_for_each_entry_safe(ex, tmp, &dev_cgroup->exceptions, list) {
+               list_del_rcu(&ex->list);
+               kfree_rcu(ex, rcu);
+       }
+}
+
+/**
+ * dev_exception_clean - frees all entries of the exception list
+ * @dev_cgroup: dev_cgroup with the exception list to be cleaned
+ *
+ * called under devcgroup_mutex
+ */
+static void dev_exception_clean(struct dev_cgroup *dev_cgroup)
+{
+       lockdep_assert_held(&devcgroup_mutex);
+
+       __dev_exception_clean(dev_cgroup);
+}
+
+static inline bool is_devcg_online(const struct dev_cgroup *devcg)
+{
+       return (devcg->behavior != DEVCG_DEFAULT_NONE);
+}
+
+/**
+ * devcgroup_online - initializes devcgroup's behavior and exceptions based on
+ *                   parent's
+ * @cgroup: cgroup getting online
+ * returns 0 in case of success, error code otherwise
+ */
+static int devcgroup_online(struct cgroup *cgroup)
+{
+       struct dev_cgroup *dev_cgroup, *parent_dev_cgroup = NULL;
+       int ret = 0;
+
+       mutex_lock(&devcgroup_mutex);
+       dev_cgroup = cgroup_to_devcgroup(cgroup);
+       if (cgroup->parent)
+               parent_dev_cgroup = cgroup_to_devcgroup(cgroup->parent);
+
+       if (parent_dev_cgroup == NULL)
+               dev_cgroup->behavior = DEVCG_DEFAULT_ALLOW;
+       else {
+               ret = dev_exceptions_copy(&dev_cgroup->exceptions,
+                                         &parent_dev_cgroup->exceptions);
+               if (!ret)
+                       dev_cgroup->behavior = parent_dev_cgroup->behavior;
+       }
+       mutex_unlock(&devcgroup_mutex);
+
+       return ret;
+}
+
+static void devcgroup_offline(struct cgroup *cgroup)
+{
+       struct dev_cgroup *dev_cgroup = cgroup_to_devcgroup(cgroup);
+
+       mutex_lock(&devcgroup_mutex);
+       dev_cgroup->behavior = DEVCG_DEFAULT_NONE;
+       mutex_unlock(&devcgroup_mutex);
 }
 
 /*
  * called from kernel/cgroup.c with cgroup_lock() held.
  */
-static struct cgroup_subsys_state *devcgroup_create(struct cgroup_subsys *ss,
-                                               struct cgroup *cgroup)
+static struct cgroup_subsys_state *devcgroup_css_alloc(struct cgroup *cgroup)
 {
-       struct dev_cgroup *dev_cgroup, *parent_dev_cgroup;
+       struct dev_cgroup *dev_cgroup;
        struct cgroup *parent_cgroup;
-       int ret;
 
        dev_cgroup = kzalloc(sizeof(*dev_cgroup), GFP_KERNEL);
        if (!dev_cgroup)
                return ERR_PTR(-ENOMEM);
-       INIT_LIST_HEAD(&dev_cgroup->whitelist);
+       INIT_LIST_HEAD(&dev_cgroup->exceptions);
+       INIT_LIST_HEAD(&dev_cgroup->propagate_pending);
+       dev_cgroup->behavior = DEVCG_DEFAULT_NONE;
        parent_cgroup = cgroup->parent;
 
-       if (parent_cgroup == NULL) {
-               struct dev_whitelist_item *wh;
-               wh = kmalloc(sizeof(*wh), GFP_KERNEL);
-               if (!wh) {
-                       kfree(dev_cgroup);
-                       return ERR_PTR(-ENOMEM);
-               }
-               wh->minor = wh->major = ~0;
-               wh->type = DEV_ALL;
-               wh->access = ACC_MASK;
-               list_add(&wh->list, &dev_cgroup->whitelist);
-       } else {
-               parent_dev_cgroup = cgroup_to_devcgroup(parent_cgroup);
-               ret = dev_whitelist_copy(&dev_cgroup->whitelist,
-                               &parent_dev_cgroup->whitelist);
-               if (ret) {
-                       kfree(dev_cgroup);
-                       return ERR_PTR(ret);
-               }
-       }
-
-       spin_lock_init(&dev_cgroup->lock);
        return &dev_cgroup->css;
 }
 
-static void devcgroup_destroy(struct cgroup_subsys *ss,
-                       struct cgroup *cgroup)
+static void devcgroup_css_free(struct cgroup *cgroup)
 {
        struct dev_cgroup *dev_cgroup;
-       struct dev_whitelist_item *wh, *tmp;
 
        dev_cgroup = cgroup_to_devcgroup(cgroup);
-       list_for_each_entry_safe(wh, tmp, &dev_cgroup->whitelist, list) {
-               list_del(&wh->list);
-               kfree(wh);
-       }
+       __dev_exception_clean(dev_cgroup);
        kfree(dev_cgroup);
 }
 
@@ -274,81 +300,240 @@ static int devcgroup_seq_read(struct cgroup *cgroup, struct cftype *cft,
                                struct seq_file *m)
 {
        struct dev_cgroup *devcgroup = cgroup_to_devcgroup(cgroup);
-       struct dev_whitelist_item *wh;
+       struct dev_exception_item *ex;
        char maj[MAJMINLEN], min[MAJMINLEN], acc[ACCLEN];
 
        rcu_read_lock();
-       list_for_each_entry_rcu(wh, &devcgroup->whitelist, list) {
-               set_access(acc, wh->access);
-               set_majmin(maj, wh->major);
-               set_majmin(min, wh->minor);
-               seq_printf(m, "%c %s:%s %s\n", type_to_char(wh->type),
+       /*
+        * To preserve the compatibility:
+        * - Only show the "all devices" when the default policy is to allow
+        * - List the exceptions in case the default policy is to deny
+        * This way, the file remains as a "whitelist of devices"
+        */
+       if (devcgroup->behavior == DEVCG_DEFAULT_ALLOW) {
+               set_access(acc, ACC_MASK);
+               set_majmin(maj, ~0);
+               set_majmin(min, ~0);
+               seq_printf(m, "%c %s:%s %s\n", type_to_char(DEV_ALL),
                           maj, min, acc);
+       } else {
+               list_for_each_entry_rcu(ex, &devcgroup->exceptions, list) {
+                       set_access(acc, ex->access);
+                       set_majmin(maj, ex->major);
+                       set_majmin(min, ex->minor);
+                       seq_printf(m, "%c %s:%s %s\n", type_to_char(ex->type),
+                                  maj, min, acc);
+               }
        }
        rcu_read_unlock();
 
        return 0;
 }
 
-/*
- * may_access_whitelist:
- * does the access granted to dev_cgroup c contain the access
- * requested in whitelist item refwh.
- * return 1 if yes, 0 if no.
- * call with c->lock held
+/**
+ * may_access - verifies if a new exception is part of what is allowed
+ *             by a dev cgroup based on the default policy +
+ *             exceptions. This is used to make sure a child cgroup
+ *             won't have more privileges than its parent or to
+ *             verify if a certain access is allowed.
+ * @dev_cgroup: dev cgroup to be tested against
+ * @refex: new exception
+ * @behavior: behavior of the exception
  */
-static int may_access_whitelist(struct dev_cgroup *c,
-                                      struct dev_whitelist_item *refwh)
+static bool may_access(struct dev_cgroup *dev_cgroup,
+                      struct dev_exception_item *refex,
+                      enum devcg_behavior behavior)
 {
-       struct dev_whitelist_item *whitem;
+       struct dev_exception_item *ex;
+       bool match = false;
 
-       list_for_each_entry(whitem, &c->whitelist, list) {
-               if (whitem->type & DEV_ALL)
-                       return 1;
-               if ((refwh->type & DEV_BLOCK) && !(whitem->type & DEV_BLOCK))
+       rcu_lockdep_assert(rcu_read_lock_held() ||
+                          lockdep_is_held(&devcgroup_mutex),
+                          "device_cgroup::may_access() called without proper synchronization");
+
+       list_for_each_entry_rcu(ex, &dev_cgroup->exceptions, list) {
+               if ((refex->type & DEV_BLOCK) && !(ex->type & DEV_BLOCK))
                        continue;
-               if ((refwh->type & DEV_CHAR) && !(whitem->type & DEV_CHAR))
+               if ((refex->type & DEV_CHAR) && !(ex->type & DEV_CHAR))
                        continue;
-               if (whitem->major != ~0 && whitem->major != refwh->major)
+               if (ex->major != ~0 && ex->major != refex->major)
                        continue;
-               if (whitem->minor != ~0 && whitem->minor != refwh->minor)
+               if (ex->minor != ~0 && ex->minor != refex->minor)
                        continue;
-               if (refwh->access & (~whitem->access))
+               if (refex->access & (~ex->access))
                        continue;
-               return 1;
+               match = true;
+               break;
        }
-       return 0;
+
+       if (dev_cgroup->behavior == DEVCG_DEFAULT_ALLOW) {
+               if (behavior == DEVCG_DEFAULT_ALLOW) {
+                       /* the exception will deny access to certain devices */
+                       return true;
+               } else {
+                       /* the exception will allow access to certain devices */
+                       if (match)
+                               /*
+                                * a new exception allowing access shouldn't
+                                * match an parent's exception
+                                */
+                               return false;
+                       return true;
+               }
+       } else {
+               /* only behavior == DEVCG_DEFAULT_DENY allowed here */
+               if (match)
+                       /* parent has an exception that matches the proposed */
+                       return true;
+               else
+                       return false;
+       }
+       return false;
 }
 
 /*
  * parent_has_perm:
- * when adding a new allow rule to a device whitelist, the rule
+ * when adding a new allow rule to a device exception list, the rule
  * must be allowed in the parent device
  */
 static int parent_has_perm(struct dev_cgroup *childcg,
-                                 struct dev_whitelist_item *wh)
+                                 struct dev_exception_item *ex)
 {
        struct cgroup *pcg = childcg->css.cgroup->parent;
        struct dev_cgroup *parent;
-       int ret;
 
        if (!pcg)
                return 1;
        parent = cgroup_to_devcgroup(pcg);
-       spin_lock(&parent->lock);
-       ret = may_access_whitelist(parent, wh);
-       spin_unlock(&parent->lock);
-       return ret;
+       return may_access(parent, ex, childcg->behavior);
+}
+
+/**
+ * may_allow_all - checks if it's possible to change the behavior to
+ *                allow based on parent's rules.
+ * @parent: device cgroup's parent
+ * returns: != 0 in case it's allowed, 0 otherwise
+ */
+static inline int may_allow_all(struct dev_cgroup *parent)
+{
+       if (!parent)
+               return 1;
+       return parent->behavior == DEVCG_DEFAULT_ALLOW;
+}
+
+/**
+ * revalidate_active_exceptions - walks through the active exception list and
+ *                               revalidates the exceptions based on parent's
+ *                               behavior and exceptions. The exceptions that
+ *                               are no longer valid will be removed.
+ *                               Called with devcgroup_mutex held.
+ * @devcg: cgroup which exceptions will be checked
+ *
+ * This is one of the three key functions for hierarchy implementation.
+ * This function is responsible for re-evaluating all the cgroup's active
+ * exceptions due to a parent's exception change.
+ * Refer to Documentation/cgroups/devices.txt for more details.
+ */
+static void revalidate_active_exceptions(struct dev_cgroup *devcg)
+{
+       struct dev_exception_item *ex;
+       struct list_head *this, *tmp;
+
+       list_for_each_safe(this, tmp, &devcg->exceptions) {
+               ex = container_of(this, struct dev_exception_item, list);
+               if (!parent_has_perm(devcg, ex))
+                       dev_exception_rm(devcg, ex);
+       }
+}
+
+/**
+ * get_online_devcg - walks the cgroup tree and fills a list with the online
+ *                   groups
+ * @root: cgroup used as starting point
+ * @online: list that will be filled with online groups
+ *
+ * Must be called with devcgroup_mutex held. Grabs RCU lock.
+ * Because devcgroup_mutex is held, no devcg will become online or offline
+ * during the tree walk (see devcgroup_online, devcgroup_offline)
+ * A separated list is needed because propagate_behavior() and
+ * propagate_exception() need to allocate memory and can block.
+ */
+static void get_online_devcg(struct cgroup *root, struct list_head *online)
+{
+       struct cgroup *pos;
+       struct dev_cgroup *devcg;
+
+       lockdep_assert_held(&devcgroup_mutex);
+
+       rcu_read_lock();
+       cgroup_for_each_descendant_pre(pos, root) {
+               devcg = cgroup_to_devcgroup(pos);
+               if (is_devcg_online(devcg))
+                       list_add_tail(&devcg->propagate_pending, online);
+       }
+       rcu_read_unlock();
+}
+
+/**
+ * propagate_exception - propagates a new exception to the children
+ * @devcg_root: device cgroup that added a new exception
+ * @ex: new exception to be propagated
+ *
+ * returns: 0 in case of success, != 0 in case of error
+ */
+static int propagate_exception(struct dev_cgroup *devcg_root,
+                              struct dev_exception_item *ex)
+{
+       struct cgroup *root = devcg_root->css.cgroup;
+       struct dev_cgroup *devcg, *parent, *tmp;
+       int rc = 0;
+       LIST_HEAD(pending);
+
+       get_online_devcg(root, &pending);
+
+       list_for_each_entry_safe(devcg, tmp, &pending, propagate_pending) {
+               parent = cgroup_to_devcgroup(devcg->css.cgroup->parent);
+
+               /*
+                * in case both root's behavior and devcg is allow, a new
+                * restriction means adding to the exception list
+                */
+               if (devcg_root->behavior == DEVCG_DEFAULT_ALLOW &&
+                   devcg->behavior == DEVCG_DEFAULT_ALLOW) {
+                       rc = dev_exception_add(devcg, ex);
+                       if (rc)
+                               break;
+               } else {
+                       /*
+                        * in the other possible cases:
+                        * root's behavior: allow, devcg's: deny
+                        * root's behavior: deny, devcg's: deny
+                        * the exception will be removed
+                        */
+                       dev_exception_rm(devcg, ex);
+               }
+               revalidate_active_exceptions(devcg);
+
+               list_del_init(&devcg->propagate_pending);
+       }
+       return rc;
+}
+
+static inline bool has_children(struct dev_cgroup *devcgroup)
+{
+       struct cgroup *cgrp = devcgroup->css.cgroup;
+
+       return !list_empty(&cgrp->children);
 }
 
 /*
- * Modify the whitelist using allow/deny rules.
+ * Modify the exception list using allow/deny rules.
  * CAP_SYS_ADMIN is needed for this.  It's at least separate from CAP_MKNOD
  * so we can give a container CAP_MKNOD to let it create devices but not
- * modify the whitelist.
+ * modify the exception list.
  * It seems likely we'll want to add a CAP_CONTAINER capability to allow
  * us to also grant CAP_SYS_ADMIN to containers without giving away the
- * device whitelist controls, but for now we'll stick with CAP_SYS_ADMIN
+ * device exception list controls, but for now we'll stick with CAP_SYS_ADMIN
  *
  * Taking rules away is always allowed (given CAP_SYS_ADMIN).  Granting
  * new access is only allowed if you're in the top-level cgroup, or your
@@ -357,32 +542,57 @@ static int parent_has_perm(struct dev_cgroup *childcg,
 static int devcgroup_update_access(struct dev_cgroup *devcgroup,
                                   int filetype, const char *buffer)
 {
-       struct dev_cgroup *cur_devcgroup;
        const char *b;
-       char *endp;
-       int retval = 0, count;
-       struct dev_whitelist_item wh;
+       char temp[12];          /* 11 + 1 characters needed for a u32 */
+       int count, rc = 0;
+       struct dev_exception_item ex;
+       struct cgroup *p = devcgroup->css.cgroup;
+       struct dev_cgroup *parent = NULL;
 
        if (!capable(CAP_SYS_ADMIN))
                return -EPERM;
 
-       cur_devcgroup = task_devcgroup(current);
+       if (p->parent)
+               parent = cgroup_to_devcgroup(p->parent);
 
-       memset(&wh, 0, sizeof(wh));
+       memset(&ex, 0, sizeof(ex));
        b = buffer;
 
        switch (*b) {
        case 'a':
-               wh.type = DEV_ALL;
-               wh.access = ACC_MASK;
-               wh.major = ~0;
-               wh.minor = ~0;
-               goto handle;
+               switch (filetype) {
+               case DEVCG_ALLOW:
+                       if (has_children(devcgroup))
+                               return -EINVAL;
+
+                       if (!may_allow_all(parent))
+                               return -EPERM;
+                       dev_exception_clean(devcgroup);
+                       devcgroup->behavior = DEVCG_DEFAULT_ALLOW;
+                       if (!parent)
+                               break;
+
+                       rc = dev_exceptions_copy(&devcgroup->exceptions,
+                                                &parent->exceptions);
+                       if (rc)
+                               return rc;
+                       break;
+               case DEVCG_DENY:
+                       if (has_children(devcgroup))
+                               return -EINVAL;
+
+                       dev_exception_clean(devcgroup);
+                       devcgroup->behavior = DEVCG_DEFAULT_DENY;
+                       break;
+               default:
+                       return -EINVAL;
+               }
+               return 0;
        case 'b':
-               wh.type = DEV_BLOCK;
+               ex.type = DEV_BLOCK;
                break;
        case 'c':
-               wh.type = DEV_CHAR;
+               ex.type = DEV_CHAR;
                break;
        default:
                return -EINVAL;
@@ -392,11 +602,19 @@ static int devcgroup_update_access(struct dev_cgroup *devcgroup,
                return -EINVAL;
        b++;
        if (*b == '*') {
-               wh.major = ~0;
+               ex.major = ~0;
                b++;
        } else if (isdigit(*b)) {
-               wh.major = simple_strtoul(b, &endp, 10);
-               b = endp;
+               memset(temp, 0, sizeof(temp));
+               for (count = 0; count < sizeof(temp) - 1; count++) {
+                       temp[count] = *b;
+                       b++;
+                       if (!isdigit(*b))
+                               break;
+               }
+               rc = kstrtou32(temp, 10, &ex.major);
+               if (rc)
+                       return -EINVAL;
        } else {
                return -EINVAL;
        }
@@ -406,11 +624,19 @@ static int devcgroup_update_access(struct dev_cgroup *devcgroup,
 
        /* read minor */
        if (*b == '*') {
-               wh.minor = ~0;
+               ex.minor = ~0;
                b++;
        } else if (isdigit(*b)) {
-               wh.minor = simple_strtoul(b, &endp, 10);
-               b = endp;
+               memset(temp, 0, sizeof(temp));
+               for (count = 0; count < sizeof(temp) - 1; count++) {
+                       temp[count] = *b;
+                       b++;
+                       if (!isdigit(*b))
+                               break;
+               }
+               rc = kstrtou32(temp, 10, &ex.minor);
+               if (rc)
+                       return -EINVAL;
        } else {
                return -EINVAL;
        }
@@ -419,13 +645,13 @@ static int devcgroup_update_access(struct dev_cgroup *devcgroup,
        for (b++, count = 0; count < 3; count++, b++) {
                switch (*b) {
                case 'r':
-                       wh.access |= ACC_READ;
+                       ex.access |= ACC_READ;
                        break;
                case 'w':
-                       wh.access |= ACC_WRITE;
+                       ex.access |= ACC_WRITE;
                        break;
                case 'm':
-                       wh.access |= ACC_MKNOD;
+                       ex.access |= ACC_MKNOD;
                        break;
                case '\n':
                case '\0':
@@ -436,31 +662,52 @@ static int devcgroup_update_access(struct dev_cgroup *devcgroup,
                }
        }
 
-handle:
-       retval = 0;
        switch (filetype) {
        case DEVCG_ALLOW:
-               if (!parent_has_perm(devcgroup, &wh))
+               if (!parent_has_perm(devcgroup, &ex))
                        return -EPERM;
-               return dev_whitelist_add(devcgroup, &wh);
+               /*
+                * If the default policy is to allow by default, try to remove
+                * an matching exception instead. And be silent about it: we
+                * don't want to break compatibility
+                */
+               if (devcgroup->behavior == DEVCG_DEFAULT_ALLOW) {
+                       dev_exception_rm(devcgroup, &ex);
+                       return 0;
+               }
+               rc = dev_exception_add(devcgroup, &ex);
+               break;
        case DEVCG_DENY:
-               dev_whitelist_rm(devcgroup, &wh);
+               /*
+                * If the default policy is to deny by default, try to remove
+                * an matching exception instead. And be silent about it: we
+                * don't want to break compatibility
+                */
+               if (devcgroup->behavior == DEVCG_DEFAULT_DENY)
+                       dev_exception_rm(devcgroup, &ex);
+               else
+                       rc = dev_exception_add(devcgroup, &ex);
+
+               if (rc)
+                       break;
+               /* we only propagate new restrictions */
+               rc = propagate_exception(devcgroup, &ex);
                break;
        default:
-               return -EINVAL;
+               rc = -EINVAL;
        }
-       return 0;
+       return rc;
 }
 
 static int devcgroup_access_write(struct cgroup *cgrp, struct cftype *cft,
                                  const char *buffer)
 {
        int retval;
-       if (!cgroup_lock_live_group(cgrp))
-               return -ENODEV;
+
+       mutex_lock(&devcgroup_mutex);
        retval = devcgroup_update_access(cgroup_to_devcgroup(cgrp),
                                         cft->private, buffer);
-       cgroup_unlock();
+       mutex_unlock(&devcgroup_mutex);
        return retval;
 }
 
@@ -480,92 +727,93 @@ static struct cftype dev_cgroup_files[] = {
                .read_seq_string = devcgroup_seq_read,
                .private = DEVCG_LIST,
        },
+       { }     /* terminate */
 };
 
-static int devcgroup_populate(struct cgroup_subsys *ss,
-                               struct cgroup *cgroup)
-{
-       return cgroup_add_files(cgroup, ss, dev_cgroup_files,
-                                       ARRAY_SIZE(dev_cgroup_files));
-}
-
 struct cgroup_subsys devices_subsys = {
        .name = "devices",
        .can_attach = devcgroup_can_attach,
-       .create = devcgroup_create,
-       .destroy  = devcgroup_destroy,
-       .populate = devcgroup_populate,
+       .css_alloc = devcgroup_css_alloc,
+       .css_free = devcgroup_css_free,
+       .css_online = devcgroup_online,
+       .css_offline = devcgroup_offline,
        .subsys_id = devices_subsys_id,
+       .base_cftypes = dev_cgroup_files,
+
+       /*
+        * While devices cgroup has the rudimentary hierarchy support which
+        * checks the parent's restriction, it doesn't properly propagates
+        * config changes in ancestors to their descendents.  A child
+        * should only be allowed to add more restrictions to the parent's
+        * configuration.  Fix it and remove the following.
+        */
+       .broken_hierarchy = true,
 };
 
-int devcgroup_inode_permission(struct inode *inode, int mask)
+/**
+ * __devcgroup_check_permission - checks if an inode operation is permitted
+ * @dev_cgroup: the dev cgroup to be tested against
+ * @type: device type
+ * @major: device major number
+ * @minor: device minor number
+ * @access: combination of ACC_WRITE, ACC_READ and ACC_MKNOD
+ *
+ * returns 0 on success, -EPERM case the operation is not permitted
+ */
+static int __devcgroup_check_permission(short type, u32 major, u32 minor,
+                                       short access)
 {
        struct dev_cgroup *dev_cgroup;
-       struct dev_whitelist_item *wh;
+       struct dev_exception_item ex;
+       int rc;
 
-       dev_t device = inode->i_rdev;
-       if (!device)
-               return 0;
-       if (!S_ISBLK(inode->i_mode) && !S_ISCHR(inode->i_mode))
-               return 0;
-       dev_cgroup = css_to_devcgroup(task_subsys_state(current,
-                               devices_subsys_id));
-       if (!dev_cgroup)
-               return 0;
+       memset(&ex, 0, sizeof(ex));
+       ex.type = type;
+       ex.major = major;
+       ex.minor = minor;
+       ex.access = access;
 
        rcu_read_lock();
-       list_for_each_entry_rcu(wh, &dev_cgroup->whitelist, list) {
-               if (wh->type & DEV_ALL)
-                       goto acc_check;
-               if ((wh->type & DEV_BLOCK) && !S_ISBLK(inode->i_mode))
-                       continue;
-               if ((wh->type & DEV_CHAR) && !S_ISCHR(inode->i_mode))
-                       continue;
-               if (wh->major != ~0 && wh->major != imajor(inode))
-                       continue;
-               if (wh->minor != ~0 && wh->minor != iminor(inode))
-                       continue;
-acc_check:
-               if ((mask & MAY_WRITE) && !(wh->access & ACC_WRITE))
-                       continue;
-               if ((mask & MAY_READ) && !(wh->access & ACC_READ))
-                       continue;
-               rcu_read_unlock();
-               return 0;
-       }
+       dev_cgroup = task_devcgroup(current);
+       rc = may_access(dev_cgroup, &ex, dev_cgroup->behavior);
        rcu_read_unlock();
 
-       return -EPERM;
+       if (!rc)
+               return -EPERM;
+
+       return 0;
+}
+
+int __devcgroup_inode_permission(struct inode *inode, int mask)
+{
+       short type, access = 0;
+
+       if (S_ISBLK(inode->i_mode))
+               type = DEV_BLOCK;
+       if (S_ISCHR(inode->i_mode))
+               type = DEV_CHAR;
+       if (mask & MAY_WRITE)
+               access |= ACC_WRITE;
+       if (mask & MAY_READ)
+               access |= ACC_READ;
+
+       return __devcgroup_check_permission(type, imajor(inode), iminor(inode),
+                       access);
 }
 
 int devcgroup_inode_mknod(int mode, dev_t dev)
 {
-       struct dev_cgroup *dev_cgroup;
-       struct dev_whitelist_item *wh;
+       short type;
 
-       dev_cgroup = css_to_devcgroup(task_subsys_state(current,
-                               devices_subsys_id));
-       if (!dev_cgroup)
+       if (!S_ISBLK(mode) && !S_ISCHR(mode))
                return 0;
 
-       rcu_read_lock();
-       list_for_each_entry(wh, &dev_cgroup->whitelist, list) {
-               if (wh->type & DEV_ALL)
-                       goto acc_check;
-               if ((wh->type & DEV_BLOCK) && !S_ISBLK(mode))
-                       continue;
-               if ((wh->type & DEV_CHAR) && !S_ISCHR(mode))
-                       continue;
-               if (wh->major != ~0 && wh->major != MAJOR(dev))
-                       continue;
-               if (wh->minor != ~0 && wh->minor != MINOR(dev))
-                       continue;
-acc_check:
-               if (!(wh->access & ACC_MKNOD))
-                       continue;
-               rcu_read_unlock();
-               return 0;
-       }
-       rcu_read_unlock();
-       return -EPERM;
+       if (S_ISBLK(mode))
+               type = DEV_BLOCK;
+       else
+               type = DEV_CHAR;
+
+       return __devcgroup_check_permission(type, MAJOR(dev), MINOR(dev),
+                       ACC_MKNOD);
+
 }