Commit a275f9dd authored by Muchun Song's avatar Muchun Song Committed by Kleber Sacilotto de Souza

driver core: Fix use-after-free and double free on glue directory

BugLink: https://bugs.launchpad.net/bugs/1845405

commit ac43432c upstream.

There is a race condition between removing glue directory and adding a new
device under the glue dir. It can be reproduced in following test:

CPU1:                                         CPU2:

device_add()
  get_device_parent()
    class_dir_create_and_add()
      kobject_add_internal()
        create_dir()    // create glue_dir

                                              device_add()
                                                get_device_parent()
                                                  kobject_get() // get glue_dir

device_del()
  cleanup_glue_dir()
    kobject_del(glue_dir)

                                                kobject_add()
                                                  kobject_add_internal()
                                                    create_dir() // in glue_dir
                                                      sysfs_create_dir_ns()
                                                        kernfs_create_dir_ns(sd)

      sysfs_remove_dir() // glue_dir->sd=NULL
      sysfs_put()        // free glue_dir->sd

                                                          // sd is freed
                                                          kernfs_new_node(sd)
                                                            kernfs_get(glue_dir)
                                                            kernfs_add_one()
                                                            kernfs_put()

Before CPU1 remove last child device under glue dir, if CPU2 add a new
device under glue dir, the glue_dir kobject reference count will be
increase to 2 via kobject_get() in get_device_parent(). And CPU2 has
been called kernfs_create_dir_ns(), but not call kernfs_new_node().
Meanwhile, CPU1 call sysfs_remove_dir() and sysfs_put(). This result in
glue_dir->sd is freed and it's reference count will be 0. Then CPU2 call
kernfs_get(glue_dir) will trigger a warning in kernfs_get() and increase
it's reference count to 1. Because glue_dir->sd is freed by CPU1, the next
call kernfs_add_one() by CPU2 will fail(This is also use-after-free)
and call kernfs_put() to decrease reference count. Because the reference
count is decremented to 0, it will also call kmem_cache_free() to free
the glue_dir->sd again. This will result in double free.

In order to avoid this happening, we also should make sure that kernfs_node
for glue_dir is released in CPU1 only when refcount for glue_dir kobj is
1 to fix this race.

The following calltrace is captured in kernel 4.14 with the following patch
applied:

commit 726e4109 ("drivers: core: Remove glue dirs from sysfs earlier")

--------------------------------------------------------------------------
[    3.633703] WARNING: CPU: 4 PID: 513 at .../fs/kernfs/dir.c:494
                Here is WARN_ON(!atomic_read(&kn->count) in kernfs_get().
....
[    3.633986] Call trace:
[    3.633991]  kernfs_create_dir_ns+0xa8/0xb0
[    3.633994]  sysfs_create_dir_ns+0x54/0xe8
[    3.634001]  kobject_add_internal+0x22c/0x3f0
[    3.634005]  kobject_add+0xe4/0x118
[    3.634011]  device_add+0x200/0x870
[    3.634017]  _request_firmware+0x958/0xc38
[    3.634020]  request_firmware_into_buf+0x4c/0x70
....
[    3.634064] kernel BUG at .../mm/slub.c:294!
                Here is BUG_ON(object == fp) in set_freepointer().
....
[    3.634346] Call trace:
[    3.634351]  kmem_cache_free+0x504/0x6b8
[    3.634355]  kernfs_put+0x14c/0x1d8
[    3.634359]  kernfs_create_dir_ns+0x88/0xb0
[    3.634362]  sysfs_create_dir_ns+0x54/0xe8
[    3.634366]  kobject_add_internal+0x22c/0x3f0
[    3.634370]  kobject_add+0xe4/0x118
[    3.634374]  device_add+0x200/0x870
[    3.634378]  _request_firmware+0x958/0xc38
[    3.634381]  request_firmware_into_buf+0x4c/0x70
--------------------------------------------------------------------------

Fixes: 726e4109 ("drivers: core: Remove glue dirs from sysfs earlier")
Signed-off-by: default avatarMuchun Song <smuchun@gmail.com>
Reviewed-by: default avatarMukesh Ojha <mojha@codeaurora.org>
Signed-off-by: default avatarPrateek Sood <prsood@codeaurora.org>
Link: https://lore.kernel.org/r/20190727032122.24639-1-smuchun@gmail.comSigned-off-by: default avatarGreg Kroah-Hartman <gregkh@linuxfoundation.org>
Signed-off-by: default avatarConnor Kuehl <connor.kuehl@canonical.com>
Signed-off-by: default avatarKleber Sacilotto de Souza <kleber.souza@canonical.com>
parent 62169799
......@@ -857,12 +857,63 @@ static inline struct kobject *get_glue_dir(struct device *dev)
*/
static void cleanup_glue_dir(struct device *dev, struct kobject *glue_dir)
{
unsigned int ref;
/* see if we live in a "glue" directory */
if (!live_in_glue_dir(glue_dir, dev))
return;
mutex_lock(&gdp_mutex);
if (!kobject_has_children(glue_dir))
/**
* There is a race condition between removing glue directory
* and adding a new device under the glue directory.
*
* CPU1: CPU2:
*
* device_add()
* get_device_parent()
* class_dir_create_and_add()
* kobject_add_internal()
* create_dir() // create glue_dir
*
* device_add()
* get_device_parent()
* kobject_get() // get glue_dir
*
* device_del()
* cleanup_glue_dir()
* kobject_del(glue_dir)
*
* kobject_add()
* kobject_add_internal()
* create_dir() // in glue_dir
* sysfs_create_dir_ns()
* kernfs_create_dir_ns(sd)
*
* sysfs_remove_dir() // glue_dir->sd=NULL
* sysfs_put() // free glue_dir->sd
*
* // sd is freed
* kernfs_new_node(sd)
* kernfs_get(glue_dir)
* kernfs_add_one()
* kernfs_put()
*
* Before CPU1 remove last child device under glue dir, if CPU2 add
* a new device under glue dir, the glue_dir kobject reference count
* will be increase to 2 in kobject_get(k). And CPU2 has been called
* kernfs_create_dir_ns(). Meanwhile, CPU1 call sysfs_remove_dir()
* and sysfs_put(). This result in glue_dir->sd is freed.
*
* Then the CPU2 will see a stale "empty" but still potentially used
* glue dir around in kernfs_new_node().
*
* In order to avoid this happening, we also should make sure that
* kernfs_node for glue_dir is released in CPU1 only when refcount
* for glue_dir kobj is 1.
*/
ref = atomic_read(&glue_dir->kref.refcount);
if (!kobject_has_children(glue_dir) && !--ref)
kobject_del(glue_dir);
kobject_put(glue_dir);
mutex_unlock(&gdp_mutex);
......
Markdown is supported
0%
or
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment