hdev->mesh_pending is extended and walked from syscall context under
hdev->lock, and unlinked and freed from the hci_cmd_sync worker under
hci_req_sync_lock, so the list has no protection; the objects on it are
also handed to hci_cmd_sync_queue() as raw pointers, and those work
entries outlive the list.  Three use-after-frees follow, all reproduced
under KASAN.

Patch 1 removes mgmt_cleanup(), the one walker of the list that cannot
take hdev->lock.  Patch 2 puts the list under hdev->lock on the worker
side and asserts it in the helpers.  Patch 3 gives the object a
reference count for the work entry.  The order matters: the assertions
in patch 2 have no violating caller left once patch 1 is in, and the
reference in patch 3 is only well defined once every unlink happens
under one lock.

Reaching any of this needs CAP_NET_ADMIN in the init user namespace.
hci_sock.c sets HCI_SOCK_TRUSTED at bind under capable() rather than
ns_capable() and gates every later mgmt command on it, and the mesh
commands additionally need HCI_MESH_EXPERIMENTAL.  Nothing in
hci_event.c touches hdev->mesh_pending.

Each reproducer was run at identical parameters on the unpatched and the
patched kernel, built with KASAN, PROVE_LOCKING, DEBUG_LIST and
DEBUG_ATOMIC_SLEEP: every splat the unpatched kernel produces is gone.

Baul Lee (3):
  Bluetooth: MGMT: remove the mesh walk from the socket destructor
  Bluetooth: MGMT: protect hdev->mesh_pending with hdev->lock
  Bluetooth: MGMT: reference-count struct mgmt_mesh_tx

 include/net/bluetooth/bluetooth.h |  1 -
 net/bluetooth/hci_sock.c          |  1 -
 net/bluetooth/mgmt.c              | 72 ++++++++++++++++++++++++---------------
 net/bluetooth/mgmt_util.c         | 31 +++++++++++++++--
 net/bluetooth/mgmt_util.h         |  3 ++
 5 files changed, 76 insertions(+), 32 deletions(-)

-- 
2.50.1 (Apple Git-155)