[PATCH v2 0/2] Bluetooth: Fix data-races in SCO/ISO connect paths

SeungJu Cheon posted 2 patches 6 days, 18 hours ago
net/bluetooth/iso.c | 60 +++++++++++++++++++++++++++++++++------------
net/bluetooth/sco.c | 20 +++++++++++----
2 files changed, 59 insertions(+), 21 deletions(-)
[PATCH v2 0/2] Bluetooth: Fix data-races in SCO/ISO connect paths
Posted by SeungJu Cheon 6 days, 18 hours ago
The connect paths read socket address and config fields without
lock_sock() and pass them to hci_get_route() and hci_connect_*(),
while connect()/bind()/setsockopt() can update them concurrently.

Patch 1 covers ISO (iso_connect_bis/cis, iso_listen_bis,
iso_conn_big_sync), patch 2 covers SCO (sco_connect).

The added lock_sock() sections are reached with the lock not held and
released before hci_get_route()/hci_dev_lock(), so no recursive locking
or new lock ordering is introduced.

Tested on KCSAN + PROVE_LOCKING with VHCI reproducers on the SCO and
ISO CIS connect paths; the hci_get_route() race no longer reproduces
and no lockdep splat is seen.

Changes in v2:
- ISO: cache bc_sid too, and pass cached src/dst/bc_sid to
  __iso_get_sock_listen_by_sid() in iso_listen_bis() (missed in v1)
- ISO: use cached bc_sid in BT_DBG() in iso_connect_bis()
- SCO: also snapshot src, setting and codec; v1 only did dst
- reword: the fix stops torn reads, it does not close the TOCTOU window
- fix the SCO Fixes: tag title

SeungJu Cheon (2):
  Bluetooth: ISO: Fix data-race on iso_pi fields in hci_get_route calls
  Bluetooth: SCO: Fix data-race on sco_pi fields in sco_connect

 net/bluetooth/iso.c | 60 +++++++++++++++++++++++++++++++++------------
 net/bluetooth/sco.c | 20 +++++++++++----
 2 files changed, 59 insertions(+), 21 deletions(-)

-- 
2.52.0