diff options
author | 2015-03-28 11:17:36 +0200 | |
---|---|---|
committer | 2015-03-28 20:05:11 +0100 | |
commit | 600b21507eb64bcd85bb87d06c2c2c451b565f02 (patch) | |
tree | 74e5e9ab26ecd95f15ad855c3026800b424e6889 /tools/perf/scripts/python/export-to-postgresql.py | |
parent | Bluetooth: btusb: Consolidate Broadcom based device entries (diff) | |
download | wireguard-linux-600b21507eb64bcd85bb87d06c2c2c451b565f02.tar.xz wireguard-linux-600b21507eb64bcd85bb87d06c2c2c451b565f02.zip |
Bluetooth: Fix race condition with HCI_RESET flag
During the HCI init phase a completed request might be the last part of
the setup procedure after which the actual init procedure starts. The
init procedure begins with a call to hci_reset_req() which sets the
HCI_RESET flag. The purpose of this flag is to make us ignore any
updates to ncmd/cmd_cnt as long as we haven't received the command
complete event for the HCI_Reset. There's a potential race with this
however:
hci_req_cmd_complete(hdev, opcode, status);
if (ev->ncmd && !test_bit(HCI_RESET, &hdev->flags)) {
atomic_set(&hdev->cmd_cnt, 1);
if (!skb_queue_empty(&hdev->cmd_q))
queue_work(hdev->workqueue, &hdev->cmd_work);
}
Since the hci_req_cmd_complete() will trigger the completion of the
setup stage, it's possible that hci_reset_req() gets called before we
try to read ev->ncmd and the HCI_RESET flag. Because of this the cmd_cnt
would never be updated and the hci_reset_req() in practice ends up
blocking itself.
This patch fixes the issue by updating cmd_cnt before notifying the
request completion, and then reading it again to determine whether the
cmd_work should be queued or not.
Signed-off-by: Johan Hedberg <johan.hedberg@intel.com>
Signed-off-by: Marcel Holtmann <marcel@holtmann.org>
Diffstat (limited to 'tools/perf/scripts/python/export-to-postgresql.py')
0 files changed, 0 insertions, 0 deletions