Hi Archie,
On Fri, Jul 22, 2022 at 3:23 AM Archie Pusaka <apusaka@xxxxxxxxxx> wrote:
>
> From: Archie Pusaka <apusaka@xxxxxxxxxxxx>
>
> When in the middle of adding accept list, the userspace can still
> remove devices, therefore causing crash if the removed device is
> the one being processed.
>
> Use a safe loop mechanism to guard against deletion while iterating
> the pending items.
>
> Below is a sample btsnoop log when user enters wrong passkey when
> pairing a LE keyboard and the corresponding stacktrace.
> @ MGMT Event: Command Complete (0x0001) plen 10
> Add Device (0x0033) plen 7
> Status: Success (0x00)
> LE Address: CA:CA:BD:78:37:F9 (Static)
> < HCI Command: LE Add Device To Accept List (0x08|0x0011) plen 7
> Address type: Random (0x01)
> Address: CA:CA:BD:78:37:F9 (Static)
> @ MGMT Event: Device Removed (0x001b) plen 7
> LE Address: CA:CA:BD:78:37:F9 (Static)
> > HCI Event: Command Complete (0x0e) plen 4
> LE Add Device To Accept List (0x08|0x0011) ncmd 1
> Status: Success (0x00)
>
> [ 167.409813] Call trace:
> [ 167.409983] hci_le_add_accept_list_sync+0x64/0x26c
> [ 167.410150] hci_update_passive_scan_sync+0x5f0/0x6dc
> [ 167.410318] add_device_sync+0x18/0x24
> [ 167.410486] hci_cmd_sync_work+0xe8/0x150
> [ 167.410509] process_one_work+0x140/0x4d0
> [ 167.410526] worker_thread+0x134/0x2e4
> [ 167.410544] kthread+0x148/0x160
> [ 167.410562] ret_from_fork+0x10/0x30
>
> Signed-off-by: Archie Pusaka <apusaka@xxxxxxxxxxxx>
> Reviewed-by: Zhengping Jiang <jiangzp@xxxxxxxxxx>
> Reviewed-by: Michael Sun <michaelfsun@xxxxxxxxxx>
>
> ---
>
> net/bluetooth/hci_sync.c | 6 +++---
> 1 file changed, 3 insertions(+), 3 deletions(-)
>
> diff --git a/net/bluetooth/hci_sync.c b/net/bluetooth/hci_sync.c
> index 3067d94e7a8e..8e843d34f7de 100644
> --- a/net/bluetooth/hci_sync.c
> +++ b/net/bluetooth/hci_sync.c
> @@ -1863,7 +1863,7 @@ struct sk_buff *hci_read_local_oob_data_sync(struct hci_dev *hdev,
> */
> static u8 hci_update_accept_list_sync(struct hci_dev *hdev)
> {
> - struct hci_conn_params *params;
> + struct hci_conn_params *params, *tmp;
> struct bdaddr_list *b, *t;
> u8 num_entries = 0;
> bool pend_conn, pend_report;
> @@ -1930,7 +1930,7 @@ static u8 hci_update_accept_list_sync(struct hci_dev *hdev)
> * just abort and return filer policy value to not use the
> * accept list.
> */
> - list_for_each_entry(params, &hdev->pend_le_conns, action) {
> + list_for_each_entry_safe(params, tmp, &hdev->pend_le_conns, action) {
> err = hci_le_add_accept_list_sync(hdev, params, &num_entries);
> if (err)
> goto done;
> @@ -1940,7 +1940,7 @@ static u8 hci_update_accept_list_sync(struct hci_dev *hdev)
> * the list of pending reports and also add these to the
> * accept list if there is still space. Abort if space runs out.
> */
> - list_for_each_entry(params, &hdev->pend_le_reports, action) {
> + list_for_each_entry_safe(params, tmp, &hdev->pend_le_reports, action) {
> err = hci_le_add_accept_list_sync(hdev, params, &num_entries);
> if (err)
> goto done;
Hmm if this happens it means other threads are actually interfering
with cmd_sync queue which is something that is probably a bug since
the whole point of cmd_sync is to serialize the commands making it
easier to do more complex state updates (such accept+resolve list
updates), we could perhaps still apply this change as a workaround but
ultimately I think it would be better to add a mgmt-tester reproducing
the issue and have a proper fix of the code updating the list from a
different thread.
> --
> 2.37.1.359.gd136c6c3e2-goog
>
--
Luiz Augusto von Dentz
