Hi Zhengping,

On Sun, Feb 26, 2023 at 11:18 PM Zhengping Jiang [off-list ref] wrote:

Hi Luiz,

I have a question. Given that each command in the cmd_sync queue
should clean up the memory in a callback function. I was wondering if
the call to cmd_complete_rsp in __mgmt_power_off function is still
necessary? Will this always risk a race condition that cmd has been
released when the complete callback or _sync function is run?

Not sure I follow you here, do you have a stack trace when the user
after free occurs?

Thanks,
Zhengping

On Fri, Feb 24, 2023 at 2:37 PM Zhengping Jiang [off-list ref] wrote:

quoted

Hi Luiz,

quoted

Any particular reason why you are not using hci_cmd_sync_clear
instead?

That is a good question and we used hci_cmd_sync_clear in the first
version, but it will clear the queue and also close the timer. As a
result, when the adapter is turned on again, the timer will not
schedule any new jobs. So the option is to use hci_cmd_sync_clear and
re-initiate the queue or to write a new function which only clears the
queue.

quoted

We also may want to move the clearing logic to
hci_dev_close_sync since it should be equivalent to
hci_request_cancel_all.

I actually have a question here. I saw
"drain_workqueue(hdev->workqueue)" in hci_dev_close_sync and thought
it should force clearing the cmd_sync queue. But it seems cannot
prevent the use-after-free situation.

Any suggestions to improve the solution?

Thanks,
Zhengping


On Fri, Feb 24, 2023 at 1:02 PM Luiz Augusto von Dentz
[off-list ref] wrote:

quoted

Hi Zhengping,

On Fri, Feb 24, 2023 at 11:53 AM Zhengping Jiang [off-list ref] wrote:

quoted

Clear cmd_sync_work queue before clearing the mgmt cmd list to avoid
racing conditions which cause use-after-free.

When powering off the adapter, the mgmt cmd list will be cleared. If a
work is queued in the cmd_sync_work queue at the same time, it will
cause the risk of use-after-free, as the cmd pointer is not checked
before use.

Signed-off-by: Zhengping Jiang <redacted>
---

Changes in v2:
- Add function to clear the queue without stop the timer

Changes in v1:
- Clear cmd_sync_work queue before clearing the mgmt cmd list

 net/bluetooth/hci_sync.c | 21 ++++++++++++++++++++-
 1 file changed, 20 insertions(+), 1 deletion(-)

diff --git a/net/bluetooth/hci_sync.c b/net/bluetooth/hci_sync.c
index 117eedb6f709..b70365dfff0c 100644
--- a/net/bluetooth/hci_sync.c
+++ b/net/bluetooth/hci_sync.c

@@ -636,6 +636,23 @@ void hci_cmd_sync_init(struct hci_dev *hdev)
        INIT_DELAYED_WORK(&hdev->adv_instance_expire, adv_timeout_expire);
 }

+static void hci_pend_cmd_sync_clear(struct hci_dev *hdev)
+{
+       struct hci_cmd_sync_work_entry *entry, *tmp;
+
+       mutex_lock(&hdev->cmd_sync_work_lock);
+       list_for_each_entry_safe(entry, tmp, &hdev->cmd_sync_work_list, list) {
+               if (entry->destroy) {
+                       hci_req_sync_lock(hdev);
+                       entry->destroy(hdev, entry->data, -ECANCELED);
+                       hci_req_sync_unlock(hdev);
+               }
+               list_del(&entry->list);
+               kfree(entry);
+       }
+       mutex_unlock(&hdev->cmd_sync_work_lock);
+}
+
 void hci_cmd_sync_clear(struct hci_dev *hdev)
 {
        struct hci_cmd_sync_work_entry *entry, *tmp;

@@ -4842,8 +4859,10 @@ int hci_dev_close_sync(struct hci_dev *hdev)

        if (!auto_off && hdev->dev_type == HCI_PRIMARY &&
            !hci_dev_test_flag(hdev, HCI_USER_CHANNEL) &&
-           hci_dev_test_flag(hdev, HCI_MGMT))
+           hci_dev_test_flag(hdev, HCI_MGMT)) {
+               hci_pend_cmd_sync_clear(hdev);

Any particular reason why you are not using hci_cmd_sync_clear
instead? We also may want to move the clearing logic to
hci_dev_close_sync since it should be equivalent to
hci_request_cancel_all.

quoted

                __mgmt_power_off(hdev);
+       }

        hci_inquiry_cache_flush(hdev);
        hci_pend_le_actions_clear(hdev);
--
2.39.2.722.g9855ee24e9-goog

--
Luiz Augusto von Dentz

-- 
Luiz Augusto von Dentz