[PATCH v3] Bluetooth: hci_serdev: stop write_work before closing the port

From: Zhao Li

Date: Sat Nov 21 2026 - 10:57:24 EST


With HCI_QUIRK_NON_PERSISTENT_SETUP, hci_uart_close() closes the serdev
port without keeping write_work off. The hci_uart_flush() call right
before closing re-enables write_work and may replay a recorded wakeup,
and hci_uart_tx_wakeup() queues it from any caller that passed the
HCI_UART_PROTO_READY test, such as the serdev write_wakeup callback.
write_work can then call ttyport_write_buf() while ttyport_close()
releases the tty, so it writes through a released tty.

Move the flush into __hci_uart_flush() with a resume_tx argument.
hci_uart_close() passes false when it is about to close the port, so
write_work stays disabled and HCI_UART_SENDING stays set after the
flush. It then clears HCI_UART_PROTO_READY under percpu_down_write(),
which waits for any hci_uart_tx_wakeup() already past its
HCI_UART_PROTO_READY test, and closes the port. Only then does it clear
HCI_UART_TX_WAKEUP, re-enable write_work and clear HCI_UART_SENDING: the
next open reuses hu and does not reset tx_state, and a stale SENDING
would stop every later wakeup from queuing write_work.

Found by code review. Compile-tested with W=1.

Fixes: de8892df72be ("Bluetooth: hci_serdev: Close UART port if NON_PERSISTENT_SETUP is set")
Suggested-by: Loic Poulain <loic.poulain@xxxxxxxxxxxxxxxx>
Cc: stable@xxxxxxxxxxxxxxx
Signed-off-by: Zhao Li <enderaoelyther@xxxxxxxxx>
---
v3:
- Add __hci_uart_flush() with a resume_tx argument and call it from
hci_uart_close() instead of repeating the flush. write_work is
re-enabled only after serdev_device_close() (Loic Poulain's review of
v2).
v2:
- Clear HCI_UART_SENDING and HCI_UART_TX_WAKEUP after cancelling
write_work. A pending write_work cancelled there left SENDING set, so
the next power-on never queued write_work (Sashiko's review of v1).

v2: https://lore.kernel.org/linux-bluetooth/20261120062048.40310-1-enderaoelyther@xxxxxxxxx/
v1: https://lore.kernel.org/linux-bluetooth/20261118124821.3990-1-enderaoelyther@xxxxxxxxx/

Merging with Michael McAllister's hci_serdev v3 series (Patchwork
14989810-14989814):
https://lore.kernel.org/linux-bluetooth/179484638195.54009.16928291772826938293.wcn3950-poweroff-v3-1@xxxxxxxxxxxxxxxxxxxxx/
- 1/5 adds "if (hci_dev_test_flag(hdev, HCI_UNREGISTER)) return 0;" at the
top of hci_uart_flush() and hci_uart_close(). The flush check has to
stay in hci_uart_flush(), not move into __hci_uart_flush().
- Once hci_uart_close() is past its own check, __hci_uart_flush() must
always run disable_work_sync(), which the enable_work() after
serdev_device_close() balances. Otherwise write_work is not disabled
during the close and the enable underflows the disable count.
- 2/5: use hci_uart_serdev_close(hu) in place of
serdev_device_close(hu->serdev) in the close_port branch.

This needs ee43bade37c4, which no stable branch has yet, and
disable_work_sync() from v6.10. With the flush body of ee43bade37c4
backported it applies as is to 7.3.y and 6.18.y and, with the older
test_bit(HCI_QUIRK_NON_PERSISTENT_SETUP, &hdev->quirks) context line, to
6.12.y; 6.6.y has no disable_work_sync().

No new W=1, sparse or smatch warnings.

AI assistance was used.

drivers/bluetooth/hci_serdev.c | 26 +++++++++++++++++++++++---
1 file changed, 23 insertions(+), 3 deletions(-)

diff --git a/drivers/bluetooth/hci_serdev.c b/drivers/bluetooth/hci_serdev.c
index 194f11251cc5..9f038ab65a73 100644
--- a/drivers/bluetooth/hci_serdev.c
+++ b/drivers/bluetooth/hci_serdev.c
@@ -91,7 +91,7 @@ static void hci_uart_write_work(struct work_struct *work)
/* ------- Interface to HCI layer ------ */

/* Reset device */
-static int hci_uart_flush(struct hci_dev *hdev)
+static void __hci_uart_flush(struct hci_dev *hdev, bool resume_tx)
{
struct hci_uart *hu = hci_get_drvdata(hdev);

@@ -112,11 +112,19 @@ static int hci_uart_flush(struct hci_dev *hdev)
hu->proto->flush(hu);
}

+ if (!resume_tx)
+ return;
+
/* Resume TX and replay a wakeup recorded while it was stopped. */
enable_work(&hu->write_work);
clear_bit(HCI_UART_SENDING, &hu->tx_state);
if (test_bit(HCI_UART_TX_WAKEUP, &hu->tx_state))
hci_uart_tx_wakeup(hu);
+}
+
+static int hci_uart_flush(struct hci_dev *hdev)
+{
+ __hci_uart_flush(hdev, true);

return 0;
}
@@ -150,22 +158,34 @@ static int hci_uart_open(struct hci_dev *hdev)
static int hci_uart_close(struct hci_dev *hdev)
{
struct hci_uart *hu = hci_get_drvdata(hdev);
+ bool close_port;

BT_DBG("hdev %p", hdev);

if (!test_bit(HCI_UART_PROTO_READY, &hu->flags))
return 0;

- hci_uart_flush(hdev);
+ close_port = hci_test_quirk(hdev, HCI_QUIRK_NON_PERSISTENT_SETUP);
+
+ /* Keep write_work disabled until the port is closed. */
+ __hci_uart_flush(hdev, !close_port);
hdev->flush = NULL;

/* When QUIRK HCI_QUIRK_NON_PERSISTENT_SETUP is set by driver,
* BT SOC is completely powered OFF during BT OFF, holding port
* open may drain the battery.
*/
- if (hci_test_quirk(hdev, HCI_QUIRK_NON_PERSISTENT_SETUP)) {
+ if (close_port) {
+ /* Drain wakeups that passed the PROTO_READY test. */
+ percpu_down_write(&hu->proto_lock);
clear_bit(HCI_UART_PROTO_READY, &hu->flags);
+ percpu_up_write(&hu->proto_lock);
+
serdev_device_close(hu->serdev);
+
+ clear_bit(HCI_UART_TX_WAKEUP, &hu->tx_state);
+ enable_work(&hu->write_work);
+ clear_bit(HCI_UART_SENDING, &hu->tx_state);
}

return 0;

base-commit: cf357cfc025e7f72851479ed3fb60d5f16413c1b
--
2.56.1