Bluetooth: btmtksdio: fix deadlock in close and reset paths

btmtksdio_close() and btmtksdio_reset() call cancel_work_sync() on
bdev->txrx_work while holding the sdio host lock, which is also acquired
by btmtksdio_txrx_work().  If txrx_work is queued when close/reset runs,
a worker thread may start it after the host lock is taken and block in
sdio_claim_host(), while cancel_work_sync() waits for the work to
finish.  The host lock is only released after cancel_work_sync()
returns, so both sides wait forever, deadlocking close/reset.

Fix this by releasing the sdio host lock before calling
cancel_work_sync(), then re-acquiring it afterwards.

In btmtksdio_close() the interrupt is already disabled by
sdio_release_irq(), which also unregisters the IRQ handler, so no new
work can be scheduled and cancel_work_sync() fully quiesces txrx_work.

btmtksdio_reset() must additionally unregister the IRQ handler before
dropping the host lock: btmtksdio_txrx_work() unconditionally re-enables
the device interrupt (C_INT_EN_SET) when the handler is still registered,
so an in-flight worker would re-enable interrupts and be rescheduled
while the device is being reset, defeating the cancellation.  The IRQ is
re-claimed by btmtksdio_open() when the HCI device is re-opened after
the reset.

This mirrors the pattern already used by btmtksdio_flush(), which
cancels the work without holding the host lock.

Signed-off-by: ZhaoJinming <zhaojinming@uniontech.com>
Signed-off-by: Luiz Augusto von Dentz <luiz.von.dentz@intel.com>

1 files changed, 22 insertions, 0 deletions

diff --git a/drivers/bluetooth/btmtksdio.c b/drivers/bluetooth/btmtksdio.c
index 4e1012e90979..67d055d5f197 100644
--- a/drivers/bluetooth/btmtksdio.c
+++ b/drivers/bluetooth/btmtksdio.c
@@ -746,8 +746,16 @@ static int btmtksdio_close(struct hci_dev *hdev)
 
 	sdio_release_irq(bdev->func);
 
+	/* No new work can be scheduled after sdio_release_irq(), so cancel the
+	 * work outside the sdio host lock. btmtksdio_txrx_work() also claims
+	 * the host, so canceling it while holding the lock would deadlock.
+	 */
+	sdio_release_host(bdev->func);
+
 	cancel_work_sync(&bdev->txrx_work);
 
+	sdio_claim_host(bdev->func);
+
 	btmtksdio_fw_pmctrl(bdev);
 
 	clear_bit(BTMTKSDIO_FUNC_ENABLED, &bdev->tx_state);
@@ -1293,8 +1301,22 @@ static void btmtksdio_reset(struct hci_dev *hdev)
 
 	sdio_writel(bdev->func, C_INT_EN_CLR, MTK_REG_CHLPCR, NULL);
 	skb_queue_purge(&bdev->txq);
+
+	/* Unregister the IRQ before releasing the host lock so that a
+	 * concurrently running btmtksdio_txrx_work() cannot re-enable the
+	 * device interrupt (C_INT_EN_SET) and be rescheduled while the device
+	 * is being reset. btmtksdio_txrx_work() also claims the host, so the
+	 * work must be cancelled outside the sdio host lock to avoid a
+	 * deadlock. The IRQ is re-claimed by btmtksdio_open() when the HCI
+	 * device is re-opened after the reset.
+	 */
+	sdio_release_irq(bdev->func);
+	sdio_release_host(bdev->func);
+
 	cancel_work_sync(&bdev->txrx_work);
 
+	sdio_claim_host(bdev->func);
+
 	gpiod_set_value_cansleep(bdev->reset, 1);
 	msleep(100);
 	gpiod_set_value_cansleep(bdev->reset, 0);