[PATCH v1 05/10] Bluetooth: btmtksdio: move interrupt service to work

From: sean.wang
Date: Fri Oct 08 2021 - 18:22:06 EST


From: Sean Wang <sean.wang@xxxxxxxxxxxx>

btmtksdio belongs to WIFI/BT combo chip that would serve two radios in one
sdio_irq so that we have to move interrupt service to worker to ensure ISR
as short as possible.

The worker would serve the both Tx and Rx in a batch to effectively reduce
many interrupts to the host and to avoid excessive sdio lock contention
between various context (even from WiFi driver) and help to be more
efficient to complete command/event transation.

Co-developed-by: Mark-yw Chen <mark-yw.chen@xxxxxxxxxxxx>
Signed-off-by: Mark-yw Chen <mark-yw.chen@xxxxxxxxxxxx>
Signed-off-by: Sean Wang <sean.wang@xxxxxxxxxxxx>
---
drivers/bluetooth/btmtksdio.c | 132 ++++++++++++++++------------------
1 file changed, 63 insertions(+), 69 deletions(-)

diff --git a/drivers/bluetooth/btmtksdio.c b/drivers/bluetooth/btmtksdio.c
index 57126a95e292..f712b52e5797 100644
--- a/drivers/bluetooth/btmtksdio.c
+++ b/drivers/bluetooth/btmtksdio.c
@@ -98,7 +98,7 @@ struct btmtksdio_dev {
struct sdio_func *func;
struct device *dev;

- struct work_struct tx_work;
+ struct work_struct txrx_work;
unsigned long tx_state;
struct sk_buff_head txq;

@@ -249,32 +249,6 @@ static u32 btmtksdio_drv_own_query(struct btmtksdio_dev *bdev)
return sdio_readl(bdev->func, MTK_REG_CHLPCR, NULL);
}

-static void btmtksdio_tx_work(struct work_struct *work)
-{
- struct btmtksdio_dev *bdev = container_of(work, struct btmtksdio_dev,
- tx_work);
- struct sk_buff *skb;
- int err;
-
- pm_runtime_get_sync(bdev->dev);
-
- sdio_claim_host(bdev->func);
-
- while ((skb = skb_dequeue(&bdev->txq))) {
- err = btmtksdio_tx_packet(bdev, skb);
- if (err < 0) {
- bdev->hdev->stat.err_tx++;
- skb_queue_head(&bdev->txq, skb);
- break;
- }
- }
-
- sdio_release_host(bdev->func);
-
- pm_runtime_mark_last_busy(bdev->dev);
- pm_runtime_put_autosuspend(bdev->dev);
-}
-
static int btmtksdio_recv_event(struct hci_dev *hdev, struct sk_buff *skb)
{
struct btmtksdio_dev *bdev = hci_get_drvdata(hdev);
@@ -425,63 +399,81 @@ static int btmtksdio_rx_packet(struct btmtksdio_dev *bdev, u16 rx_size)
return err;
}

-static void btmtksdio_interrupt(struct sdio_func *func)
+static void btmtksdio_txrx_work(struct work_struct *work)
{
- struct btmtksdio_dev *bdev = sdio_get_drvdata(func);
+ struct btmtksdio_dev *bdev = container_of(work, struct btmtksdio_dev,
+ txrx_work);
+ unsigned long txrx_timeout;
+ struct sk_buff *skb;
u32 int_status;
u16 rx_size;
-
- /* It is required that the host gets ownership from the device before
- * accessing any register, however, if SDIO host is not being released,
- * a potential deadlock probably happens in a circular wait between SDIO
- * IRQ work and PM runtime work. So, we have to explicitly release SDIO
- * host here and claim again after the PM runtime work is all done.
- */
- sdio_release_host(bdev->func);
+ int err;

pm_runtime_get_sync(bdev->dev);

sdio_claim_host(bdev->func);

/* Disable interrupt */
- sdio_writel(func, C_INT_EN_CLR, MTK_REG_CHLPCR, NULL);
+ sdio_writel(bdev->func, C_INT_EN_CLR, MTK_REG_CHLPCR, 0);

- int_status = sdio_readl(func, MTK_REG_CHISR, NULL);
+ while ((skb = skb_dequeue(&bdev->txq))) {
+ err = btmtksdio_tx_packet(bdev, skb);
+ if (err < 0) {
+ bdev->hdev->stat.err_tx++;
+ skb_queue_head(&bdev->txq, skb);
+ break;
+ }
+ }

- /* Ack an interrupt as soon as possible before any operation on
- * hardware.
- *
- * Note that we don't ack any status during operations to avoid race
- * condition between the host and the device such as it's possible to
- * mistakenly ack RX_DONE for the next packet and then cause interrupts
- * not be raised again but there is still pending data in the hardware
- * FIFO.
- */
- sdio_writel(func, int_status, MTK_REG_CHISR, NULL);
+ txrx_timeout = jiffies + 5 * HZ;
+
+ do {
+ int_status = sdio_readl(bdev->func, MTK_REG_CHISR, NULL);
+
+ /* Ack an interrupt as soon as possible before any operation on
+ * hardware.
+ *
+ * Note that we don't ack any status during operations to avoid race
+ * condition between the host and the device such as it's possible to
+ * mistakenly ack RX_DONE for the next packet and then cause interrupts
+ * not be raised again but there is still pending data in the hardware
+ * FIFO.
+ */
+ sdio_writel(bdev->func, int_status, MTK_REG_CHISR, NULL);
+
+ if (int_status & FW_OWN_BACK_INT)
+ bt_dev_dbg(bdev->hdev, "Get fw own back");
+
+ if (int_status & TX_EMPTY)
+ schedule_work(&bdev->txrx_work);
+ else if (unlikely(int_status & TX_FIFO_OVERFLOW))
+ bt_dev_warn(bdev->hdev, "Tx fifo overflow");
+
+ if (int_status & RX_DONE_INT) {
+ rx_size = (int_status & RX_PKT_LEN) >> 16;
+ if (btmtksdio_rx_packet(bdev, rx_size) < 0)
+ bdev->hdev->stat.err_rx++;
+ }

- if (unlikely(!int_status))
- bt_dev_err(bdev->hdev, "CHISR is 0");
+ } while (int_status || time_is_before_jiffies(txrx_timeout));

- if (int_status & FW_OWN_BACK_INT)
- bt_dev_dbg(bdev->hdev, "Get fw own back");
+ /* Enable interrupt */
+ sdio_writel(bdev->func, C_INT_EN_SET, MTK_REG_CHLPCR, 0);

- if (int_status & TX_EMPTY)
- schedule_work(&bdev->tx_work);
- else if (unlikely(int_status & TX_FIFO_OVERFLOW))
- bt_dev_warn(bdev->hdev, "Tx fifo overflow");
+ sdio_release_host(bdev->func);

- if (int_status & RX_DONE_INT) {
- rx_size = (int_status & RX_PKT_LEN) >> 16;
+ pm_runtime_mark_last_busy(bdev->dev);
+ pm_runtime_put_autosuspend(bdev->dev);
+}

- if (btmtksdio_rx_packet(bdev, rx_size) < 0)
- bdev->hdev->stat.err_rx++;
- }
+static void btmtksdio_interrupt(struct sdio_func *func)
+{
+ struct btmtksdio_dev *bdev = sdio_get_drvdata(func);

- /* Enable interrupt */
- sdio_writel(func, C_INT_EN_SET, MTK_REG_CHLPCR, NULL);
+ /* Disable interrupt */
+ sdio_writel(bdev->func, C_INT_EN_CLR, MTK_REG_CHLPCR, 0);

- pm_runtime_mark_last_busy(bdev->dev);
- pm_runtime_put_autosuspend(bdev->dev);
+ schedule_work(&bdev->txrx_work);
}

static int btmtksdio_open(struct hci_dev *hdev)
@@ -583,6 +575,8 @@ static int btmtksdio_close(struct hci_dev *hdev)

sdio_release_irq(bdev->func);

+ cancel_work_sync(&bdev->txrx_work);
+
/* Return ownership to the device */
sdio_writel(bdev->func, C_FW_OWN_REQ_SET, MTK_REG_CHLPCR, NULL);

@@ -604,7 +598,7 @@ static int btmtksdio_flush(struct hci_dev *hdev)

skb_queue_purge(&bdev->txq);

- cancel_work_sync(&bdev->tx_work);
+ cancel_work_sync(&bdev->txrx_work);

return 0;
}
@@ -795,7 +789,7 @@ static int btmtksdio_send_frame(struct hci_dev *hdev, struct sk_buff *skb)

skb_queue_tail(&bdev->txq, skb);

- schedule_work(&bdev->tx_work);
+ schedule_work(&bdev->txrx_work);

return 0;
}
@@ -818,7 +812,7 @@ static int btmtksdio_probe(struct sdio_func *func,
bdev->dev = &func->dev;
bdev->func = func;

- INIT_WORK(&bdev->tx_work, btmtksdio_tx_work);
+ INIT_WORK(&bdev->txrx_work, btmtksdio_txrx_work);
skb_queue_head_init(&bdev->txq);

/* Initialize and register HCI device */
--
2.25.1