2024-05-14 11:28:47

by Vitor Soares

[permalink] [raw]
Subject: [PATCH v5] can: mcp251xfd: fix infinite loop when xmit fails

From: Vitor Soares <[email protected]>

When the mcp251xfd_start_xmit() function fails, the driver stops
processing messages, and the interrupt routine does not return,
running indefinitely even after killing the running application.

Error messages:
[ 441.298819] mcp251xfd spi2.0 can0: ERROR in mcp251xfd_start_xmit: -16
[ 441.306498] mcp251xfd spi2.0 can0: Transmit Event FIFO buffer not empty. (seq=0x000017c7, tef_tail=0x000017cf, tef_head=0x000017d0, tx_head=0x000017d3).
.. and repeat forever.

The issue can be triggered when multiple devices share the same
SPI interface. And there is concurrent access to the bus.

The problem occurs because tx_ring->head increments even if
mcp251xfd_start_xmit() fails. Consequently, the driver skips one
TX package while still expecting a response in
mcp251xfd_handle_tefif_one().

This patch resolves the issue by starting a workqueue to write
the tx obj synchronously if err = -EBUSY. In case of another error,
it decrements tx_ring->head, removes skb from the echo stack, and
drops the message.

Fixes: 55e5b97f003e ("can: mcp25xxfd: add driver for Microchip MCP25xxFD SPI CAN")
Cc: [email protected]
Signed-off-by: Vitor Soares <[email protected]>
---

V4->V5:
- Start a workqueue to write tx obj with spi_sync() when spi_async() == -EBUSY.

V3->V4:
- Leave can_put_echo_skb() and stop the queue if needed, before mcp251xfd_tx_obj_write().
- Re-sync head and remove echo skb if mcp251xfd_tx_obj_write() fails.
- Revert -> return NETDEV_TX_BUSY if mcp251xfd_tx_obj_write() == -EBUSY.

V2->V3:
- Add tx_dropped stats.
- netdev_sent_queue() only if can_put_echo_skb() succeed.

V1->V2:
- Return NETDEV_TX_BUSY if mcp251xfd_tx_obj_write() == -EBUSY.
- Rework the commit message to address the change above.
- Change can_put_echo_skb() to be called after mcp251xfd_tx_obj_write() succeed.
Otherwise, we get Kernel NULL pointer dereference error.

.../net/can/spi/mcp251xfd/mcp251xfd-core.c | 13 ++++-
drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c | 51 ++++++++++++++++---
drivers/net/can/spi/mcp251xfd/mcp251xfd.h | 5 ++
3 files changed, 60 insertions(+), 9 deletions(-)

diff --git a/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c b/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c
index 1d9057dc44f2..6cca853f2b1e 100644
--- a/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c
+++ b/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c
@@ -2141,15 +2141,25 @@ static int mcp251xfd_probe(struct spi_device *spi)
if (err)
goto out_free_candev;

+ priv->tx_work_obj = NULL;
+ priv->wq = alloc_workqueue("mcp251xfd_wq", WQ_FREEZABLE, 0);
+ if (!priv->wq) {
+ err = -ENOMEM;
+ goto out_can_rx_offload_del;
+ }
+ INIT_WORK(&priv->tx_work, mcp251xfd_tx_obj_write_sync);
+
err = mcp251xfd_register(priv);
if (err) {
dev_err_probe(&spi->dev, err, "Failed to detect %s.\n",
mcp251xfd_get_model_str(priv));
- goto out_can_rx_offload_del;
+ goto out_can_free_wq;
}

return 0;

+ out_can_free_wq:
+ destroy_workqueue(priv->wq);
out_can_rx_offload_del:
can_rx_offload_del(&priv->offload);
out_free_candev:
@@ -2165,6 +2175,7 @@ static void mcp251xfd_remove(struct spi_device *spi)
struct mcp251xfd_priv *priv = spi_get_drvdata(spi);
struct net_device *ndev = priv->ndev;

+ destroy_workqueue(priv->wq);
can_rx_offload_del(&priv->offload);
mcp251xfd_unregister(priv);
spi->max_speed_hz = priv->spi_max_speed_hz_orig;
diff --git a/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c b/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c
index 160528d3cc26..1e7ddf316643 100644
--- a/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c
+++ b/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c
@@ -131,6 +131,41 @@ mcp251xfd_tx_obj_from_skb(const struct mcp251xfd_priv *priv,
tx_obj->xfer[0].len = len;
}

+static void mcp251xfd_tx_failure_drop(const struct mcp251xfd_priv *priv,
+ struct mcp251xfd_tx_ring *tx_ring,
+ int err)
+{
+ struct net_device *ndev = priv->ndev;
+ struct net_device_stats *stats = &ndev->stats;
+ unsigned int frame_len = 0;
+ u8 tx_head;
+
+ tx_ring->head--;
+ stats->tx_dropped++;
+ tx_head = mcp251xfd_get_tx_head(tx_ring);
+ can_free_echo_skb(ndev, tx_head, &frame_len);
+ netdev_completed_queue(ndev, 1, frame_len);
+ netif_wake_queue(ndev);
+
+ if (net_ratelimit())
+ netdev_err(priv->ndev, "ERROR in %s: %d\n", __func__, err);
+}
+
+void mcp251xfd_tx_obj_write_sync(struct work_struct *ws)
+{
+ struct mcp251xfd_priv *priv = container_of(ws, struct mcp251xfd_priv,
+ tx_work);
+ struct mcp251xfd_tx_obj *tx_obj = priv->tx_work_obj;
+ struct mcp251xfd_tx_ring *tx_ring = priv->tx;
+ int err;
+
+ err = spi_sync(priv->spi, &tx_obj->msg);
+ if (err)
+ mcp251xfd_tx_failure_drop(priv, tx_ring, err);
+
+ priv->tx_work_obj = NULL;
+}
+
static int mcp251xfd_tx_obj_write(const struct mcp251xfd_priv *priv,
struct mcp251xfd_tx_obj *tx_obj)
{
@@ -175,7 +210,7 @@ netdev_tx_t mcp251xfd_start_xmit(struct sk_buff *skb,
if (can_dev_dropped_skb(ndev, skb))
return NETDEV_TX_OK;

- if (mcp251xfd_tx_busy(priv, tx_ring))
+ if (mcp251xfd_tx_busy(priv, tx_ring) || priv->tx_work_obj)
return NETDEV_TX_BUSY;

tx_obj = mcp251xfd_get_tx_obj_next(tx_ring);
@@ -193,13 +228,13 @@ netdev_tx_t mcp251xfd_start_xmit(struct sk_buff *skb,
netdev_sent_queue(priv->ndev, frame_len);

err = mcp251xfd_tx_obj_write(priv, tx_obj);
- if (err)
- goto out_err;
-
- return NETDEV_TX_OK;
-
- out_err:
- netdev_err(priv->ndev, "ERROR in %s: %d\n", __func__, err);
+ if (err == -EBUSY) {
+ priv->tx_work_obj = tx_obj;
+ netif_stop_queue(ndev);
+ queue_work(priv->wq, &priv->tx_work);
+ } else if (err) {
+ mcp251xfd_tx_failure_drop(priv, tx_ring, err);
+ }

return NETDEV_TX_OK;
}
diff --git a/drivers/net/can/spi/mcp251xfd/mcp251xfd.h b/drivers/net/can/spi/mcp251xfd/mcp251xfd.h
index 24510b3b8020..4e27a33f4030 100644
--- a/drivers/net/can/spi/mcp251xfd/mcp251xfd.h
+++ b/drivers/net/can/spi/mcp251xfd/mcp251xfd.h
@@ -633,6 +633,10 @@ struct mcp251xfd_priv {
struct mcp251xfd_rx_ring *rx[MCP251XFD_FIFO_RX_NUM];
struct mcp251xfd_tx_ring tx[MCP251XFD_FIFO_TX_NUM];

+ struct workqueue_struct *wq;
+ struct work_struct tx_work;
+ struct mcp251xfd_tx_obj *tx_work_obj;
+
DECLARE_BITMAP(flags, __MCP251XFD_FLAGS_SIZE__);

u8 rx_ring_num;
@@ -952,6 +956,7 @@ void mcp251xfd_skb_set_timestamp(const struct mcp251xfd_priv *priv,
void mcp251xfd_timestamp_init(struct mcp251xfd_priv *priv);
void mcp251xfd_timestamp_stop(struct mcp251xfd_priv *priv);

+void mcp251xfd_tx_obj_write_sync(struct work_struct *ws);
netdev_tx_t mcp251xfd_start_xmit(struct sk_buff *skb,
struct net_device *ndev);

--
2.34.1



2024-05-14 12:24:22

by Marc Kleine-Budde

[permalink] [raw]
Subject: Re: [PATCH v5] can: mcp251xfd: fix infinite loop when xmit fails

On 14.05.2024 11:58:22, Vitor Soares wrote:
> From: Vitor Soares <[email protected]>
>
> When the mcp251xfd_start_xmit() function fails, the driver stops
> processing messages, and the interrupt routine does not return,
> running indefinitely even after killing the running application.
>
> Error messages:
> [ 441.298819] mcp251xfd spi2.0 can0: ERROR in mcp251xfd_start_xmit: -16
> [ 441.306498] mcp251xfd spi2.0 can0: Transmit Event FIFO buffer not empty. (seq=0x000017c7, tef_tail=0x000017cf, tef_head=0x000017d0, tx_head=0x000017d3).
> ... and repeat forever.
>
> The issue can be triggered when multiple devices share the same
> SPI interface. And there is concurrent access to the bus.
>
> The problem occurs because tx_ring->head increments even if
> mcp251xfd_start_xmit() fails. Consequently, the driver skips one
> TX package while still expecting a response in
> mcp251xfd_handle_tefif_one().
>
> This patch resolves the issue by starting a workqueue to write
> the tx obj synchronously if err = -EBUSY. In case of another error,
> it decrements tx_ring->head, removes skb from the echo stack, and
> drops the message.

This looks quite good! Good work!

I think you better move the allocation/destroy of the wq into the open()
and stop() callbacks. You have to destroy the workqueue before putting
the interface to sleep.

>
> Fixes: 55e5b97f003e ("can: mcp25xxfd: add driver for Microchip MCP25xxFD SPI CAN")
> Cc: [email protected]
> Signed-off-by: Vitor Soares <[email protected]>
> ---
>
> V4->V5:
> - Start a workqueue to write tx obj with spi_sync() when spi_async() == -EBUSY.
>
> V3->V4:
> - Leave can_put_echo_skb() and stop the queue if needed, before mcp251xfd_tx_obj_write().
> - Re-sync head and remove echo skb if mcp251xfd_tx_obj_write() fails.
> - Revert -> return NETDEV_TX_BUSY if mcp251xfd_tx_obj_write() == -EBUSY.
>
> V2->V3:
> - Add tx_dropped stats.
> - netdev_sent_queue() only if can_put_echo_skb() succeed.
>
> V1->V2:
> - Return NETDEV_TX_BUSY if mcp251xfd_tx_obj_write() == -EBUSY.
> - Rework the commit message to address the change above.
> - Change can_put_echo_skb() to be called after mcp251xfd_tx_obj_write() succeed.
> Otherwise, we get Kernel NULL pointer dereference error.
>
> .../net/can/spi/mcp251xfd/mcp251xfd-core.c | 13 ++++-
> drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c | 51 ++++++++++++++++---
> drivers/net/can/spi/mcp251xfd/mcp251xfd.h | 5 ++
> 3 files changed, 60 insertions(+), 9 deletions(-)
>
> diff --git a/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c b/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c
> index 1d9057dc44f2..6cca853f2b1e 100644
> --- a/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c
> +++ b/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c
> @@ -2141,15 +2141,25 @@ static int mcp251xfd_probe(struct spi_device *spi)
> if (err)
> goto out_free_candev;
>
> + priv->tx_work_obj = NULL;
> + priv->wq = alloc_workqueue("mcp251xfd_wq", WQ_FREEZABLE, 0);

The m_can driver uses a ordered workqueue and you can add the name of
the spi device to the wq's name :)

priv->wq = alloc_ordered_workqueue("%s-mcp251xfd_wq", WQ_FREEZABLE | WQ_MEM_RECLAIM, dev_name(&spi->dev));

> + if (!priv->wq) {
> + err = -ENOMEM;
> + goto out_can_rx_offload_del;
> + }
> + INIT_WORK(&priv->tx_work, mcp251xfd_tx_obj_write_sync);
> +
> err = mcp251xfd_register(priv);
> if (err) {
> dev_err_probe(&spi->dev, err, "Failed to detect %s.\n",
> mcp251xfd_get_model_str(priv));
> - goto out_can_rx_offload_del;
> + goto out_can_free_wq;

nitpick:
out_destroy_workqueue;

to match the function call.

> }
>
> return 0;
>
> + out_can_free_wq:
> + destroy_workqueue(priv->wq);
> out_can_rx_offload_del:
> can_rx_offload_del(&priv->offload);
> out_free_candev:
> @@ -2165,6 +2175,7 @@ static void mcp251xfd_remove(struct spi_device *spi)
> struct mcp251xfd_priv *priv = spi_get_drvdata(spi);
> struct net_device *ndev = priv->ndev;
>
> + destroy_workqueue(priv->wq);
> can_rx_offload_del(&priv->offload);
> mcp251xfd_unregister(priv);
> spi->max_speed_hz = priv->spi_max_speed_hz_orig;
> diff --git a/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c b/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c
> index 160528d3cc26..1e7ddf316643 100644
> --- a/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c
> +++ b/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c
> @@ -131,6 +131,41 @@ mcp251xfd_tx_obj_from_skb(const struct mcp251xfd_priv *priv,
> tx_obj->xfer[0].len = len;
> }
>
> +static void mcp251xfd_tx_failure_drop(const struct mcp251xfd_priv *priv,
> + struct mcp251xfd_tx_ring *tx_ring,
> + int err)
> +{
> + struct net_device *ndev = priv->ndev;
> + struct net_device_stats *stats = &ndev->stats;
> + unsigned int frame_len = 0;
> + u8 tx_head;
> +
> + tx_ring->head--;
> + stats->tx_dropped++;
> + tx_head = mcp251xfd_get_tx_head(tx_ring);
> + can_free_echo_skb(ndev, tx_head, &frame_len);
> + netdev_completed_queue(ndev, 1, frame_len);
> + netif_wake_queue(ndev);
> +
> + if (net_ratelimit())
> + netdev_err(priv->ndev, "ERROR in %s: %d\n", __func__, err);
> +}
> +
> +void mcp251xfd_tx_obj_write_sync(struct work_struct *ws)
> +{
> + struct mcp251xfd_priv *priv = container_of(ws, struct mcp251xfd_priv,
> + tx_work);
> + struct mcp251xfd_tx_obj *tx_obj = priv->tx_work_obj;
> + struct mcp251xfd_tx_ring *tx_ring = priv->tx;
> + int err;
> +
> + err = spi_sync(priv->spi, &tx_obj->msg);
> + if (err)
> + mcp251xfd_tx_failure_drop(priv, tx_ring, err);
> +
> + priv->tx_work_obj = NULL;

Race condition:
- after spi_sync() the CAN frame is send
- after the TX complete IRQ the TX queue is restarted
- the xmit handler might get BUSY
- fill the tx_work_obj again

> +}
> +
> static int mcp251xfd_tx_obj_write(const struct mcp251xfd_priv *priv,
> struct mcp251xfd_tx_obj *tx_obj)
> {
> @@ -175,7 +210,7 @@ netdev_tx_t mcp251xfd_start_xmit(struct sk_buff *skb,
> if (can_dev_dropped_skb(ndev, skb))
> return NETDEV_TX_OK;
>
> - if (mcp251xfd_tx_busy(priv, tx_ring))
> + if (mcp251xfd_tx_busy(priv, tx_ring) || priv->tx_work_obj)

This should not happen, but better save than sorry.

> return NETDEV_TX_BUSY;
>
> tx_obj = mcp251xfd_get_tx_obj_next(tx_ring);
> @@ -193,13 +228,13 @@ netdev_tx_t mcp251xfd_start_xmit(struct sk_buff *skb,
> netdev_sent_queue(priv->ndev, frame_len);
>
> err = mcp251xfd_tx_obj_write(priv, tx_obj);
> - if (err)
> - goto out_err;
> -
> - return NETDEV_TX_OK;
> -
> - out_err:
> - netdev_err(priv->ndev, "ERROR in %s: %d\n", __func__, err);
> + if (err == -EBUSY) {
> + priv->tx_work_obj = tx_obj;
> + netif_stop_queue(ndev);
> + queue_work(priv->wq, &priv->tx_work);

nitpick: I would do "netif_stop_queue(ndev);" first.

My mental idea of the code flow is:
- stop the queue
- do everything to start the workqueue.

> + } else if (err) {
> + mcp251xfd_tx_failure_drop(priv, tx_ring, err);
> + }
>
> return NETDEV_TX_OK;
> }
> diff --git a/drivers/net/can/spi/mcp251xfd/mcp251xfd.h b/drivers/net/can/spi/mcp251xfd/mcp251xfd.h
> index 24510b3b8020..4e27a33f4030 100644
> --- a/drivers/net/can/spi/mcp251xfd/mcp251xfd.h
> +++ b/drivers/net/can/spi/mcp251xfd/mcp251xfd.h
> @@ -633,6 +633,10 @@ struct mcp251xfd_priv {
> struct mcp251xfd_rx_ring *rx[MCP251XFD_FIFO_RX_NUM];
> struct mcp251xfd_tx_ring tx[MCP251XFD_FIFO_TX_NUM];
>
> + struct workqueue_struct *wq;
> + struct work_struct tx_work;
> + struct mcp251xfd_tx_obj *tx_work_obj;
> +
> DECLARE_BITMAP(flags, __MCP251XFD_FLAGS_SIZE__);
>
> u8 rx_ring_num;
> @@ -952,6 +956,7 @@ void mcp251xfd_skb_set_timestamp(const struct mcp251xfd_priv *priv,
> void mcp251xfd_timestamp_init(struct mcp251xfd_priv *priv);
> void mcp251xfd_timestamp_stop(struct mcp251xfd_priv *priv);
>
> +void mcp251xfd_tx_obj_write_sync(struct work_struct *ws);
> netdev_tx_t mcp251xfd_start_xmit(struct sk_buff *skb,
> struct net_device *ndev);
>
> --
> 2.34.1
>
>

regards,
Marc

--
Pengutronix e.K. | Marc Kleine-Budde |
Embedded Linux | https://www.pengutronix.de |
Vertretung Nürnberg | Phone: +49-5121-206917-129 |
Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-9 |


Attachments:
(No filename) (8.32 kB)
signature.asc (499.00 B)
Download all attachments

2024-05-14 14:34:55

by Vitor Soares

[permalink] [raw]
Subject: Re: [PATCH v5] can: mcp251xfd: fix infinite loop when xmit fails

Hi Marc,

Appreciate your feedback.

On Tue, 2024-05-14 at 14:23 +0200, Marc Kleine-Budde wrote:
> On 14.05.2024 11:58:22, Vitor Soares wrote:
> > From: Vitor Soares <[email protected]>
> >
> > When the mcp251xfd_start_xmit() function fails, the driver stops
> > processing messages, and the interrupt routine does not return,
> > running indefinitely even after killing the running application.
> >
> > Error messages:
> > [  441.298819] mcp251xfd spi2.0 can0: ERROR in mcp251xfd_start_xmit: -16
> > [  441.306498] mcp251xfd spi2.0 can0: Transmit Event FIFO buffer not empty.
> > (seq=0x000017c7, tef_tail=0x000017cf, tef_head=0x000017d0,
> > tx_head=0x000017d3).
> > ... and repeat forever.
> >
> > The issue can be triggered when multiple devices share the same
> > SPI interface. And there is concurrent access to the bus.
> >
> > The problem occurs because tx_ring->head increments even if
> > mcp251xfd_start_xmit() fails. Consequently, the driver skips one
> > TX package while still expecting a response in
> > mcp251xfd_handle_tefif_one().
> >
> > This patch resolves the issue by starting a workqueue to write
> > the tx obj synchronously if err = -EBUSY. In case of another error,
> > it decrements tx_ring->head, removes skb from the echo stack, and
> > drops the message.
>
> This looks quite good! Good work!
>
> I think you better move the allocation/destroy of the wq into the open()
> and stop() callbacks. You have to destroy the workqueue before putting
> the interface to sleep.
>
> >
> > Fixes: 55e5b97f003e ("can: mcp25xxfd: add driver for Microchip MCP25xxFD SPI
> > CAN")
> > Cc: [email protected]
> > Signed-off-by: Vitor Soares <[email protected]>
> > ---
> >
> > V4->V5:
> >   - Start a workqueue to write tx obj with spi_sync() when spi_async() == -
> > EBUSY.
> >
> > V3->V4:
> >   - Leave can_put_echo_skb() and stop the queue if needed, before
> > mcp251xfd_tx_obj_write().
> >   - Re-sync head and remove echo skb if mcp251xfd_tx_obj_write() fails.
> >   - Revert -> return NETDEV_TX_BUSY if mcp251xfd_tx_obj_write() == -EBUSY.
> >
> > V2->V3:
> >   - Add tx_dropped stats.
> >   - netdev_sent_queue() only if can_put_echo_skb() succeed.
> >
> > V1->V2:
> >   - Return NETDEV_TX_BUSY if mcp251xfd_tx_obj_write() == -EBUSY.
> >   - Rework the commit message to address the change above.
> >   - Change can_put_echo_skb() to be called after mcp251xfd_tx_obj_write()
> > succeed.
> >     Otherwise, we get Kernel NULL pointer dereference error.
> >
> >  .../net/can/spi/mcp251xfd/mcp251xfd-core.c    | 13 ++++-
> >  drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c  | 51 ++++++++++++++++---
> >  drivers/net/can/spi/mcp251xfd/mcp251xfd.h     |  5 ++
> >  3 files changed, 60 insertions(+), 9 deletions(-)
> >
> > diff --git a/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c
> > b/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c
> > index 1d9057dc44f2..6cca853f2b1e 100644
> > --- a/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c
> > +++ b/drivers/net/can/spi/mcp251xfd/mcp251xfd-core.c
> > @@ -2141,15 +2141,25 @@ static int mcp251xfd_probe(struct spi_device *spi)
> >         if (err)
> >                 goto out_free_candev;
> >  
> > +       priv->tx_work_obj = NULL;
> > +       priv->wq = alloc_workqueue("mcp251xfd_wq", WQ_FREEZABLE, 0);
>
> The m_can driver uses a ordered workqueue and you can add the name of
> the spi device to the wq's name :)
>
>         priv->wq = alloc_ordered_workqueue("%s-mcp251xfd_wq", WQ_FREEZABLE |
> WQ_MEM_RECLAIM, dev_name(&spi->dev));
>
> > +       if (!priv->wq) {
> > +               err = -ENOMEM;
> > +               goto out_can_rx_offload_del;
> > +       }
> > +       INIT_WORK(&priv->tx_work, mcp251xfd_tx_obj_write_sync);
> > +
> >         err = mcp251xfd_register(priv);
> >         if (err) {
> >                 dev_err_probe(&spi->dev, err, "Failed to detect %s.\n",
> >                               mcp251xfd_get_model_str(priv));
> > -               goto out_can_rx_offload_del;
> > +               goto out_can_free_wq;
>
> nitpick:
>                      out_destroy_workqueue;
>
> to match the function call.
>
> >         }
> >  
> >         return 0;
> >  
> > + out_can_free_wq:
> > +       destroy_workqueue(priv->wq);
> >   out_can_rx_offload_del:
> >         can_rx_offload_del(&priv->offload);
> >   out_free_candev:
> > @@ -2165,6 +2175,7 @@ static void mcp251xfd_remove(struct spi_device *spi)
> >         struct mcp251xfd_priv *priv = spi_get_drvdata(spi);
> >         struct net_device *ndev = priv->ndev;
> >  
> > +       destroy_workqueue(priv->wq);
> >         can_rx_offload_del(&priv->offload);
> >         mcp251xfd_unregister(priv);
> >         spi->max_speed_hz = priv->spi_max_speed_hz_orig;
> > diff --git a/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c
> > b/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c
> > index 160528d3cc26..1e7ddf316643 100644
> > --- a/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c
> > +++ b/drivers/net/can/spi/mcp251xfd/mcp251xfd-tx.c
> > @@ -131,6 +131,41 @@ mcp251xfd_tx_obj_from_skb(const struct mcp251xfd_priv
> > *priv,
> >         tx_obj->xfer[0].len = len;
> >  }
> >  
> > +static void mcp251xfd_tx_failure_drop(const struct mcp251xfd_priv *priv,
> > +                                     struct mcp251xfd_tx_ring *tx_ring,
> > +                                     int err)
> > +{
> > +       struct net_device *ndev = priv->ndev;
> > +       struct net_device_stats *stats = &ndev->stats;
> > +       unsigned int frame_len = 0;
> > +       u8 tx_head;
> > +
> > +       tx_ring->head--;
> > +       stats->tx_dropped++;
> > +       tx_head = mcp251xfd_get_tx_head(tx_ring);
> > +       can_free_echo_skb(ndev, tx_head, &frame_len);
> > +       netdev_completed_queue(ndev, 1, frame_len);
> > +       netif_wake_queue(ndev);
> > +
> > +       if (net_ratelimit())
> > +               netdev_err(priv->ndev, "ERROR in %s: %d\n", __func__, err);
> > +}
> > +
> > +void mcp251xfd_tx_obj_write_sync(struct work_struct *ws)
> > +{
> > +       struct mcp251xfd_priv *priv = container_of(ws, struct
> > mcp251xfd_priv,
> > +                                                  tx_work);
> > +       struct mcp251xfd_tx_obj *tx_obj = priv->tx_work_obj;
> > +       struct mcp251xfd_tx_ring *tx_ring = priv->tx;
> > +       int err;
> > +
> > +       err = spi_sync(priv->spi, &tx_obj->msg);
> > +       if (err)
> > +               mcp251xfd_tx_failure_drop(priv, tx_ring, err);
> > +
> > +       priv->tx_work_obj = NULL;
>
> Race condition:
> - after spi_sync() the CAN frame is send
> - after the TX complete IRQ the TX queue is restarted
> - the xmit handler might get BUSY
> - fill the tx_work_obj again
>
> > +}
> > +
> >  static int mcp251xfd_tx_obj_write(const struct mcp251xfd_priv *priv,
> >                                   struct mcp251xfd_tx_obj *tx_obj)
> >  {
> > @@ -175,7 +210,7 @@ netdev_tx_t mcp251xfd_start_xmit(struct sk_buff *skb,
> >         if (can_dev_dropped_skb(ndev, skb))
> >                 return NETDEV_TX_OK;
> >  
> > -       if (mcp251xfd_tx_busy(priv, tx_ring))
> > +       if (mcp251xfd_tx_busy(priv, tx_ring) || priv->tx_work_obj)
>
> This should not happen, but better save than sorry.


As there is the race condition you mentioned above, on this condition:
priv->tx_work_obj = tx_obj --> xmit will return NETDEV_TX_BUSY

or

priv->tx_work_obj = NULL --> It goes through the rest of the code or the
workqueue may sleep after setting tx_work_obj to NULL. Should I use work_busy()
here instead or do you have another suggestion?


Everything else is clear to me and I'm addressing it for the v6.

Best regards,
Vitor Soares

2024-05-15 11:13:03

by Marc Kleine-Budde

[permalink] [raw]
Subject: Re: [PATCH v5] can: mcp251xfd: fix infinite loop when xmit fails

On 14.05.2024 15:34:01, Vitor Soares wrote:
> > > +void mcp251xfd_tx_obj_write_sync(struct work_struct *ws)
> > > +{
> > > + struct mcp251xfd_priv *priv = container_of(ws, struct
> > > mcp251xfd_priv,
> > > + tx_work);
> > > + struct mcp251xfd_tx_obj *tx_obj = priv->tx_work_obj;
> > > + struct mcp251xfd_tx_ring *tx_ring = priv->tx;
> > > + int err;
> > > +
> > > + err = spi_sync(priv->spi, &tx_obj->msg);
> > > + if (err)
> > > + mcp251xfd_tx_failure_drop(priv, tx_ring, err);
> > > +
> > > + priv->tx_work_obj = NULL;
> >
> > Race condition:
> > - after spi_sync() the CAN frame is send
> > - after the TX complete IRQ the TX queue is restarted
> > - the xmit handler might get BUSY
> > - fill the tx_work_obj again

You can avoid the race condition by moving "priv->tx_work_obj = NULL;"
in front of the "spi_sync();". Right?

> > > +}
> > > +
> > > static int mcp251xfd_tx_obj_write(const struct mcp251xfd_priv *priv,
> > > struct mcp251xfd_tx_obj *tx_obj)
> > > {
> > > @@ -175,7 +210,7 @@ netdev_tx_t mcp251xfd_start_xmit(struct sk_buff *skb,
> > > if (can_dev_dropped_skb(ndev, skb))
> > > return NETDEV_TX_OK;
> > >
> > > - if (mcp251xfd_tx_busy(priv, tx_ring))
> > > + if (mcp251xfd_tx_busy(priv, tx_ring) || priv->tx_work_obj)
> >
> > This should not happen, but better save than sorry.
>
> As there is the race condition you mentioned above, on this condition:
> priv->tx_work_obj = tx_obj --> xmit will return NETDEV_TX_BUSY
>
> or
>
> priv->tx_work_obj = NULL --> It goes through the rest of the code or
> the workqueue may sleep after setting tx_work_obj to NULL. Should I
> use work_busy() here instead or do you have another suggestion?

Yes, introduce mcp251xfd_work_busy().

I'm not sure what happens if the xmit is called between the
"priv->tx_work_obj = NULL" and the end of the work. Will queue_work()
return false, as the queue is still running?

> Everything else is clear to me and I'm addressing it for the v6.

Marc

--
Pengutronix e.K. | Marc Kleine-Budde |
Embedded Linux | https://www.pengutronix.de |
Vertretung Nürnberg | Phone: +49-5121-206917-129 |
Amtsgericht Hildesheim, HRA 2686 | Fax: +49-5121-206917-9 |


Attachments:
(No filename) (2.39 kB)
signature.asc (499.00 B)
Download all attachments

2024-05-15 16:29:56

by Vitor Soares

[permalink] [raw]
Subject: Re: [PATCH v5] can: mcp251xfd: fix infinite loop when xmit fails

On Wed, 2024-05-15 at 13:12 +0200, Marc Kleine-Budde wrote:
> On 14.05.2024 15:34:01, Vitor Soares wrote:
> > > > +void mcp251xfd_tx_obj_write_sync(struct work_struct *ws)
> > > > +{
> > > > +       struct mcp251xfd_priv *priv = container_of(ws, struct
> > > > mcp251xfd_priv,
> > > > +                                                  tx_work);
> > > > +       struct mcp251xfd_tx_obj *tx_obj = priv->tx_work_obj;
> > > > +       struct mcp251xfd_tx_ring *tx_ring = priv->tx;
> > > > +       int err;
> > > > +
> > > > +       err = spi_sync(priv->spi, &tx_obj->msg);
> > > > +       if (err)
> > > > +               mcp251xfd_tx_failure_drop(priv, tx_ring, err);
> > > > +
> > > > +       priv->tx_work_obj = NULL;
> > >
> > > Race condition:
> > > - after spi_sync() the CAN frame is send
> > > - after the TX complete IRQ the TX queue is restarted
> > > - the xmit handler might get BUSY
> > > - fill the tx_work_obj again
>
> You can avoid the race condition by moving "priv->tx_work_obj = NULL;"
> in front of the "spi_sync();". Right?
>
> > > > +}
> > > > +
> > > >  static int mcp251xfd_tx_obj_write(const struct mcp251xfd_priv *priv,
> > > >                                   struct mcp251xfd_tx_obj *tx_obj)
> > > >  {
> > > > @@ -175,7 +210,7 @@ netdev_tx_t mcp251xfd_start_xmit(struct sk_buff
> > > > *skb,
> > > >         if (can_dev_dropped_skb(ndev, skb))
> > > >                 return NETDEV_TX_OK;
> > > >  
> > > > -       if (mcp251xfd_tx_busy(priv, tx_ring))
> > > > +       if (mcp251xfd_tx_busy(priv, tx_ring) || priv->tx_work_obj)
> > >
> > > This should not happen, but better save than sorry.
> >
> > As there is the race condition you mentioned above, on this condition:
> > priv->tx_work_obj = tx_obj --> xmit will return NETDEV_TX_BUSY
> >
> > or
> >
> > priv->tx_work_obj = NULL --> It goes through the rest of the code or
> > the workqueue may sleep after setting tx_work_obj to NULL. Should I
> > use work_busy() here instead or do you have another suggestion?
>
> Yes, introduce mcp251xfd_work_busy().
>

I'll implement it.

> I'm not sure what happens if the xmit is called between the
> "priv->tx_work_obj = NULL" and the end of the work. Will queue_work()
> return false, as the queue is still running?

From the test I did so far, my understanding is the following:

If mcp251xfd_tx_obj_write doesn't fail, everything is OK.

if mcp251xfd_tx_obj_write fails with EBUSY
- stop netif queue
- fill the tx_work_obj
- start worker

queue_work() doesn't return false even when work_busy() = true.
- xmit handler return, and wait netif_wake_queue()
- the work handler waits until the previous job gets done before starting the
next one.
- after the TX completes IRQ, the TX queue is restarted

If the TX queue is restarted immediately after queue_work(), tx_work_obj is
filled, making the xmit handler return NETDEV_TX_BUSY.

The tests were done with a delay after priv->tx_work_obj = NULL.


Best regards,
Vitor Soares