Re: [PATCH v7 1/4] can: m_can: Create a m_can platform framework

[Wolfgang Grandegger]

Hello Dan,

Am 08.03.19 um 18:25 schrieb Dan Murphy:
> On 3/8/19 11:08 AM, Wolfgang Grandegger wrote:
>> Hello,
>>
>> Am 08.03.19 um 16:48 schrieb Dan Murphy:
>>> Wolfgang
>>>
>>> On 3/8/19 8:41 AM, Wolfgang Grandegger wrote:
>>>> Hello Dan,
>>>>
>>>> thinking more about it...
>>>>
>>>> Am 08.03.19 um 14:29 schrieb Wolfgang Grandegger:
>>>>> Hello Dan,
>>>>>
>>>>> Am 08.03.19 um 13:44 schrieb Dan Murphy:
>>>>>> Wolfgang
>>>>>>
>>>>>> On 3/8/19 4:10 AM, Wolfgang Grandegger wrote:
>>>>>>> Hallo Dan,
>>>>>>>
>>>>>>> Am 05.03.19 um 16:52 schrieb Dan Murphy:
>>>>>>>> Create a m_can platform framework that peripherial
>>>>>>>> devices can register to and use common code and register sets.
>>>>>>>> The peripherial devices may provide read/write and configuration
>>>>>>>> support of the IP.
>>>>>>>>
>>>>>>>> Signed-off-by: Dan Murphy <dmurphy@xxxxxx>
>>>>>>>> ---
>>>>>>>>
>>>>>>>>
>>>>>>>> v7 - Fixed remaining new checkpatch issues, removed CSR setting, fixed tx hard
>>>>>>>> start function to return tx_busy, and renamed device callbacks - https://lore.kernel.org/patchwork/patch/1047220/
>>>>>>>>
>>>>>>>> v6 - Squashed platform patch to this patch for bissectablity, fixed coding style
>>>>>>>> issues, updated Kconfig help, placed mcan reg offsets back into c file, renamed
>>>>>>>> priv->skb to priv->tx_skb and cleared perp interrupts at ISR start -
>>>>>>>> Patch 1 comments - https://lore.kernel.org/patchwork/patch/1042446/
>>>>>>>> Patch 2 comments - https://lore.kernel.org/patchwork/patch/1042442/
>>>>>>>>
>>>>>>>>  drivers/net/can/m_can/Kconfig          |  13 +-
>>>>>>>>  drivers/net/can/m_can/Makefile         |   1 +
>>>>>>>>  drivers/net/can/m_can/m_can.c          | 700 +++++++++++++------------
>>>>>>>>  drivers/net/can/m_can/m_can.h          | 110 ++++
>>>>>>>>  drivers/net/can/m_can/m_can_platform.c | 202 +++++++
>>>>>>>>  5 files changed, 682 insertions(+), 344 deletions(-)
>>>>>>>>  create mode 100644 drivers/net/can/m_can/m_can.h
>>>>>>>>  create mode 100644 drivers/net/can/m_can/m_can_platform.c
>>>>>>>>
>>>>>>>> diff --git a/drivers/net/can/m_can/Kconfig b/drivers/net/can/m_can/Kconfig
>>>>>>>> index 04f20dd39007..f7119fd72df4 100644
>>>>>>>> --- a/drivers/net/can/m_can/Kconfig
>>>>>>>> +++ b/drivers/net/can/m_can/Kconfig
>>>>>>>> @@ -1,5 +1,14 @@
>>>>>>>>  config CAN_M_CAN
>>>>>>>> +	tristate "Bosch M_CAN support"
>>>>>>>> +	---help---
>>>>>>>> +	  Say Y here if you want support for Bosch M_CAN controller framework.
>>>>>>>> +	  This is common support for devices that embed the Bosch M_CAN IP.
>>>>>>>> +
>>>>>>>> +config CAN_M_CAN_PLATFORM
>>>>>>>> +	tristate "Bosch M_CAN support for io-mapped devices"
>>>>>>>>  	depends on HAS_IOMEM
>>>>>>>> -	tristate "Bosch M_CAN devices"
>>>>>>>> +	depends on CAN_M_CAN
>>>>>>>>  	---help---
>>>>>>>> -	  Say Y here if you want to support for Bosch M_CAN controller.
>>>>>>>> +	  Say Y here if you want support for IO Mapped Bosch M_CAN controller.
>>>>>>>> +	  This support is for devices that have the Bosch M_CAN controller
>>>>>>>> +	  IP embedded into the device and the IP is IO Mapped to the processor.
>>>>>>>> diff --git a/drivers/net/can/m_can/Makefile b/drivers/net/can/m_can/Makefile
>>>>>>>> index 8bbd7f24f5be..057bbcdb3c74 100644
>>>>>>>> --- a/drivers/net/can/m_can/Makefile
>>>>>>>> +++ b/drivers/net/can/m_can/Makefile
>>>>>>>> @@ -3,3 +3,4 @@
>>>>>>>>  #
>>>>>>>>  
>>>>>>>>  obj-$(CONFIG_CAN_M_CAN) += m_can.o
>>>>>>>> +obj-$(CONFIG_CAN_M_CAN_PLATFORM) += m_can_platform.o
>>>>>>>> diff --git a/drivers/net/can/m_can/m_can.c b/drivers/net/can/m_can/m_can.c
>>>>>>>> index 9b449400376b..a60278d94126 100644
>>>>>>>> --- a/drivers/net/can/m_can/m_can.c
>>>>>>>> +++ b/drivers/net/can/m_can/m_can.c
>>>>>>>
>>>>>>> ... snip...
>>>>>>>
>>>>>>>> +static netdev_tx_t m_can_start_xmit(struct sk_buff *skb,
>>>>>>>> +				    struct net_device *dev)
>>>>>>>> +{
>>>>>>>> +	struct m_can_priv *priv = netdev_priv(dev);
>>>>>>>> +
>>>>>>>> +	if (can_dropped_invalid_skb(dev, skb))
>>>>>>>> +		return NETDEV_TX_OK;
>>>>>>>> +
>>>>>>>> +	if (priv->is_peripherial) {
>>>>>>>> +		if (priv->tx_skb) {
>>>>>>>> +			netdev_err(dev, "hard_xmit called while tx busy\n");
>>>>>>>> +			return NETDEV_TX_BUSY;
>>>>>>>> +		}
>>>>>>>
>>>>>>> The problem with that approach is, that the upper layer will try to
>>>>>>> resubmit the current "skb" but not the previous "tx_skb". And the
>>>>>>> previous "tx_skb" has not been freed yet. I would just drop and free the
>>>>>>> skb and return NETDEV_TX_OK in m_can_tx_handler() for peripheral devices
>>>>>>> (like can_dropped_invalid_skb() does).
>>>>>>>
>>>>>>
>>>>>> OK.
>>>>>>
>>>>>> So would this also be a bug in the hi3110 and mcp251x drivers (line 521) as well because besides checking tx_length
>>>>>> this is how these drivers are written.
>>>>>
>>>>> This is different. When entering the "start_xmit" routine, the previous
>>>>> TX is still in progress. It will (hopefully) complete soon. Therefore
>>>>> returning NETDEV_TX_BUSY is OK. The "start_xmit" routine will be
>>>>> recalled soon with the same "skb". That scenario should/could also not
>>>>> happen.
>>>>
>>>> In principle, this also applies to the m_can peripheral devices. If
>>>> tx_skb is not NULL, the TX is still in progress and returning
>>>> NETDEV_TX_BUSY is just fine.
>>>>
>>>>>
>>>>> In contrast, in "m_can_tx_handler()", the skb could not be handled
>>>>> because the FIFO is full. The "start_xmit" routine for peripheral
>>>>> devices for that skb already returned NETDEV_TX_OK. Therefore the only
>>>>> meaningful action is to drop the skb. Also this error should not happen
>>>>> and if, something is going really wrong. Therefore I think, a
>>>>> WARN_ONCE() would be even more appropriate. But that should be a
>>>>> separate patch.
>>>>
>>>> But that's a different issue/error. The tx_skb cannot be processed in
>>>> "m_can_tx_handler()". Either we drop it or we re-queue it (retry later).
>>>>
>>>
>>> OK I am a bit confused on this.  Are you saying this is not an issue?
>>> Or are you saying I need to check for tx_len like the other code?
>>
>> If you check for tx_skb in the "start_xmit" routine like the hi3110 and
>> mcp251x, it will work the same way. But only, if the "tx_handler()" has
>> fully processed the message. It simple means, the TX is still in
>> progress and will complete soon. But in "m_can_tx_handler()" we return
>> without handling the message! It will never be sent and freed. Or will
>> the "m_can_tx_handler()" retry?
>>
> 
> I am not seeing where we are not handling the message in the m_can_tx_handler()

static void m_can_tx_handler(struct m_can_classdev *priv)
{
		...
		/* Check if FIFO full */
		if (m_can_tx_fifo_full(priv)) {
			/* This shouldn't happen */
			netif_stop_queue(dev);
			netdev_warn(dev,
				    "TX queue active although FIFO is full.");
			return;
		}

We simply return here. When is the message (tx_skb) processed (sent or freed)?
What happens with tx_skb?


For the hi3110, we have:

static void hi3110_tx_work_handler(struct work_struct *ws)
{
        struct hi3110_priv *priv = container_of(ws, struct hi3110_priv,
                                                tx_work);
        struct spi_device *spi = priv->spi;
        struct net_device *net = priv->net;
        struct can_frame *frame;

        mutex_lock(&priv->hi3110_lock);
        if (priv->tx_skb) {
                if (priv->can.state == CAN_STATE_BUS_OFF) {
                        hi3110_clean(net);
                } else {
                        frame = (struct can_frame *)priv->tx_skb->data;
                        hi3110_hw_tx(spi, frame);
                        priv->tx_len = 1 + frame->can_dlc;
                        can_put_echo_skb(priv->tx_skb, net, 0);
                        priv->tx_skb = NULL;
                }
        }
        mutex_unlock(&priv->hi3110_lock);
}

Either the tx_skb is sent or cleanup (dropped and freed) in case of bus-off.
Also "hi3110_clean" sets "priv->tx_skb = NULL"! The "tx_len" handles a pending 
"echo_skb".

> 
> In the peripheral code the work is queued up.  And the work thread is m_can_tx_work_queue.
> 
> This in turn calls the m_can_tx_handler and the worker is blocked until return which means the message
> would have been processed.
> 
> If there is no issue and the handler returns OK then the skb is set to null.
> Otherwise the only other time that the skb will not be null is if the FIFO was full.
> 
> Plus there can only be one work queue at a time so the processing is synchronous.
> If the upper layer decides to send another packet before the prior one is complete then it will get
> a TX busy return.
> 
> IOmapped calls are blocked on return so this is not an issue.  We cannot do it the same way with peripherals due to the
> atomic context of the request.
Wolfgang.