Re: [PATCH V2] firmware: arm_scmi: Make scmi core independent of transport type

[Arnd Bergmann <arnd@arndb.de>]

On Fri, Jan 10, 2020 at 10:43 AM Viresh Kumar <viresh.kumar@linaro.org> wrote:
>
> The SCMI specification is fairly independent of the transport protocol,
> which can be a simple mailbox (already implemented) or anything else.
> The current Linux implementation however is very much dependent of the
> mailbox transport layer.
>
> This patch makes the SCMI core code (driver.c) independent of the
> mailbox transport layer and moves all mailbox related code to a new
> file: mailbox.c.
>
> We can now implement more transport protocols to transport SCMI
> messages, some of the transport protocols getting discussed currently
> are SMC/HVC, SPCI (built on top of SMC/HVC), OPTEE based mailbox
> (similar to SPCI), and vitio based transport as alternative to mailbox.
>
> The transport protocols just need to provide struct scmi_desc, which
> also implements the struct scmi_transport_ops.
>
> Signed-off-by: Viresh Kumar <viresh.kumar@linaro.org>
> ---
> V2:
> - Dropped __iomem from payload data.

Simply dropping the __iomem isn't much better, now you get other
type mismatches.

> - Moved transport ops to scmi_desc, and that has a per transport
>   instance now which is differentiated using the compatible string.
> - Converted IS_ERR_OR_NULL to IS_ERR.

These look good to me.

> + * @payload: Transmit/Receive payload area
> + * @dev: Reference to device in the SCMI hierarchy corresponding to this
> + *      channel
> + * @handle: Pointer to SCMI entity handle
> + * @transport_info: Transport layer related information
> + */
> +struct scmi_chan_info {
> +       void *payload;
> +       struct device *dev;
> +       struct scmi_handle *handle;
> +       void *transport_info;
> +};

Maybe you can wrap the scmi_chan_info inside of another
structure that contains  the payload pointer, and use container_of
to convert between them?

It's not obvious which parts of the structure should be shared and
which are transport specific.

> -static void scmi_rx_callback(struct mbox_client *cl, void *m)
> +void scmi_rx_callback(struct scmi_chan_info *cinfo, struct scmi_xfer *t)
>  {
>         u8 msg_type;
>         u32 msg_hdr;
>         u16 xfer_id;
>         struct scmi_xfer *xfer;
> -       struct scmi_chan_info *cinfo = client_to_scmi_chan_info(cl);
>         struct device *dev = cinfo->dev;
>         struct scmi_info *info = handle_to_scmi_info(cinfo->handle);
>         struct scmi_xfers_info *minfo = &info->tx_minfo;
> -       struct scmi_shared_mem __iomem *mem = cinfo->payload;
> +       struct scmi_shared_mem *mem = cinfo->payload;
>
>         msg_hdr = ioread32(&mem->msg_header);

This is where it goes wrong: you cannot pass a kernel pointer
without __iomem into ioread32(). Building the driver with sparse
(using "make C=1") should show you this and possibly other
related conversion bugs.

       Arnd