Re: [PATCH v2 01/15] drm/dsi: Make mipi_dsi_dcs_{read, write}() symmetrical

[Andrzej Hajda <a.hajda@samsung.com>]

On 10/14/2014 04:16 PM, Thierry Reding wrote:
> On Tue, Oct 14, 2014 at 03:53:26PM +0200, Andrzej Hajda wrote:
>> On 10/14/2014 01:29 PM, Thierry Reding wrote:
>>> On Tue, Oct 14, 2014 at 01:25:44PM +0200, Andrzej Hajda wrote:
>>>> On 10/14/2014 12:57 PM, Thierry Reding wrote:
>>>>> On Tue, Oct 14, 2014 at 12:38:15PM +0200, Andrzej Hajda wrote:
>>>>>> On 10/14/2014 11:44 AM, Thierry Reding wrote:
>>>>>>> On Tue, Oct 14, 2014 at 10:00:32AM +0200, Andrzej Hajda wrote:
>>>>>>>> On 10/13/2014 12:16 PM, Thierry Reding wrote:
>>>>>>>>> From: Thierry Reding <treding@nvidia.com>
>>>>>>>>>
>>>>>>>>> Currently the mipi_dsi_dcs_write() function requires the DCS command
>>>>>>>>> byte to be embedded within the write buffer whereas mipi_dsi_dcs_read()
>>>>>>>>> has a separate parameter. Make them more symmetrical by adding an extra
>>>>>>>>> command parameter to mipi_dsi_dcs_write().
>>>>>>>>>
>>>>>>>>> The S6E8AA0 driver relies on the old asymmetric API and there's concern
>>>>>>>>> that moving to the new API may be less efficient. Provide a new function
>>>>>>>>> with the old semantics for those cases and make the S6E8AA0 driver use
>>>>>>>>> it instead.
>>>>>>>>>
>>>>>>>>> Signed-off-by: Thierry Reding <treding@nvidia.com>
>>>>>>>>> ---
>>>>>>>>> Changes in v2:
>>>>>>>>> - provide mipi_dsi_dcs_write_buffer() for backwards compatibility
>>>>>>>>>
>>>>>>>>>  drivers/gpu/drm/drm_mipi_dsi.c        | 127 +++++++++++++++++++++++++++++-----
>>>>>>>>>  drivers/gpu/drm/panel/panel-s6e8aa0.c |   2 +-
>>>>>>>>>  include/drm/drm_mipi_dsi.h            |   6 +-
>>>>>>>>>  3 files changed, 114 insertions(+), 21 deletions(-)
>>>>>>>>>
>>>>>>>>> diff --git a/drivers/gpu/drm/drm_mipi_dsi.c b/drivers/gpu/drm/drm_mipi_dsi.c
>>>>>>>>> index eb6dfe52cab2..1702ffd07986 100644
>>>>>>>>> --- a/drivers/gpu/drm/drm_mipi_dsi.c
>>>>>>>>> +++ b/drivers/gpu/drm/drm_mipi_dsi.c
>>>>>>>>> @@ -199,33 +199,120 @@ int mipi_dsi_detach(struct mipi_dsi_device *dsi)
>>>>>>>>>  EXPORT_SYMBOL(mipi_dsi_detach);
>>>>>>>>>  
>>>>>>>>>  /**
>>>>>>>>> - * mipi_dsi_dcs_write - send DCS write command
>>>>>>>>> - * @dsi: DSI device
>>>>>>>>> - * @data: pointer to the command followed by parameters
>>>>>>>>> - * @len: length of @data
>>>>>>>>> + * mipi_dsi_dcs_write_buffer() - transmit a DCS command with payload
>>>>>>>>> + * @dsi: DSI peripheral device
>>>>>>>>> + * @data: buffer containing data to be transmitted
>>>>>>>>> + * @len: size of transmission buffer
>>>>>>>>> + *
>>>>>>>>> + * This function will automatically choose the right data type depending on
>>>>>>>>> + * the command payload length.
>>>>>>>>> + *
>>>>>>>>> + * Return: The number of bytes successfully transmitted or a negative error
>>>>>>>>> + * code on failure.
>>>>>>>>>   */
>>>>>>>>> -ssize_t mipi_dsi_dcs_write(struct mipi_dsi_device *dsi, const void *data,
>>>>>>>>> -			    size_t len)
>>>>>>>>> +ssize_t mipi_dsi_dcs_write_buffer(struct mipi_dsi_device *dsi,
>>>>>>>>> +				  const void *data, size_t len)
>>>>>>>>>  {
>>>>>>>>> -	const struct mipi_dsi_host_ops *ops = dsi->host->ops;
>>>>>>>>>  	struct mipi_dsi_msg msg = {
>>>>>>>>>  		.channel = dsi->channel,
>>>>>>>>>  		.tx_buf = data,
>>>>>>>>>  		.tx_len = len
>>>>>>>>>  	};
>>>>>>>>>  
>>>>>>>>> -	if (!ops || !ops->transfer)
>>>>>>>>> +	if (!dsi->host->ops || !dsi->host->ops->transfer)
>>>>>>>>>  		return -ENOSYS;
>>>>>>>>>  
>>>>>>>>>  	switch (len) {
>>>>>>>>>  	case 0:
>>>>>>>>>  		return -EINVAL;
>>>>>>>>> +
>>>>>>>>>  	case 1:
>>>>>>>>>  		msg.type = MIPI_DSI_DCS_SHORT_WRITE;
>>>>>>>>>  		break;
>>>>>>>>> +
>>>>>>>>>  	case 2:
>>>>>>>>>  		msg.type = MIPI_DSI_DCS_SHORT_WRITE_PARAM;
>>>>>>>>>  		break;
>>>>>>>>> +
>>>>>>>>> +	default:
>>>>>>>>> +		msg.type = MIPI_DSI_DCS_LONG_WRITE;
>>>>>>>>> +		break;
>>>>>>>>> +	}
>>>>>>>>> +
>>>>>>>>> +	return dsi->host->ops->transfer(dsi->host, &msg);
>>>>>>>>> +}
>>>>>>>>> +EXPORT_SYMBOL(mipi_dsi_dcs_write_buffer);
>>>>>>>>> +
>>>>>>>>> +/**
>>>>>>>>> + * mipi_dsi_dcs_write() - send DCS write command
>>>>>>>>> + * @dsi: DSI peripheral device
>>>>>>>>> + * @cmd: DCS command
>>>>>>>>> + * @data: buffer containing the command payload
>>>>>>>>> + * @len: command payload length
>>>>>>>>> + *
>>>>>>>>> + * This function will automatically choose the right data type depending on
>>>>>>>>> + * the command payload length.
>>>>>>>>> + *
>>>>>>>>> + * Return: The number of bytes successfully transmitted or a negative error
>>>>>>>>> + * code on failure.
>>>>>>>>> + */
>>>>>>>>> +ssize_t mipi_dsi_dcs_write(struct mipi_dsi_device *dsi, u8 cmd,
>>>>>>>>> +			   const void *data, size_t len)
>>>>>>>>> +{
>>>>>>>>> +	struct mipi_dsi_msg msg;
>>>>>>>>> +	ssize_t err;
>>>>>>>>> +	size_t size;
>>>>>>>>> +	u8 *tx;
>>>>>>>>> +
>>>>>>>>> +	if (!dsi->host->ops || !dsi->host->ops->transfer)
>>>>>>>>> +		return -ENOSYS;
>>>>>>>>> +
>>>>>>>>> +	if (len > 0) {
>>>>>>>>> +		unsigned int offset = 0;
>>>>>>>>> +
>>>>>>>>> +		/*
>>>>>>>>> +		 * DCS long write packets contain the word count in the header
>>>>>>>>> +		 * bytes 1 and 2 and have a payload containing the DCS command
>>>>>>>>> +		 * byte folowed by word count minus one bytes.
>>>>>>>>> +		 *
>>>>>>>>> +		 * DCS short write packets encode the DCS command and up to
>>>>>>>>> +		 * one parameter in header bytes 1 and 2.
>>>>>>>>> +		 */
>>>>>>>>> +		if (len > 1)
>>>>>>>>> +			size = 3 + len;
>>>>>>>>> +		else
>>>>>>>>> +			size = 1 + len;
>>>>>>>> I guess "size = 2" would be better here.
>>>>>>> This is on purpose because it documents the format. If len > 1, then the
>>>>>>> packet is a long write, so we need three bytes (command & word count) in
>>>>>>> addition to the payload. For short writes, len <= 1 and we only need one
>>>>>>> extra byte (command).
>>>>>>>
>>>>>>>>> +
>>>>>>>>> +		tx = kmalloc(size, GFP_KERNEL);
>>>>>>>>> +		if (!tx)
>>>>>>>>> +			return -ENOMEM;
>>>>>>>>> +
>>>>>>>>> +		/* write word count to header for DCS long write packets */
>>>>>>>>> +		if (len > 1) {
>>>>>>>>> +			tx[offset++] = ((1 + len) >> 0) & 0xff;
>>>>>>>>> +			tx[offset++] = ((1 + len) >> 8) & 0xff;
>>>>>>>>> +		}
>>>>>>>>> +
>>>>>>>>> +		/* write the DCS command byte followed by the payload */
>>>>>>>>> +		tx[offset++] = cmd;
>>>>>>>>> +		memcpy(tx + offset, data, len);
>>>>>>>>> +	} else {
>>>>>>>>> +		tx = &cmd;
>>>>>>>>> +		size = 1;
>>>>>>>>> +	}
>>>>>>>> Contents of this conditional is incompatible with the current API.
>>>>>>>> mipi_dsi_msg.tx_buf contains only data and mipi_dsi_msg.tx_len contains
>>>>>>>> lenght of this data. Now you try to embed length of data into tx_buf and
>>>>>>>> this breaks the API.
>>>>>>> Huh? Of course the new API has different semantics, but that's the whole
>>>>>>> point of it. The else branch here is to optimize for the case where a
>>>>>>> command has no payload. In that case there is no need for allocating an
>>>>>>> extra buffer, since the command byte is the only data transferred.
>>>>>> If this is the whole point of it why patch description says nothing
>>>>>> about it?
>>>>> I thought the patch description said this. What do you think needs to be
>>>>> added?
>>>> In short, that new API assumes that transfer callback must interpret
>>>> write buffer
>>>> differently than before :) Ie that sometimes at the beginning of the buffer
>>>> there will be additional bytes.
>>> Right, we never defined exactly which part of code would take which data
>>> and into what data it would be transformed. That obviously breaks as
>>> soon as two implementations make different assumptions. =)
>> In previous version of this patch [1] you have made different assumption,
>> and in the discussion you were clearly aware of the current implementation,
>> so your reaction here surprises me little bit. Maybe some misunderstanding.
>> Anyway I am glad we are now both aware of the problem.
>>
>> [1]: http://permalink.gmane.org/gmane.comp.video.dri.devel/110647
> It's possible I didn't realize the full complexity of the problem at the
> time. To summarize, I think the helpers in the core should do as much
> work as they possibly can, so that drivers don't have to duplicate the
> same over and over again. That is, the DCS helpers that are under
> discussion here should create a buffer that reflects the packet that is
> to be sent to the DSI peripheral, including the specific layout of the
> header. So a struct mipi_dsi_msg contains the following information:
>
> 	- .channel + .type make up the DI (Data ID) field in the packet
> 	  header
>
> 	for short packets:
> 	- .tx_buf[0] and .tx_buf[1] correspond to Data 0 and Data 1
> 	  fields in the packet header (both of these are only present if
> 	  .tx_len is larger than 0 and 1, and default to 0 otherwise)
>
> 	for long packets:
> 	- .tx_buf[0] and .tx_buf[1] correspond to the word count
> 	- .tx_buf[2..] represent the payload (word count - 2 bytes)
>
> That's pretty much what section 8.4 General Packet Structure of the DSI
> specification describes. This intentionally leaves out the header ECC
> and 16-bit packet footer (checksum). These are typically implemented in
> hardware, and if they aren't we can provide helpers that compute them
> based on the header, and the payload in .tx_buf. That way all the packet
> composition defined in the specification is handled by common code and a
> driver only needs to have the hardware-specific knowledge, namely where
> the various pieces need to be written in order to transmit them as
> desired.
>
> Does that make sense?
According to DSI specification we can describe DSI
message/command/transaction
on two levels:
1. Application layer - message is described by a triplet (data_type,
channel_id, data).
2. Protocol layer - message is described as a four byte packet header,
optionally
followed by packet data (payload) and checksum (which we can skip it
here as it should be generated by HW).

In the current API the 1st approach is used. There is no defined
standard how to program
DSI host to generate specific message, so this approach seems to be the
most natural in general case.

On the other side all DSI hosts I looked at use approach based on
protocol layer, ie.
packet header is written to one FIFO register and payload to another
(exynos, intel, omap) or the same (tegra).

Your proposition is something close to 2nd approach, maybe it would be
better to convert to completely to 2nd approach:

struct mipi_dsi_msg {
    u8 header[4]; /* u32 header;  ??? */
    void *payload; /* NULL in case of short packets */
    size_t payload_len;
    ...
};

Anyway, I think conversion to protocol layer should be done by helper
but this helper should be called rather from dsi host,
otherwise we can encounter some day dsi host which we need to feed with
data differently and we will need to perform
back-conversion from protocol layer to app layer, it will not be
difficult it will be just ugly :)

What about creating helpers to make dsi packets from current dsi
message. Sth like:

... drm_mipi_create_packet(struct mipi_dsi_packet *pkt, struct
mipi_dsi_msg *msg)
{
    if (long packet) {
        pkt->header = ...;
        pkt->payload = msg->tx_buf;
        pkt->len = msg->tx_len;
    } else {
        pkt->header = ...;
        pkt->payload = NULL;
        pkt->len = 0;
    }
}

This way in dsi host we will have sth like:

host_transfer(...msg) {
    struct mipi_dsi_packet pkt;

    drm_mipi_create_packet(&pkt, msg);

    writel(msg.header, REG_HDR);

   for (i = 0; i < pkt.len; i += 4)
        writel(pkt.payload[i..i+3], REG_PAYLOAD);
}

Please note that this way we can avoid dynamic memory
allocation/copy/deallocation, I know it is cheap, but it does not seems
to be necessary.


Regards
Andrzej