Re: [PATCH 2/2] rpmsg: glink: Consolidate TX_DATA and TX_DATA_CONT

[Bjorn Andersson]

On Fri, Apr 07, 2023 at 03:10:45PM -0700, Chris Lew wrote:
> 
> 
> On 3/27/2023 7:41 AM, Bjorn Andersson wrote:
> > Rather than duplicating most of the code for constructing the initial
> > TX_DATA and subsequent TX_DATA_CONT packets, roll them into a single
> > loop.
> > 
> > Signed-off-by: Bjorn Andersson <quic_bjorande@xxxxxxxxxxx>
> > ---
> >   drivers/rpmsg/qcom_glink_native.c | 46 +++++++++----------------------
> >   1 file changed, 13 insertions(+), 33 deletions(-)
> > 
> > diff --git a/drivers/rpmsg/qcom_glink_native.c b/drivers/rpmsg/qcom_glink_native.c
> > index 62634d020d13..082cf7f4888e 100644
> > --- a/drivers/rpmsg/qcom_glink_native.c
> > +++ b/drivers/rpmsg/qcom_glink_native.c
> > @@ -1309,7 +1309,7 @@ static int __qcom_glink_send(struct glink_channel *channel,
> >   	int ret;
> >   	unsigned long flags;
> >   	int chunk_size = len;
> > -	int left_size = 0;
> > +	size_t offset = 0;
> >   	if (!glink->intentless) {
> >   		while (!intent) {
> > @@ -1343,49 +1343,29 @@ static int __qcom_glink_send(struct glink_channel *channel,
> >   		iid = intent->id;
> >   	}
> > -	if (wait && chunk_size > SZ_8K) {
> > -		chunk_size = SZ_8K;
> > -		left_size = len - chunk_size;
> > -	}
> > -	req.msg.cmd = cpu_to_le16(GLINK_CMD_TX_DATA);
> > -	req.msg.param1 = cpu_to_le16(channel->lcid);
> > -	req.msg.param2 = cpu_to_le32(iid);
> > -	req.chunk_size = cpu_to_le32(chunk_size);
> > -	req.left_size = cpu_to_le32(left_size);
> > -
> > -	ret = qcom_glink_tx(glink, &req, sizeof(req), data, chunk_size, wait);
> > -
> > -	/* Mark intent available if we failed */
> > -	if (ret) {
> > -		if (intent)
> > -			intent->in_use = false;
> > -		return ret;
> > -	}
> > -
> > -	while (left_size > 0) {
> > -		data = (void *)((char *)data + chunk_size);
> > -		chunk_size = left_size;
> > -		if (chunk_size > SZ_8K)
> > +	while (offset < len) {
> > +		chunk_size = len - offset;
> > +		if (chunk_size > SZ_8K && (wait || offset > 0))
> 
> offset > 0 seems to be a new condition compared to the previous logic.
> Are we adding this as a cached check because we know if offset is set then
> fragmented sends are allowed?
> 

You're right, I believe my intention was to retain the two inquiries of
the original code; for the first block, don't split it if we're not
waiting and for any subsequent blocks always split.

> I don't think wait would have changed during the loop, so I'm not sure if
> offset > 0 is adding any extra value to the check.
> 

But you're totally right, offset > 0 would only occur if wait is set and
wait will not have changed for subsequent blocks.

So while capturing the original conditions, it seems superfluous.

Thanks,
Bjorn

> >   			chunk_size = SZ_8K;
> > -		left_size -= chunk_size;
> > -		req.msg.cmd = cpu_to_le16(GLINK_CMD_TX_DATA_CONT);
> > +		req.msg.cmd = cpu_to_le16(offset == 0 ? GLINK_CMD_TX_DATA : GLINK_CMD_TX_DATA_CONT);
> >   		req.msg.param1 = cpu_to_le16(channel->lcid);
> >   		req.msg.param2 = cpu_to_le32(iid);
> >   		req.chunk_size = cpu_to_le32(chunk_size);
> > -		req.left_size = cpu_to_le32(left_size);
> > +		req.left_size = cpu_to_le32(len - offset - chunk_size);
> > -		ret = qcom_glink_tx(glink, &req, sizeof(req), data,
> > -				    chunk_size, wait);
> > -
> > -		/* Mark intent available if we failed */
> > +		ret = qcom_glink_tx(glink, &req, sizeof(req), data + offset, chunk_size, wait);
> >   		if (ret) {
> > +			/* Mark intent available if we failed */
> >   			if (intent)
> >   				intent->in_use = false;
> > -			break;
> > +			return ret;
> >   		}
> > +
> > +		offset += chunk_size;
> >   	}
> > -	return ret;
> > +
> > +	return 0;
> >   }
> >   static int qcom_glink_send(struct rpmsg_endpoint *ept, void *data, int len)