Re: [PATCH v3 23/25] bus: mhi: ep: Add support for queueing SKBs to the host

[Alex Elder <elder@linaro.org>]

On 2/22/22 8:38 AM, Manivannan Sadhasivam wrote:
> On Tue, Feb 15, 2022 at 04:40:29PM -0600, Alex Elder wrote:
>> On 2/12/22 12:21 PM, Manivannan Sadhasivam wrote:
>>> Add support for queueing SKBs to the host over the transfer ring of the
>>> relevant channel. The mhi_ep_queue_skb() API will be used by the client
>>> networking drivers to queue the SKBs to the host over MHI bus.
>>>
>>> The host will add ring elements to the transfer ring periodically for
>>> the device and the device will write SKBs to the ring elements. If a
>>> single SKB doesn't fit in a ring element (TRE), it will be placed in
>>> multiple ring elements and the overflow event will be sent for all ring
>>> elements except the last one. For the last ring element, the EOT event
>>> will be sent indicating the packet boundary.
>>>
>>> Signed-off-by: Manivannan Sadhasivam <manivannan.sadhasivam@linaro.org>
>>
>> I'm a little confused by this, so maybe you can provide
>> a better explanation somehow.
>>
>> 					-Alex
>>
>>> ---
>>>    drivers/bus/mhi/ep/main.c | 102 ++++++++++++++++++++++++++++++++++++++
>>>    include/linux/mhi_ep.h    |  13 +++++
>>>    2 files changed, 115 insertions(+)
>>>
>>> diff --git a/drivers/bus/mhi/ep/main.c b/drivers/bus/mhi/ep/main.c
>>> index baf383a4857b..e4186b012257 100644
>>> --- a/drivers/bus/mhi/ep/main.c
>>> +++ b/drivers/bus/mhi/ep/main.c
>>> @@ -488,6 +488,108 @@ int mhi_ep_process_tre_ring(struct mhi_ep_ring *ring, struct mhi_ep_ring_element
>>>    	return 0;
>>>    }
>>> +int mhi_ep_queue_skb(struct mhi_ep_device *mhi_dev, enum dma_data_direction dir,
>>> +		     struct sk_buff *skb, size_t len, enum mhi_flags mflags)
>>
>> Why are both skb and len supplied?  Will an skb be supplied
>> without wanting to send all of it?  Must len be less than
>> skb->len?  I'm a little confused about the interface.
>>
>> Also, the data direction is *out*, right?  You'll never
>> be queueing a "receive" SKB?
>>
> 
> This was done to be compatible with the MHI host API where the host can queue
> SKBs in both directions. But I think I should stop doing this.


OK.

>>> +{
>>> +	struct mhi_ep_chan *mhi_chan = (dir == DMA_FROM_DEVICE) ? mhi_dev->dl_chan :
>>> +								mhi_dev->ul_chan;
>>> +	struct mhi_ep_cntrl *mhi_cntrl = mhi_dev->mhi_cntrl;
>>> +	struct device *dev = &mhi_chan->mhi_dev->dev;
>>> +	struct mhi_ep_ring_element *el;
>>> +	struct mhi_ep_ring *ring;
>>> +	size_t bytes_to_write;
>>> +	enum mhi_ev_ccs code;
>>> +	void *read_from_loc;
>>> +	u32 buf_remaining;
>>> +	u64 write_to_loc;
>>> +	u32 tre_len;
>>> +	int ret = 0;
>>> +
>>> +	if (dir == DMA_TO_DEVICE)
>>> +		return -EINVAL;
>>
>> Can't you just preclude this from happening, or
>> know it won't happen by inspection?
>>
>>> +
>>> +	buf_remaining = len;
>>> +	ring = &mhi_cntrl->mhi_chan[mhi_chan->chan].ring;
>>> +
>>> +	mutex_lock(&mhi_chan->lock);
>>> +
>>> +	do {
>>> +		/* Don't process the transfer ring if the channel is not in RUNNING state */
>>> +		if (mhi_chan->state != MHI_CH_STATE_RUNNING) {
>>> +			dev_err(dev, "Channel not available\n");
>>> +			ret = -ENODEV;
>>> +			goto err_exit;
>>> +		}
>>> +
>>
>> It would be nice if the caller could know whether there
>> was enough room *before* you start transferring things.
>> It's probably a lot of work to get to that point though.
>>
> 
> No, the caller will do this check but the check is included here so that we
> don't run out of buffers when the packet needs to be splitted.
> 
>>> +		if (mhi_ep_queue_is_empty(mhi_dev, dir)) {
>>> +			dev_err(dev, "TRE not available!\n");
>>> +			ret = -EINVAL;
>>> +			goto err_exit;
>>> +		}
>>> +
>>> +		el = &ring->ring_cache[ring->rd_offset];
>>> +		tre_len = MHI_EP_TRE_GET_LEN(el);
>>> +		if (skb->len > tre_len) {
>>> +			dev_err(dev, "Buffer size (%d) is too large for TRE (%d)!\n",
>>> +				skb->len, tre_len);
>>
>> This means the receive buffer must be big enough to hold
>> any incoming SKB.  This is *without* checking for the
>> CHAIN flag in the TRE, so what you describe in the
>> patch description seems not to be true.  I.e., multiple
>> TREs in a TRD will *not* be consumed if the SKB data
>> requires more than what's left in the current TRE.
>>
> 
> I think I removed this check for v3 but somehow the change got lost :/

Looking at this now, it's possible I got confused about
which direction the data was moving; but I'm not really
sure.

 From the perspective of the endpoint device, this is the
*transmit* function.  But when the device is transmitting,
it is moving data into the *receive* buffers that the host
has allocated and supplied via the transfer ring.

My statement seems to be correct though, with this logic,
the host must supply a buffer large enough to receive the
entire next SKB, or it will get an error back.  I no longer
know what happens when this function (mhi_ep_queue_skb())
returns an error--is the skb dropped?

> But anyway, there is no need to check for CHAIN flag while writing to host.
> CHAIN flag is only used or even make sense when host writes data to device, so

I'm not sure that's correct, but I don't want to get into that issue here.
We can talk about that separately.

> that it knows the packet boundary and could use the CHAIN flag to tell the
> device where the boundary lies.

This doesn't sound to me like what the purpose of the CHAIN flag is,
but perhaps I'm misunderstanding you.  Let's have a quick private
chat about this so we don't waste any more e-mail bandwidth.

					-Alex

> But when the device writes to host, it already has the pre-queued elements from
> host that has no idea where the packet boundary lies. So the host would've set
> only EOT on all TREs and expects the device to send OVERFLOW event for TREs that
> don't have the complete packet. Then finally, when device sends EOT event, the
> host will detect the boundary.
> 
> Thanks,
> Mani