by Manivannan Sadhasivam

[permalink] [raw]

Subject: Re: [PATCH v2 11/16] bus: mhi: core: Add support for data transfer

Hi Arnd,

On Mon, Feb 17, 2020 at 10:17:51PM +0530, Manivannan Sadhasivam wrote:
> Hi Arnd,
>
> On Mon, Feb 17, 2020 at 05:13:37PM +0100, Arnd Bergmann wrote:
> > On Fri, Jan 31, 2020 at 2:51 PM Manivannan Sadhasivam
> > <[email protected]> wrote:
> >
> > > @@ -648,6 +715,31 @@ static int parse_ch_cfg(struct mhi_controller *mhi_cntrl,
> > > mhi_chan->db_cfg.pollcfg = ch_cfg->pollcfg;
> > > mhi_chan->xfer_type = ch_cfg->data_type;
> > >
> > > + switch (mhi_chan->xfer_type) {
> > > + case MHI_BUF_RAW:
> > > + mhi_chan->gen_tre = mhi_gen_tre;
> > > + mhi_chan->queue_xfer = mhi_queue_buf;
> > > + break;
> > > + case MHI_BUF_SKB:
> > > + mhi_chan->queue_xfer = mhi_queue_skb;
> > > + break;
> > > + case MHI_BUF_SCLIST:
> > > + mhi_chan->gen_tre = mhi_gen_tre;
> > > + mhi_chan->queue_xfer = mhi_queue_sclist;
> > > + break;
> > > + case MHI_BUF_NOP:
> > > + mhi_chan->queue_xfer = mhi_queue_nop;
> > > + break;
> > > + case MHI_BUF_DMA:
> > > + case MHI_BUF_RSC_DMA:
> > > + mhi_chan->queue_xfer = mhi_queue_dma;
> > > + break;
> > > + default:
> > > + dev_err(mhi_cntrl->dev,
> > > + "Channel datatype not supported\n");
> > > + goto error_chan_cfg;
> > > + }
> > > +
> >
> > While looking through the driver to see how the DMA gets handled, I came
> > across the multitude of mhi_queue_* functions, which seems like a
> > layering violation to me, given that they are all implemented by the
> > core code as well, and the client driver needs to be aware of
> > which one to call. Are you able to lift these out of the common interface
> > and make the client driver call these directly, or maybe provide a direct
> > interface based on mhi_buf_info to replace these?
> >
>
> It sounds reasonable to me. Let me discuss this internally with Qcom guys to
> see if they have any objections.
>

I looked into this in detail and found that the queue_xfer callbacks are tied
to the MHI channels. For instance, the callback gets attached to ul_chan (uplink
channel) and dl_chan (downlink channel) during controller registration. And
when the device driver calls the callback, the MHI stack calls respective queue
function for the relevant channel. For instance,

```
int mhi_queue_transfer(struct mhi_device *mhi_dev,
enum dma_data_direction dir, void *buf, size_t len,
enum mhi_flags mflags)
{
if (dir == DMA_TO_DEVICE)
return mhi_dev->ul_chan->queue_xfer(mhi_dev, mhi_dev->ul_chan,
buf, len, mflags);
else
return mhi_dev->dl_chan->queue_xfer(mhi_dev, mhi_dev->dl_chan,
buf, len, mflags);
}
```

If we use the direct queue API's this will become hard to handle. So, I'll keep
it as it is.

Thanks,
Mani

> Thanks,
> Mani
>
> > Arnd

2020-02-18 14:35:02

by Arnd Bergmann

[permalink] [raw]

Subject: Re: [PATCH v2 11/16] bus: mhi: core: Add support for data transfer

On Tue, Feb 18, 2020 at 6:51 AM Manivannan Sadhasivam
<[email protected]> wrote:
> On Mon, Feb 17, 2020 at 10:17:51PM +0530, Manivannan Sadhasivam wrote:
> > On Mon, Feb 17, 2020 at 05:13:37PM +0100, Arnd Bergmann wrote:
> > > On Fri, Jan 31, 2020 at 2:51 PM Manivannan Sadhasivam
> > > <[email protected]> wrote:
> > > While looking through the driver to see how the DMA gets handled, I came
> > > across the multitude of mhi_queue_* functions, which seems like a
> > > layering violation to me, given that they are all implemented by the
> > > core code as well, and the client driver needs to be aware of
> > > which one to call. Are you able to lift these out of the common interface
> > > and make the client driver call these directly, or maybe provide a direct
> > > interface based on mhi_buf_info to replace these?
> > >
> >
> > It sounds reasonable to me. Let me discuss this internally with Qcom guys to
> > see if they have any objections.
> >
>
> I looked into this in detail and found that the queue_xfer callbacks are tied
> to the MHI channels. For instance, the callback gets attached to ul_chan (uplink
> channel) and dl_chan (downlink channel) during controller registration. And
> when the device driver calls the callback, the MHI stack calls respective queue
> function for the relevant channel. For instance,
>
> ```
> int mhi_queue_transfer(struct mhi_device *mhi_dev,
> enum dma_data_direction dir, void *buf, size_t len,
> enum mhi_flags mflags)
> {
> if (dir == DMA_TO_DEVICE)
> return mhi_dev->ul_chan->queue_xfer(mhi_dev, mhi_dev->ul_chan,
> buf, len, mflags);
> else
> return mhi_dev->dl_chan->queue_xfer(mhi_dev, mhi_dev->dl_chan,
> buf, len, mflags);
> }
> ```
>
> If we use the direct queue API's this will become hard to handle. So, I'll keep
> it as it is.

Please have another look, this is exactly the part of the subsystem
that I think should be replaced. For the caller, there should not be much
difference between passing ul_chan/dl_chan or
DMA_TO_DEVICE/DMA_FROM_DEVICE, so that too can be lifted
to a higher level.

Arnd