On Mon, Mar 1, 2021 at 1:10 PM Andy Shevchenko
<andriy.shevchenko@xxxxxxxxxxxxxxx> wrote:
> On Mon, Mar 01, 2021 at 02:09:25PM +0200, Andy Shevchenko wrote:
> > On Mon, Mar 01, 2021 at 05:24:41PM +0530, Viresh Kumar wrote:
> > > On 01-03-21, 14:41, Jie Deng wrote:
> > > > +/**
> > > > + * struct virtio_i2c_req - the virtio I2C request structure
> > > > + * @out_hdr: the OUT header of the virtio I2C message
> > > > + * @write_buf: contains one I2C segment being written to the device
> > > > + * @read_buf: contains one I2C segment being read from the device
> > > > + * @in_hdr: the IN header of the virtio I2C message
> > > > + */
> > > > +struct virtio_i2c_req {
> > > > + struct virtio_i2c_out_hdr out_hdr;
> > > > + u8 *write_buf;
> > > > + u8 *read_buf;
> > > > + struct virtio_i2c_in_hdr in_hdr;
> > > > +};
> > >
> > > I am not able to appreciate the use of write/read bufs here as we
> > > aren't trying to read/write data in the same transaction. Why do we
> > > have two bufs here as well as in specs ?
> >
> > I涎 and SMBus support bidirectional transfers as well. I think two buffers is
> > the right thing to do.
>
> Strictly speaking "half duplex".
But the driver does not support this at all: the sglist always has three
members as Viresh says: outhdr, msgbuf and inhdr. It then uses a
bounce buffer for the actual data transfer, and this always goes either
one way or the other.
I think the more important question is: does this driver actually need
the bounce buffer? It doesn't have to worry about adjacent stack
data being clobbered by cache maintenance because virtio is always
cache coherent and, so I suspect the bounce buffer can be left out.
It might actually be simpler to just have a fixed-length array of
headers on the stack and at most the corresponding number of
transfers for one virtqueue_kick().
Is there a reasonable limit on how many transfers we would
expect to handle at once? I see that most callers of i2c_transfer()
hardcode the number to '1' or '2', rarely '3' or '4', while the proposed
implementation seems to be optimized for much larger numbers.
Arnd
