Re: [RFC 5/5] i3c: add i3cdev module to expose i3c dev in /dev

[Arnd Bergmann]

On Tue, Dec 10, 2019 at 8:15 PM Vitor Soares <Vitor.Soares@xxxxxxxxxxxx> wrote:
> From: Arnd Bergmann <arnd@xxxxxxxx>
> Date: Tue, Dec 10, 2019 at 17:51:14
>
> > On Tue, Dec 10, 2019 at 4:37 PM Vitor Soares <Vitor.Soares@xxxxxxxxxxxx> wrote:
> > >
> > > +/* IOCTL commands */
> > > +#define I3C_DEV_IOC_MAGIC      0x07
> > > +
> > > +struct i3c_ioc_priv_xfer {
> > > +       struct i3c_priv_xfer __user *xfers;     /* pointers to i3c_priv_xfer */
> > > +       __u32 nxfers;                           /* number of i3c_priv_xfer */
> > > +};
> > > +
> > > +#define I3C_IOC_PRIV_XFER      \
> > > +       _IOW(I3C_DEV_IOC_MAGIC, 30, struct i3c_ioc_priv_xfer)
> > > +
> > > +#define  I3C_IOC_PRIV_XFER_MAX_MSGS    42
> >
> > This is not a great data structure for UAPI, please see
> > https://urldefense.proofpoint.com/v2/url?u=https-3A__git.kernel.org_pub_scm_linux_kernel_git_arnd_playground.git_tree_Documentation_core-2Dapi_ioctl.rst-3Fh-3Dcompat-2Dioctl-2Dendgame-26id-3D927324b7900ee9b877691a8b237e272fabb21bf5&d=DwIBaQ&c=DPL6_X_6JkXFx7AXWqB0tg&r=qVuU64u9x77Y0Kd0PhDK_lpxFgg6PK9PateHwjb_DY0&m=5Q9WjK0o93NR7DQ9NM6So6mfdgpNnZnSaP8qMpgaC7E&s=LzzjrUQAG8fx5jkVyK73dBDrahNAvk09Cxxlx3KOiXI&e=
> >
> > for some background. I'm planning to submit that documentation for
> > mainline integration soon.
> >
> >      Arnd
>
> Thanks for sharing the document.
>
> My understanding is that I should use a data structure like the struct
> spi_ioc_transfer, with this I may also use the same ioctl command
> definition. Am I right?

Yes, that would be an example of a structure that follows the best
practices from my document. It is still rather complex, so if you
can make it any simpler, that would be ideal.

> In the documentation you also refer the compact_ioctl() and It is not
> clear to me if the compact_ioctl() is mandatory in this case. Should I
> implement it as well?

If the structure is defined like that, you just need to set
".compat_ioctl=compat_ptr_ioctl," in the file_operations structure
and it will work, but you cannot skip that step.

As your interface is basically just read/write based, I wonder
if there is a way to completely avoid the ioctl and instead
use io_submit() as the primary interface.

      Arnd