Re: [PATCH 1/2] lib/string_helpers: Introduce strsplit_u32()

[Andy Shevchenko]

On Tue, Jul 12, 2022 at 3:51 PM Cezary Rojewski
<cezary.rojewski@xxxxxxxxx> wrote:
> On 2022-07-09 10:42 PM, Andy Shevchenko wrote:
> > On Sat, Jul 09, 2022 at 10:45:49AM +0200, Cezary Rojewski wrote:
> >> On 2022-07-08 6:49 PM, Andy Shevchenko wrote:
> >>> On Fri, Jul 8, 2022 at 6:32 PM Cezary Rojewski
> >>> <cezary.rojewski@xxxxxxxxx> wrote:
> >>>> On 2022-07-08 5:25 PM, Andy Shevchenko wrote:
> >>>>> On Fri, Jul 8, 2022 at 2:34 PM Péter Ujfalusi
> >>>>> <peter.ujfalusi@xxxxxxxxxxxxxxx> wrote:

...

> >>>> A long shot, but what if we were to modify get_options() so it takes
> >>>> additional element-size parameter instead?
> >>>
> >>> But why? int / unsigned int, u32 / s32  are all compatible in the current cases.
> >>
> >> I'd like to avoid any additional operations, so that the retrieved payload
> >> can be provided to the IPC handler directly. The IPC handlers for AudioDSP
> >> drivers are expecting payload in u32s.
> >>
> >> // u32 **tkns, size_t *num_tkns as foo() arguments
> >> // u32 *ints, int nints as locals
> >>
> >>      get_options(buf, 0, &nints);
> >>      if (!nints) {
> >>              ret = -ENOENT;
> >>              goto free_buf;
> >>      }
> >>
> >>      ints = kcalloc(nints + 1, sizeof(*ints), GFP_KERNEL);
> >>      if (!ints) {
> >>              ret = -ENOMEM;
> >>              goto free_buf;
> >>      }
> >>
> >>      get_num_options(buf, nints + 1, ints, sizeof(*ints));
> >>
> >>      *tkns = ints;
> >>      *num_tkns = nints;
> >>
> >> No additional operations in between. The intermediate IPC handler can later
> >> refer to the actual payload via &tkns[1] before passing it to the generic
> >> one.
> >>
> >> Casting int array into u32 array does not feel right, or perhaps I'm missing
> >> something like in the doc case.
> >
> > C standard.
> >
> > int to unsigned int is not promoted. And standard says that "The rank of any
> > unsigned integer type shall equal the rank of the corresponding signed integer
> > type, if any."
> >
> > I don't know why one needs to have an additional churn here. int and unsigned
> > int are interoperable with the adjustment to the sign when the other argument
> > is signed or lesser rank of.
>
> I still believe that casting blindly is not the way to go. I did
> explicitly ask about int vs u32,

There is no such type in the C standard.

> not int vs unsigned int. Please note
> that these values are later passed to the IPC handlers, and this changes
> the context a bit. If hw expects u32, then u32 it shall be.

H/W doesn't expect u32, HW expects bytes or group of bytes with:
1) dedicated address alignment (if required);
2) dedicated byte order;
3) dedicated padding (if required).

Correct me if I'm wrong.

> Please correct me if I'm wrong, but there is no guarantee that int is
> always 32bits long.

There is no guarantee by the C standard, indeed, but there is an upper
level guarantee, by the Linux kernel.

> What is guaranteed though, is that int holds at
> least -/+ 32,767. Also, values larger than INT_MAX are allowed in the
> IPC payload.

Yeah... this is binary protocol, right? So, what limits are you
talking about here if they are not applicable there anyway. It's
simply different dimension of limits (i.e. bytes and bits and not C
language types).

-- 
With Best Regards,
Andy Shevchenko