On Tue, Sep 06, 2022 at 03:54:06PM +0300, Andy Shevchenko wrote:
> On Mon, Sep 05, 2022 at 12:35:42PM -0700, Dmitry Torokhov wrote:
> > On Mon, Sep 05, 2022 at 08:20:01PM +0300, Andy Shevchenko wrote:
>
> ...
>
> > > + subsys = acpi_get_subsystem_id(ACPI_HANDLE(dev));
> > > + if (IS_ERR(subsys) && PTR_ERR(subsys) != -ENODATA)
> > > + return PTR_ERR(subsys);
> > > +
> > > + if (IS_ERR(subsys) && PTR_ERR(subsys) == -ENODATA)
> > > + subsys = kstrdup_const("unknown", GFP_KERNEL);
> >
> > Do we really need kstrdup_const() here? This makes me wonder if we need
> > to also have error handling here, and if we going to tip some automated
> > tools by not having it. Why can't we simply assign the constant here
> > (and continue using kfree_const() below)?
>
> Which makes code inconsistent. But okay, no big deal.
To me the *_const() APIs are needed when the code does not really know
if it deals with a const/read-only object or not. If we know for sure we
are dealing with a const/read-only object, we can skip allocation and
freeing, so I do not see any inconsistencies.
>
> > I think this is the case where PTR_ERR_OR_ZERO() might help avoid
> > multiple IS_ERR/PTR_ERR:
> >
> > subsys = acpi_get_subsystem_id(ACPI_HANDLE(dev));
> > error = PTR_ERR_OR_ZERO(subsys);
> > if (error == -ENODATA)
> > subsys = "unknown";
> > else if (error)
> > return error;
>
> Would it matter? The generated code will be the same in both cases, no?
No, in the end I think the optimizer will reduce both variants to the
same thing. I do find mine a bit more compact and thus easier to read,
but I will not insist.
Thanks.
--
Dmitry
