Re: [PATCH v3 0/6] Add FIELD_MODIFY() helper

[Marc Zyngier <maz@kernel.org>]

On Fri, 18 Apr 2025 16:08:38 +0100,
Yury Norov <yury.norov@gmail.com> wrote:
> 
> On Thu, Apr 17, 2025 at 06:45:12PM +0100, Marc Zyngier wrote:
> > On Thu, 17 Apr 2025 18:22:29 +0100,
> > Andrew Lunn <andrew@lunn.ch> wrote:
> > > 
> > > On Thu, Apr 17, 2025 at 12:10:54PM +0100, Marc Zyngier wrote:
> > > > On Thu, 17 Apr 2025 11:47:07 +0100,
> > > > Luo Jie <quic_luoj@quicinc.com> wrote:
> > > > > 
> > > > > Add the helper FIELD_MODIFY() to the FIELD_XXX family of bitfield
> > > > > macros. It is functionally similar as xxx_replace_bits(), but adds
> > > > > the compile time checking to catch incorrect parameter type errors.
> > > > > 
> > > > > This series also converts the four instances of opencoded FIELD_MODIFY()
> > > > > that are found in the core kernel files, to instead use the new
> > > > > FIELD_MODIFY() macro. This is achieved with Coccinelle, by adding
> > > > > the script field_modify.cocci.
> > > > > 
> > > > > The changes are validated on IPQ9574 SoC which uses ARM64 architecture.
> > > > 
> > > > We already have the *_replace_bits() functions (see
> > > > include/linux/bitfield.h).
> > > > 
> > > > Why do we need extra helpers?
> > > 
> > > If you look at bitfield.h, the *_replace_bits() seem to be
> > > undocumented internal macro magic, not something you are expected to
> > > use. What you are expected to use in that file is however well
> > > documented. The macro magic also means that cross referencing tools
> > > don't find them.
> > 
> > $ git grep _replace_bits|  wc -l
> > 1514
> 
> FIELD_PREP() only is used 10 times more.

And? I'm sure that if you count "+", you'll find it to be yet a few
order of magnitudes more.

>  
> > I think a bunch of people have found them, tooling notwithstanding.
> > 
> > As for the documentation, the commit message in 00b0c9b82663ac would
> > be advantageously promoted to full-fledged kernel-doc.
> 
> The FIELD_MODIFY() and uxx_replace_bits() are simply different things.
> 
> FIELD_MODIFY() employs __BF_FIELD_CHECK(), which allows strict
> parameters checking at compile time. And people like it. See
> recent fixed-size GENMASK() series:
> 
> https://patchwork.kernel.org/comment/26283604/

I don't care much for what people like or not. I don't see this
FIELD_MODIFY() as a particular improvement on *_replace_bits().

> The _replace_bits() functions return fixed-width values, and intended
> for: "manipulating bitfields both in host- and fixed-endian", as the
> very first line in the commit message says.
> 
> Those using _replace_bits() for something else abuse the API, and
> should switch to FIELD_MODIFY().

Or not.

	M.

-- 
Jazz isn't dead. It just smells funny.