Re: [PATCH] string.h: Mark 34 functions with __must_check

From: David Sterba
Date: Thu Oct 10 2019 - 10:27:23 EST


On Wed, Oct 09, 2019 at 10:33:45AM -0700, Nick Desaulniers wrote:
> On Wed, Oct 9, 2019 at 9:38 AM Joe Perches <joe@xxxxxxxxxxx> wrote:
> >
> > On Wed, 2019-10-09 at 09:13 -0700, Nick Desaulniers wrote:
> > > On Wed, Oct 9, 2019 at 8:09 AM Steven Rostedt <rostedt@xxxxxxxxxxx> wrote:
> > > > On Wed, 9 Oct 2019 14:14:28 +0200 Markus Elfring <Markus.Elfring@xxxxxx> wrote:
> > []
> > > > > Several functions return values with which useful data processing
> > > > > should be performed. These values must not be ignored then.
> > > > > Thus use the annotation â__must_checkâ in the shown function declarations.
> > []
> > > > I'm curious. How many warnings showed up when you applied this patch?
> > >
> > > I got zero for x86_64 and arm64 defconfig builds of linux-next with
> > > this applied. Hopefully that's not an argument against the more
> > > liberal application of it? I view __must_check as a good thing, and
> > > encourage its application, unless someone can show that a certain
> > > function would be useful to call without it.
> >
> > stylistic trivia, neither agreeing nor disagreeing with the patch
> > as I generally avoid reading Markus' patches.
> >
> > I believe __must_check is best placed before the return type as
> > that makes grep for function return type easier to parse.
> >
> > i.e. prefer
> > [static inline] __must_check <type> <function>(<args...>);
> > over
> > [static inline] <type> __must_check <function>(<args...>);
> >
>
> + Miguel
> So I just checked `__cold`, and `__cold` is all over the board in
> style. I see it:
> 1. before anything fs/btrfs/super.c#L101
> 2. after static before return type (what you recommend) fs/btrfs/super.c#L2318
> 3. after return type fs/btrfs/inode.c#L9426

As you can see in the git history, case 1 is from 2015 and the newer
changes put the attribute between type and name - that's my "current"
but hopefully final preference.

> Can we pick a style and enforce it via checkpatch? (It's probably not
> fun to check for each function attribute in
> include/linux/compiler_attributes.h).

Anything that has the return type, attributes and function name on one
line works for me, but I know that there are other style preferences
that put function name as the first word on a separate line. My reasons
are for better search results, ie.

extent_map.c:void __cold extent_map_exit(void)
extent_map.h:void __cold extent_map_exit(void);
file.c:void __cold btrfs_auto_defrag_exit(void)
inode.c:void __cold btrfs_destroy_cachep(void)
ordered-data.c:void __cold ordered_data_exit(void)
ordered-data.h:void __cold ordered_data_exit(void);

is better than

send.c:__cold
super.c:__cold
super.c:__cold
super.c:__cold

which I might get to fix eventually.