Re: [PATCH] add sched_getcpu

[Alexander Monakov <amonakov@ispras.ru>]

On Mon, 29 Feb 2016, Rich Felker wrote:
> > (this include could also be dropped; I think it's a matter of policy whether
> > such includes are desirable or not, so please wait for comment from Rich)
> 
> Policy is to always include the header with the public declaration
> (and any feature test macros necessary to get it) so that the compiler
> checks the implementation against the public declaration.

This policy certain makes sense; I pointed that out because I've seen it
violated; at least the following files violate it by defining something
without including anything:

arch/arm/src/__aeabi_atexit.c
src/internal/procfdname.c
src/misc/gethostid.c
src/prng/__seed48.c
src/signal/restore.c
src/signal/sigrtmin.c
src/stdlib/abs.c
src/stdlib/labs.c
src/stdlib/llabs.c
src/time/__month_to_secs.c
src/time/__year_to_secs.c

(but e.g. in procfdname.c there's nothing to include because every caller
declares the prototype, and also hardcodes the buffer size -- see my recent
patch)

> > This is wrong, as it doesn't set errno on error, and does not produce -1. This
> > should be something like 'return s ? __syscall_ret(s) : c;' or maybe
> > 'return __syscall_ret(s ? s : c);'.
> 
> Why is an error even possible here? Is it just to account for ancient
> kernels that lack the syscall?

Apparently so -- afaics current kernels can only produce EFAULT, which cannot
happen in this case.

(I was wrong about the need to use __syscall_ret, but that's for another
subthread)

Alexander