On Fri, May 27, 2016 at 01:20:29PM -0700, Linus Torvalds wrote:
> I didn't look at the details of your patch, but I did look at several
> IS_ERR_VALUE() uses in the standard kernel, and they were basically
> all wrong. Even the ones that used it for the rigth reason (vm_brk()
> that returns a pointer or an error in an "unsigned long") had actively
> screwed up and truncated that (correct) unsigned long value to "int"
> before doing the IS_ERR_VALUE(), which made it all wrong again.
>
> And the other users just looked entirely bogus, and were all just
> "zero or negative error code" that has nothing to do with
> IS_ERR_VALUE(). The code should just check against zero, not use that
> macro that was designed for something different.
Both gfs2 ones and the one in fs/romfs are crap. AFS one is simply ridiculous -
result = generic_file_write_iter(iocb, from);
if (IS_ERR_VALUE(result)) {
_leave(" = %zd", result);
return result;
}
_leave(" = %zd", result);
return result;
In binfmt*.c some callers are valid (vm_mmap and vm_brk values; BTW, the
code in binfmt_flat.c checks for vm_mmap returning 0, which should never
happen, AFAICS). Most of binfmt_flat.c ones are pure cargo-culting.
net/9p/client.c ones are results of serious mistake in design of 9P.L
extensions to 9P - they assume that numeric values of E... are
architecture-independent and send them over the wire. The uses of
IS_ERR_VALUE there are papering over that. Badly.
A couple of sound/* ones should be simply "is negative".
netfilter ones are caused by bad calling conventions of
xt_percpu_counter_alloc().
I hadn't looked through drivers/* users, but judging by fs/, net/ and sound/,
I would expect them to be pointless garbage in the best case and red flags
for broken code in the worst.
IMO IS_ERR_VALUE() should be renamed to something less generic and be used
a lot less than it is right now.
