On Fri, Jul 12, 2019 at 2:49 AM Andrew Morton <a...@linux-foundation.org> wrote:
> On Wed, 3 Jul 2019 10:10:55 +0200 Arnd Bergmann <a...@arndb.de> wrote:
> <scratches head>
>
> Surely clang is being extraordinarily dumb here?
>
> DECLARE_WAIT_QUEUE_HEAD_ONSTACK() is effectively doing
>
> struct wait_queue_head name = ({ __init_waitqueue_head(&name) ; name;
> })
>
> which is perfectly legitimate! clang has no business assuming that
> __init_waitqueue_head() will do any reads from the pointer which it was
> passed, nor can clang assume that __init_waitqueue_head() leaves any of
> *name uninitialized.
>
> Does it also warn if code does this?
>
> struct wait_queue_head name;
> __init_waitqueue_head(&name);
> name = name;
>
> which is equivalent, isn't it?
No, it does not warn for this.
I've tried a few more variants here: https://godbolt.org/z/ykSX0r
What I think is going on here is a result of clang and gcc fundamentally
treating -Wuninitialized warnings differently. gcc tries to make the warnings
as helpful as possible, but given the NP-complete nature of this problem
it won't always get it right, and it traditionally allowed this syntax as a
workaround.
int f(void)
{
int i = i; // tell gcc not to warn
return i;
}
clang apparently implements the warnings in a way that is as
completely predictable (and won't warn in cases that it
doesn't completely understand), but decided as a result that the
gcc 'int i = i' syntax is bogus and it always warns about a variable
used in its own declaration that is later referenced, without looking
at whether the declaration does initialize it or not.
> The proposed solution is, effectively, to open-code
> __init_waitqueue_head() at each DECLARE_WAIT_QUEUE_HEAD_ONSTACK()
> callsite. That's pretty unpleasant and calls for an explanatory
> comment at the __WAIT_QUEUE_HEAD_INIT_ONSTACK() definition site as well
> as a cautionary comment at the __init_waitqueue_head() definition so we
> can keep the two versions in sync as code evolves.
Yes, makes sense.
> Hopefully clang will soon be hit with the cluebat (yes?) and this
> change becomes obsolete in the quite short term. Surely 6-12 months
> from now nobody will be using the uncluebatted version of clang on
> contemporary kernel sources so we get to remove this nastiness again.
> Which makes me wonder whether we should merge it at all.
Would it make you feel better to keep the current code but have an alternative
version guarded with e.g. "#if defined(__clang__ && (__clang_major__ <= 9)"?
While it is probably a good idea to fix clang here, this is one of the last
issues that causes a significant difference between gcc and clang in build
testing with kernelci:
https://kernelci.org/build/next/branch/master/kernel/next-20190709/
I'm trying to get all the warnings fixed there so we can spot build-time
regressions more easily.
Arnd
