On Wed, Jun 01, 2022 at 12:26:27PM +0200, Toke Høiland-Jørgensen wrote:
> Luc Van Oostenryck <luc.vanoostenryck@gmail.com> writes:
> 
> > On Mon, May 23, 2022 at 12:30:14PM +0200, Toke Høiland-Jørgensen wrote:
> >> kernel test robot <lkp@intel.com> writes:
> >> 
> >> > tree:   https://git.kernel.org/pub/scm/linux/kernel/git/torvalds/linux.git master
> >> > head:   eaea45fc0e7b6ae439526b4a41d91230c8517336
> >> > commit: 782347b6bcad07ddb574422e01e22c92e05928c8 xdp: Add proper __rcu annotations to redirect map entries
> >> > date:   11 months ago
> >> > config: ia64-randconfig-s031-20220522 (https://download.01.org/0day-ci/archive/20220522/202205222029.xpW3PM1y-lkp@intel.com/config)
> >> > compiler: ia64-linux-gcc (GCC) 11.3.0
> >> 
> >> Hmm, so this is ia64-only? Some kind of macro breakage? Paul, any ideas?
> >
> > Hi,
> >
> > It's surely IA64's cmpxchg() which contains lines like:
> > 	_r_ = ia64_cmpxchg8_##sem((__u64 *) ptr, new, _o_); 
> 
> Oh, right. Hmm, well, if the cmpxchg does an internal cast that
> complicates things a bit. My immediate thought was to move the
> unrcu_pointer() inside the calls to cmpxchg(), like:
 
> But that seems to confuse sparse because these are ptr-to-ptr
> constructs:

Yes, that can't work because it applies on the wrong level (same difference as
between "int const ** ptr" and "int * const * ptr").

I've taken a quick look and the problem is really to be solved in IA64's
macros for cmpxchg() and friends. Two things need to be done:
1) avoid casts like the "(__u64 *) ptr" here above (ideally no cast would
   be needed but a "(__u64 __force *) ptr" would be pefectly acceptable in
   such macros.
2) the value returned by these macros must match the type of the pointer
   and the old/new values. For example, on x86 such macros are written as:
	({
		__typeof__ (*(ptr)) __ret = (arg);
		switch (sizeof(*(ptr))) { 
		case ...
			... use ptr without dropping the address space ..

		__ret;
	)}

See, for example, arch/x86/include/asm/cmpxchg.h

-- Luc