Re: [PATCH] rdma/siw: avoid smp_store_mb() on a u64

[Jason Gunthorpe]

On Fri, Jul 12, 2019 at 10:51:23AM +0200, Arnd Bergmann wrote:
> The new siw driver fails to build on i386 with
> 
> drivers/infiniband/sw/siw/siw_qp.c:1025:3: error: invalid output size for constraint '+q'
>                 smp_store_mb(*cq->notify, SIW_NOTIFY_NOT);
>                 ^
> include/asm-generic/barrier.h:141:35: note: expanded from macro 'smp_store_mb'
>  #define smp_store_mb(var, value)  __smp_store_mb(var, value)
>                                   ^
> arch/x86/include/asm/barrier.h:65:47: note: expanded from macro '__smp_store_mb'
>  #define __smp_store_mb(var, value) do { (void)xchg(&var, value); } while (0)
>                                               ^
> include/asm-generic/atomic-instrumented.h:1648:2: note: expanded from macro 'xchg'
>         arch_xchg(__ai_ptr, __VA_ARGS__);                               \
>         ^
> arch/x86/include/asm/cmpxchg.h:78:27: note: expanded from macro 'arch_xchg'
>  #define arch_xchg(ptr, v)       __xchg_op((ptr), (v), xchg, "")
>                                 ^
> arch/x86/include/asm/cmpxchg.h:48:19: note: expanded from macro '__xchg_op'
>                                       : "+q" (__ret), "+m" (*(ptr))     \
>                                               ^
> drivers/infiniband/sw/siw/siw_qp.o: In function `siw_sqe_complete':
> siw_qp.c:(.text+0x1450): undefined reference to `__xchg_wrong_size'
> drivers/infiniband/sw/siw/siw_qp.o: In function `siw_rqe_complete':
> siw_qp.c:(.text+0x15b0): undefined reference to `__xchg_wrong_size'
> drivers/infiniband/sw/siw/siw_verbs.o: In function `siw_req_notify_cq':
> siw_verbs.c:(.text+0x18ff): undefined reference to `__xchg_wrong_size'
> 
> Since smp_store_mb() has to be an atomic store, but the architecture
> can only do this on 32-bit quantities or smaller, but 'cq->notify'
> is a 64-bit word.
> 
> Apparently the smp_store_mb() is paired with a READ_ONCE() here, which
> seems like an odd choice because there is only a barrier on the writer
> side and not the reader, and READ_ONCE() is already not atomic on
> quantities larger than a CPU register.
> 
> I suspect it is sufficient to use the (possibly nonatomic) WRITE_ONCE()
> and an SMP memory barrier here. If it does need to be atomic as well
> as 64-bit quantities, using an atomic64_set_release()/atomic64_read_acquire()
> may be a better choice.
> 
> Fixes: 303ae1cdfdf7 ("rdma/siw: application interface")
> Fixes: f29dd55b0236 ("rdma/siw: queue pair methods")
> Cc: Peter Zijlstra <peterz@xxxxxxxxxxxxx>
> Signed-off-by: Arnd Bergmann <arnd@xxxxxxxx>
> ---
>  drivers/infiniband/sw/siw/siw_qp.c    | 4 +++-
>  drivers/infiniband/sw/siw/siw_verbs.c | 5 +++--
>  2 files changed, 6 insertions(+), 3 deletions(-)

Bernard, please send at patch for whatever solution we settled on
against 5.3-rc1

Thanks,
Jason