Re: [Qemu-devel] [PATCH] linux-user: Fix do_store_exclusive for shared memory of interprocess.

[Emilio G. Cota]

(Adding Richard to Cc)

On Sat, Oct 15, 2016 at 23:53:48 +0800, Heiher wrote:
> From: Heiher <address@hidden>
> 
> test case: http://pastebin.com/raw/x2GW4xNW

You should check out this patchset and use it as a base for working on this
topic:

  http://lists.nongnu.org/archive/html/qemu-devel/2016-10/msg02341.html

In particular, the added tests/atomic_add-bench does a very similar thing
to what you're doing with your test case -- although with pthreads instead of
mmap(MAP_SHARED) + fork.

(more comments below)

> Signed-off-by: Heiher <address@hidden>
> ---
>  linux-user/main.c | 24 ++++++++++++++++++++++--
>  1 file changed, 22 insertions(+), 2 deletions(-)
> 
> diff --git a/linux-user/main.c b/linux-user/main.c
> index 0e31dad..81b0a49 100644
> --- a/linux-user/main.c
> +++ b/linux-user/main.c
> @@ -2312,6 +2312,23 @@ static const uint8_t mips_syscall_args[] = {
>  #  undef MIPS_SYS
>  # endif /* O32 */
>  
> +#define cmpxchg_user(old, new, gaddr, target_type)                   \
> +({                                                                   \
> +    abi_ulong __gaddr = (gaddr);                                     \
> +    target_type *__hptr;                                             \
> +    abi_long __ret = 0;                                                      
> \
> +    if ((__hptr = lock_user(VERIFY_WRITE, __gaddr, sizeof(target_type), 0))) 
> { \
> +        if ((old) != atomic_cmpxchg(__hptr, (old), (new)))           \
> +            __ret = -TARGET_EAGAIN;                                  \
> +        unlock_user(__hptr, __gaddr, sizeof(target_type));           \
> +    } else                                                           \
> +        __ret = -TARGET_EFAULT;                                              
> \
> +    __ret;                                                           \
> +})
> +
> +#define cmpxchg_user_u32(old, new, gaddr) cmpxchg_user((old), (new), 
> (gaddr), uint32_t)
> +#define cmpxchg_user_u64(old, new, gaddr) cmpxchg_user((old), (new), 
> (gaddr), uint64_t)
> +
>  static int do_store_exclusive(CPUMIPSState *env)
>  {
>      target_ulong addr;
> @@ -2342,12 +2359,15 @@ static int do_store_exclusive(CPUMIPSState *env)
>                  env->active_tc.gpr[reg] = 0;
>              } else {
>                  if (d) {
> -                    segv = put_user_u64(env->llnewval, addr);
> +                    segv = cmpxchg_user_u64(env->llval, env->llnewval, addr);
>                  } else {
> -                    segv = put_user_u32(env->llnewval, addr);
> +                    segv = cmpxchg_user_u32(env->llval, env->llnewval, addr);
>                  }
>                  if (!segv) {
>                      env->active_tc.gpr[reg] = 1;
> +                } else if (-TARGET_EAGAIN == segv) {
> +                    segv = 0;
> +                    env->active_tc.gpr[reg] = 0;
>                  }
>              }
>          }

With the atomic cmpxchg patch series referenced above, we should directly
translate to cmpxchg, thereby removing the exception--just like this
patch does for the Alpha architecture:
  http://lists.nongnu.org/archive/html/qemu-devel/2016-10/msg02373.html

Thanks

                Emilio