Re: [Qemu-devel] [PATCH 1.1] coroutine: Avoid ucontext usage on i386 Linux host

[Michael Tokarev]

On 09.05.2012 23:21, Jan Kiszka wrote:
[]
> --- a/configure
> +++ b/configure
> @@ -2777,17 +2777,22 @@ fi
>  # windows autodetected by make
>  if test "$coroutine" = "" -o "$coroutine" = "ucontext"; then
>    if test "$darwin" != "yes"; then
> -    cat > $TMPC << EOF
> +    if test "$linux" = "yes" -a "$cpu" = "i386"; then
> +      # RT signal mask corruption for 32-on-64 bit prevents ucontext usage
> +      coroutine_backend=gthread
> +    else

I'd say this is unacceptable.  We're making 32bit x86 code to use
different implementation of one of core subsystems, and 32bit code
already receives much less testing because everyone is using 64bit
code and many don't care about 32bits (this is general 32bit case,
not a more exotic 32-on-64bit case).  I think it is a sure way to
have more subtle "works for me" bugs which happens only on 32bits..

Besides, gthread is quite a bit slower than ucontext (but at least
it works even in the exotic case).

Thanks,

/mjt