On Thu, Aug 22, 2024 at 10:05 AM H.J. Lu <hjl.tools@xxxxxxxxx> wrote:
>
> On Thu, Aug 22, 2024 at 5:21 AM Adhemerval Zanella Netto
> <adhemerval.zanella@xxxxxxxxxx> wrote:
> >
> >
> >
> > On 21/08/24 16:41, Alejandro Colomar wrote:
> > > Hi Adhemerval, Elliott,
> > >
> > > On Wed, Aug 21, 2024 at 03:19:02PM GMT, Adhemerval Zanella Netto wrote:
> > >> I would use either the kernel interface:
> > >>
> > >> arch/x86/kernel/process_64.c
> > >> 961 SYSCALL_DEFINE2(arch_prctl, int, option, unsigned long, arg2)
> > >>
> > >> Where is seems to what most caller do, or maybe something like:
> > >>
> > >> union __arch_prctl_arg
> > >> {
> > >> unsigned long addr;
> > >> unsigned long *vaddr;
> > >> };
> > >>
> > >> int arch_prctl (int option, union __arch_prctl_arg *arg);
> > >
> > > Or you could use a transparent union:
> > >
> > > $ cat arch_prctl.c
> > > union __attribute__((__transparent_union__)) __arch_prctl_arg {
> > > unsigned long set;
> > > unsigned long *get;
> > > };
> > >
> > > int my_arch_prctl(int op, union __arch_prctl_arg arg2);
> > >
> > > int
> > > main(void)
> > > {
> > > unsigned long u = 0;
> > >
> > > my_arch_prctl(1, u);
> > > my_arch_prctl(1, &u);
> > > }
> > > $ gcc -Wall -Wextra -S arch_prctl.c
> > > $
> >
> > Although we do use __transparent_union__ we have to still handle very old
> > compilers:
>
> We can hide the prototype for the older compilers. Softwares need to
> deal with the missing prototype anyway.
(avoiding source incompatibility via mismatching declarations is
perhaps a reason to just go with the kernel signature, since that's
what the folks i've seen `extern` this function themselves seem to
use...)
> > socket/sys/socket.h
> >
> > 51 /* This is the type we use for generic socket address arguments.
> > 52
> > 53 With GCC 2.7 and later, the funky union causes redeclarations or
> > 54 uses with any of the listed types to be allowed without complaint.
> > 55 G++ 2.7 does not support transparent unions so there we want the
> > 56 old-style declaration, too. */
> > 57 #if defined __cplusplus || !__GNUC_PREREQ (2, 7) || !defined __USE_GNU
> >
> > I guess this is not really required anymore.
> >
> > >
> > >> And for constants, it would require a x86_64 specific sys/prctl.h header
> > >> with either the has_include tricks to include the kernel one or to just
> > >> copy the kernel one (along with possible a test to check the sync with
> > >> kernel definitions).
> > >
> > > I think including would be simpler, if it's possible.
> > >
> > > I've seen some problems arise from copying kernel stuff in glibc
> > > headers, such as being unable to include both a some kernel and some
> > > glibc headers in the same program due to redefinitions.
> > >
> > >
> > > Have a lovely night!
> > > Alex
> > >
> > >
>
>
>
> --
> H.J.
