On Fri, Jul 12, 2019 at 3:50 PM Laurent Vivier <laur...@vivier.eu> wrote:
> Le 12/07/2019 à 15:36, Arnd Bergmann a écrit :
> >> We don't do memcopy() but we set each field one by one, so the padding
> >> doesn't
> >> seem needed if we define correctly the user structure:
> >>
> >> struct target_timeval64 {
> >> abi_llong tv_sec;
> >> abi_long tv_usec;
> >> };
> >>
> >> and do something like:
> >>
> >> struct target_timeval64 *target_tv;
> >> struct timeval *host_tv;
> >> ...
> >> __put_user(host_tv->tv_sec, &target_tv->tv_sec);
> >> __put_user(host_tv->tv_usec, &target_tv->tv_usec);
> >> ...
> >
> > That still seems wrong. The user application has a definition
> > of 'timeval' that contains the padding, so your definition has
> > to match that.
>
> I don't find this definition with the padding. Where it is defined?
>
> We are at the syscall level, so structures are the ones provided by the
> target to the syscall, and they can be converted by the libc if the one
> from the user space differs.
glibc will have to create a definition that matches the kernel, which uses
struct __kernel_timespec {
__s64 tv_sec;
__s64 tv_nsec;
};
As posix requires tv_nsec to be 'long', you need padding between
tv_sec and tv_nsec to have a libc definition matching the kernel's
binary layout.
Arnd
