Re: [PATCH] linux: Use 'long int' for timespec tv_nsec on x32 (BZ #16437)

[Arnd Bergmann <arnd@arndb.de>]

On Fri, Dec 10, 2021 at 12:29 PM Florian Weimer <fweimer@redhat.com> wrote:
> > On Thu, Dec 9, 2021 at 8:35 PM Adhemerval Zanella via Libc-alpha
> > <libc-alpha@sourceware.org> wrote:
> >> On 09/12/2021 16:19, Florian Weimer wrote: * Adhemerval Zanella via Libc-alpha:
> >
> >> > The change is also not entirely ABI-compatible because existing code on
> >> > x32 depends on the implementation clearing the upper 32 bits, and if new
> >> > code copies around structs, it won't necessary copy the zero bits.
> >>
> >> Afaiu the high-bits clearing is strictly necessary on kABI boundary for
> >> set functions.  Some issues indeed might arise if user use timespec
> >> embedded in some function and mix objects built with different glibc
> >> version, but it is similar to 32-bit architectures with mixing time_t
> >> objects.
> >
> > The kernel's 32-bit ABI ignores the upper 32 bits of tv_nsec for structures
> > passed from user space, but zeroes them when returning a timespec to user
> > space.
>
> And that zeroing may get lost if the struct has padding.  That's my
> concern.

Which struct? I am talking about __kernel_timespec, which has
no padding on any architecture, and this is what gets used in the uabi
headers.

       Arnd