[PATCH v7 0/3] y2038: Linux: Introduce __clock_settime64 function

[Joseph Myers]

On Tue, 17 Sep 2019, Lukasz Majewski wrote:

> Then if you replace the condition [1] with #if __TIMESIZE == 64 you
> would have:
> 
> struct timespec
> {
>   __time_t tv_sec;		/* Seconds.  */
>   __syscall_slong_t tv_nsec;    /* Nanoseconds.  */
> }

The *public* struct timespec (defined in 
time/bits/types/struct_timespec.h) should be changed for ports that define 
__TIMESIZE == 64 while __SYSCALL_WORDSIZE == 32.

That is, if __TIMESIZE == 64, and if __SYSCALL_WORDSIZE (if defined) is 32 
or __WORDSIZE (if __SYSCALL_WORDSIZE is not defined), then struct timespec 
needs endian-dependent padding (defined as an *unnamed* 32-bit bit-field, 
so that it gets ignored for initializers).  (This is the same padding as 
would be needed in the case where __TIMESIZE == 32 but _TIME_BITS=64 is 
defined, but _TIME_BITS=64 support for headers comes later.)

RV32 has got away without that change to struct timespec because it's 
little-endian, and as long as __time_t is 8-byte-aligned implicit padding 
works as well as explicit in the little-endian case.  If BE, or if 8-byte 
__time_t is only 4-byte-aligned in structs (and so the struct ends up as 
12-byte without explicit padding), there would be problems.  I think it's 
cleanest to make the padding explicit even in the cases where in fact 
implicit padding would give the same layout.

RV32 does not need any support for clearing the padding before passing 
struct timespec to the kernel, because that's only relevant for compat 
syscalls in Linux 5.1.0 to 5.1.4 and the RISC-V kernel doesn't yet have 
compat syscall support for running RV32 binaries under RV64 kernels.

-- 
Joseph S. Myers
joseph@codesourcery.com