On Thursday 15 May 2014 11:47:03 James Bottomley wrote: > On Wed, 2014-05-14 at 13:00 +0200, Arnd Bergmann wrote: > > On Tuesday 13 May 2014 22:35:08 Geert Uytterhoeven wrote: > > > > > > On Tue, May 13, 2014 at 9:32 PM, Arnd Bergmann <arnd@xxxxxxxx> wrote: > > > > I think we have three categories: > > > > > > Thanks for the list! > > > > > > > a) interfaces that uses relative time_t/timespec/timeval: > > > > b) interfaces that don't make sense for times in the past: > > > > > > > c) interfaces that require absolute times: > > > > - stat/lstat/fstatat/ > > > > - utime/utimes/futimesat > > > > > > > > These absolutely have to use something better than time_t > > > > both in user space and in the kernel so we can deal with > > > > old files. A lot of file systems need to be fixed as well so > > > > we can actually store the times, regardless of whether we > > > > are running a 32 or 64 bit kernel. > > > > > > So these are the ones we have to worry about. > > > It looks like they all involve I/O? Apart from the case of using block data > > > from the buffer cache, the 64-bit operations should disappear in the > > > actual I/O noise, right? > > > > Right. Also there have been proposals for a better 'stat' replacement > > for years, which would solve half of the interface problem for the > > file system interfaces. > > > > However, we also need to find a solution for category b), I only put > > them into a different category above because we can treat them > > differently in the kernel. For instance, we could use ktime_t for > > the kernel code in category b) and a new struct timespec64 for > > the times in struct inode. > > On the user interface side, using timespec64 would be a reasonable > > choice for both categories, because we already have two implementations > > of all those syscalls in order to handle 32-on-64 compat tasks, > > and we could use the same set of syscall implementations for time64-on-32. > > To step back a bit, is a 64 bit time_t actually a good solution on 32 > bits? Paying a 64 bit penalty on every time operation does seem a bit > overkill. Most eventual uses are either monotonic counting or relative > addition/subtraction. If we added an additional 32 bit quantity called > epoch, this would increase once every 68 years. Within the kernel, we > could do intelligent interpolation about what the epoch is, so in 2039 > if we get a low 32 bit time_t value, we assume epoch == 1, conversely if > we see a high one, we assume epoch == 0. We could add epoch on the end > of the syscalls and detect if it's not present and fill in an > interpolated value. Interfaces which truly want 64 bit time_t would get > it from epoch + 32 bit time_t. The main advantage that I see with 64-bit time_t is that a lot of user space already works with it, because NetBSD and OpenBSD use this, and all 64-bit Linux systems already work with this without changing the ABI. If we want a POSIX extension to cover a new ABI, this would be the most likely candidate. For Linux-only interfaces, the 64-bit nanosecond timestamps seem like a good alternative, and they would also speed up the 64-bit architectures because we can skip the normalization. As mentioned earlier, between kernel and user space it's probably best to avoid time_t and timeval completely and just use timespec64 or some other safe type, but there has to be a way to port user space that relies on time_t or timespec. Arnd -- To unsubscribe from this list: send the line "unsubscribe linux-arch" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
