Andy Lutomirski
2018-Sep-18 14:01 UTC
[patch 09/11] x86/vdso: Simplify the invalid vclock case
> On Sep 18, 2018, at 12:52 AM, Thomas Gleixner <tglx at linutronix.de> wrote: > >> On Mon, 17 Sep 2018, John Stultz wrote: >>> On Mon, Sep 17, 2018 at 12:25 PM, Andy Lutomirski <luto at kernel.org> wrote: >>> Also, I'm not entirely convinced that this "last" thing is needed at >>> all. John, what's the scenario under which we need it? >> >> So my memory is probably a bit foggy, but I recall that as we >> accelerated gettimeofday, we found that even on systems that claimed >> to have synced TSCs, they were actually just slightly out of sync. >> Enough that right after cycles_last had been updated, a read on >> another cpu could come in just behind cycles_last, resulting in a >> negative interval causing lots of havoc. >> >> So the sanity check is needed to avoid that case. > > Your memory serves you right. That's indeed observable on CPUs which > lack TSC_ADJUST. > > @Andy: Welcome to the wonderful world of TSC. >Do we do better if we use signed arithmetic for the whole calculation? Then a small backwards movement would result in a small backwards result. Or we could offset everything so that we?d have to go back several hundred ms before we cross zero.
Thomas Gleixner
2018-Sep-18 22:46 UTC
[patch 09/11] x86/vdso: Simplify the invalid vclock case
On Tue, 18 Sep 2018, Andy Lutomirski wrote:> > On Sep 18, 2018, at 12:52 AM, Thomas Gleixner <tglx at linutronix.de> wrote: > > > >> On Mon, 17 Sep 2018, John Stultz wrote: > >>> On Mon, Sep 17, 2018 at 12:25 PM, Andy Lutomirski <luto at kernel.org> wrote: > >>> Also, I'm not entirely convinced that this "last" thing is needed at > >>> all. John, what's the scenario under which we need it? > >> > >> So my memory is probably a bit foggy, but I recall that as we > >> accelerated gettimeofday, we found that even on systems that claimed > >> to have synced TSCs, they were actually just slightly out of sync. > >> Enough that right after cycles_last had been updated, a read on > >> another cpu could come in just behind cycles_last, resulting in a > >> negative interval causing lots of havoc. > >> > >> So the sanity check is needed to avoid that case. > > > > Your memory serves you right. That's indeed observable on CPUs which > > lack TSC_ADJUST. > > > > @Andy: Welcome to the wonderful world of TSC. > > > > Do we do better if we use signed arithmetic for the whole calculation? > Then a small backwards movement would result in a small backwards result. > Or we could offset everything so that we?d have to go back several > hundred ms before we cross zero.That would be probably the better solution as signed math would be problematic when the resulting ns value becomes negative. As the delta is really small, otherwise the TSC sync check would have caught it, the caller should never be able to observe time going backwards. I'll have a look into that. It needs some thought vs. the fractional part of the base time, but it should be not rocket science to get that correct. Famous last words... Thanks, tglx
Andy Lutomirski
2018-Sep-18 23:03 UTC
[patch 09/11] x86/vdso: Simplify the invalid vclock case
> On Sep 18, 2018, at 3:46 PM, Thomas Gleixner <tglx at linutronix.de> wrote: > > On Tue, 18 Sep 2018, Andy Lutomirski wrote: >>> On Sep 18, 2018, at 12:52 AM, Thomas Gleixner <tglx at linutronix.de> wrote: >>> >>>>> On Mon, 17 Sep 2018, John Stultz wrote: >>>>> On Mon, Sep 17, 2018 at 12:25 PM, Andy Lutomirski <luto at kernel.org> wrote: >>>>> Also, I'm not entirely convinced that this "last" thing is needed at >>>>> all. John, what's the scenario under which we need it? >>>> >>>> So my memory is probably a bit foggy, but I recall that as we >>>> accelerated gettimeofday, we found that even on systems that claimed >>>> to have synced TSCs, they were actually just slightly out of sync. >>>> Enough that right after cycles_last had been updated, a read on >>>> another cpu could come in just behind cycles_last, resulting in a >>>> negative interval causing lots of havoc. >>>> >>>> So the sanity check is needed to avoid that case. >>> >>> Your memory serves you right. That's indeed observable on CPUs which >>> lack TSC_ADJUST. >>> >>> @Andy: Welcome to the wonderful world of TSC. >>> >> >> Do we do better if we use signed arithmetic for the whole calculation? >> Then a small backwards movement would result in a small backwards result. >> Or we could offset everything so that we?d have to go back several >> hundred ms before we cross zero. > > That would be probably the better solution as signed math would be > problematic when the resulting ns value becomes negative. As the delta is > really small, otherwise the TSC sync check would have caught it, the caller > should never be able to observe time going backwards. > > I'll have a look into that. It needs some thought vs. the fractional part > of the base time, but it should be not rocket science to get that > correct. Famous last words... >It?s also fiddly to tune. If you offset it too much, then the fancy divide-by-repeated-subtraction loop will hurt more than the comparison to last.
Thomas Gleixner
2018-Sep-19 13:29 UTC
[patch 09/11] x86/vdso: Simplify the invalid vclock case
On Wed, 19 Sep 2018, Rasmus Villemoes wrote:> On 2018-09-19 00:46, Thomas Gleixner wrote: > > On Tue, 18 Sep 2018, Andy Lutomirski wrote: > >>> > >> > >> Do we do better if we use signed arithmetic for the whole calculation? > >> Then a small backwards movement would result in a small backwards result. > >> Or we could offset everything so that we?d have to go back several > >> hundred ms before we cross zero. > > > > That would be probably the better solution as signed math would be > > problematic when the resulting ns value becomes negative. As the delta is > > really small, otherwise the TSC sync check would have caught it, the caller > > should never be able to observe time going backwards. > > > > I'll have a look into that. It needs some thought vs. the fractional part > > of the base time, but it should be not rocket science to get that > > correct. Famous last words... > > Does the sentinel need to be U64_MAX? What if vgetcyc and its minions > returned gtod->cycle_last-1 (for some value of 1), and the caller just > does "if ((s64)cycles - (s64)last < 0) return fallback; ns +> (cycles-last)* ...". That should just be a "sub ; js ; ". It's an extra > load of ->cycle_last, but only on the path where we're heading for the > fallback anyway. The value of 1 can be adjusted so that in the "js" > path, we could detect and accept an rdtsc_ordered() call that's just a > few 10s of cycles behind last and treat that as 0 and continue back on > the normal path. But maybe it's hard to get gcc to generate the expected > code.I played around with a lot of variants and GCC generates all kinds of interesting ASM. And at some point optimizing that math code is not buying anything because the LFENCE before RDTSC is dominating all of it. Thanks, tglx
Maybe Matching Threads
- [patch 09/11] x86/vdso: Simplify the invalid vclock case
- [patch 09/11] x86/vdso: Simplify the invalid vclock case
- [patch 09/11] x86/vdso: Simplify the invalid vclock case
- [patch 10/11] x86/vdso: Move cycle_last handling into the caller
- [patch 09/11] x86/vdso: Simplify the invalid vclock case