lkml.org 
[lkml]   [2006]   [Nov]   [24]   [last100]   RSS Feed
Views: [wrap][no wrap]   [headers]  [forward] 
 
Messages in this thread
/
From
SubjectRe: [patch] x86: unify/rewrite SMP TSC sync code
Date
On Friday 24 November 2006 21:46, Ingo Molnar wrote:
>
> * Andi Kleen <ak@suse.de> wrote:
>
> >
> > > yeah - the main new bit for x86-64 is the unconditional check for time
> > > warps. We shouldnt (and cannot) really trust the CPU and the board/BIOS
> > > getting it right. There were always some motherboards using Intel CPUs
> > > that had the TSCs wrong.
> >
> > In the 64bit capable generation I don't know of any run in spec
> > (except for multinode systems and there was one overclocked system
> > where the cores got unsync, but overclocking is an operator error)
>
> i have one (Intel based), 64-bit, fully in spec, which is off by
> ~3000-4000 cycles. So it happens.

More details?

> I was in fact surprised when i noticed that you removed the
> unconditional TSC check that i put there years ago

I removed it because you pointed out that it usually caused
trouble on Intel systems: we would always detect errors due to measurement errors
and then make things worse by trying to fix it.

But you're right it might have been better to keep
a check with a threshold to catch totally broken cases.

> but which apps are using RDTSC natively? Trapping isnt too good i agree

The only sure way would be to trap+printk -- but from previous
user complaints it's a substantial number.

> - if then we should remove it from the CPU features and hence apps wont
> (or shouldnt) use it.

I doubt the majority checks any cpu features first ...

>
> > > nor can the TSC really be synced up properly in the hotplug CPU
> > > case, after the fact - what if the app already read out an older TSC
> > > value and a new CPU is added. If the TSC isnt sync on SMP then it
> > > quickly gets pretty messy, and we should rather take a look at /why/
> > > these apps are using RDTSC.
> >
> > Because gettimeofday is too slow.
>
> as i indicated it in another discussion, i can fix that. Next patch will
> be that.

Well I hope it's not making it HZ resolution. As noted earlier we tried
that already and it didn't work (it violates the "forward monotonity"
that is commonly expected)

Ok I could imagine it making sense as a new CLOCK_FASTBUTLOUSYRESOLUTION timer in
clock_gettime() [together with the new vdso fastpath], but not as default.

-Andi
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@vger.kernel.org
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/

\
 
 \ /
  Last update: 2006-11-24 22:09    [W:0.040 / U:0.480 seconds]
©2003-2020 Jasper Spaans|hosted at Digital Ocean and TransIP|Read the blog|Advertise on this site