On Fri, 2005-07-29 at 16:31 -0700, Christoph Lameter wrote: > What you are dealing with is a machine that is using ITC as a time bases. > That is a special case.
The default time source for ia64 systems is a "special case"? 4 socket and smaller boxes typically do not have any other time source. > The fix should not affect machines that have a > proper time source. More below. You can circumvent the compensation for > ITC inaccuracies by specifying "nojitter" on the kernel command if you are > willing to take the risk of slightly inaccurate time. And what if you don't have any HPET and aren't willing to take that risk? We need a solution that works with all time sources. A system with the default time source should not hang or have unreasonable delays with the standard setup. Why is a synchronized ITC driven from a common clock such a terrible time source for small systems? > Well get a proper time source and do not use ITC for a time source in an > SMP system. Got HPET hardware? No, HPET on small boxes is unnecessary, we should be able to come up with something that can effectively use the ITC. Does a seqlock really make sense for the do_gettimeofday() path? This problem arises because a reader of time is actually updating and writing a part of time. It would certainly be too much overhead to obtain a write lock for every gettimeofday(), but to avoid that we have to put some kind of contention avoidance in the path. I don't know whether that should be some kind of back-off algorithm at the point of contention w/ the cmpxchg or up higher with a new seqlock read entry point that blocks when a write is in progress. In any case, I think we need to focus on a solution that works well on all systems, not just those with extra timer hardware. Thanks, Alex - To unsubscribe from this list: send the line "unsubscribe linux-kernel" in the body of a message to [EMAIL PROTECTED] More majordomo info at http://vger.kernel.org/majordomo-info.html Please read the FAQ at http://www.tux.org/lkml/