Skip site navigation (1)Skip section navigation (2)
Date:      Sun, 2 Dec 2007 07:53:34 -0800
From:      Luigi Rizzo <rizzo@icir.org>
To:        Poul-Henning Kamp <phk@phk.freebsd.dk>
Cc:        arch@freebsd.org
Subject:   Re: New "timeout" api, to replace callout
Message-ID:  <20071202075334.A11104@xorpc.icir.org>
In-Reply-To: <19518.1196609689@critter.freebsd.dk>; from phk@phk.freebsd.dk on Sun, Dec 02, 2007 at 03:34:49PM %2B0000
References:  <20071202072658.A9217@xorpc.icir.org> <19518.1196609689@critter.freebsd.dk>

next in thread | previous in thread | raw e-mail | index | archive | help
On Sun, Dec 02, 2007 at 03:34:49PM +0000, Poul-Henning Kamp wrote:
> In message <20071202072658.A9217@xorpc.icir.org>, Luigi Rizzo writes:
> >On Sun, Dec 02, 2007 at 03:08:41PM +0000, Poul-Henning Kamp wrote:
> 
> >> Using a deadline timer based in the HPET, the timeout can be scheduled
> >> to any 1/14318181th of a second
> >
> >This must obviously come for a price or there would be no reason to
> >provide 'millisecond' and 'second' precisions that you suggested.
> >I agree that the API should not assume that the hardware is based on
> >a periodic timer, but in the same way it should not assume that
> >the hardware can generate arbitrary timeouts.
> 
> If the client code needs this finegrained control, the trick is to
> extend the API to be able to provide this information somehow.
> 
> But before designing that, we need to have concrete examples 
> to talk about.
> 
> >"periodically with whatever period is convenient to you" is
> >a legitimate request that client code might have.
> >(I understand that this concept of time is very southern european :)
> 
> I disagree, that case, should be "with a period between N and M,
> at your choice".

yes surely this is better. But since your initial interface only
had one parameter for this i tried to stick with that.

> And yes, it may be that we need to extend the API for that, but
> please give a concrete example so we know what we're talking about.

DEVICE_POLLING is one example. You poll (and do resource allocation)
to avoid livelock generated by too many interrupts, but don't know
in advance what's the correct polling rate so rely on the system administrator
to set HZ appropriately.

(the same might apply for devices with broken or missing interrupt hardware
so you have to poll. You'd trash them in an ideal world, but sometimes
you have to live with them).

dummynet is another one - in principle you'd like precise timing but
with thousands of active pipes you probably can't afford that and
so again rely on some coarser resolution which is synchronized with
some system task.

> >It wouldn't make sense to do poll-based event handling at the same
> >rate on a 133MHz soekris or on a 3 GHz multicore cpu.
> 
> Hz is the same for those two, so today we poll at the same rate.

but you'd bump HZ to 2k or 4k up on a modern CPU, and lower to 500
on a soekris.

> >In any case: in terms of API it costs absolutely nothing to support 
> >system-dependent resolutions instead of absolute ones, so 
> >i don't quite understand the opposition.
> 
> I'm applying Jim Gettys 3. rule from the X11 project:
> 
> 	3.The only thing worse than generalizing from one example
> 	  is generalizing from no examples at all.

for what matters that would apply to the ns/us/ms/s resolutions that
you proposed. They may be intuitive units but it doesn't mean that they are
the most appropriate ones.

Anyways, I think you got my point.

cheers
luigi



Want to link to this message? Use this URL: <https://mail-archive.FreeBSD.org/cgi/mid.cgi?20071202075334.A11104>