Date: Thu, 25 Mar 2004 16:05:51 -0500
From: "Bill Rugolsky Jr." <brugolsky@telemetry-investments.com>
To: Sridhar Samudrala <sri@us.ibm.com>
Cc: davem@redhat.com, jgarzik@pobox.com, linux-kernel@vger.kernel.org,
       netdev@oss.sgi.com
Subject: Re: [PATCH] Consolidate multiple implementations of jiffies-msecs conversions.
Message-ID: <20040325210551.GA23993@ti19.telemetry-investments.com>
Mail-Followup-To: "Bill Rugolsky Jr." <brugolsky@telemetry-investments.com>,
	Sridhar Samudrala <sri@us.ibm.com>, davem@redhat.com,
	jgarzik@pobox.com, linux-kernel@vger.kernel.org, netdev@oss.sgi.com
References: <Pine.LNX.4.58.0403251142110.3037@localhost.localdomain>
Mime-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <Pine.LNX.4.58.0403251142110.3037@localhost.localdomain>
User-Agent: Mutt/1.4.1i
Sender: linux-kernel-owner@vger.kernel.org
Content-Length: 2381
Lines: 57

On Thu, Mar 25, 2004 at 12:17:41PM -0800, Sridhar Samudrala wrote:
> The following patch to 2.6.5-rc2 consolidates 6 different implementations
> of msecs to jiffies and 3 different implementation of jiffies to msecs.
> All of them now use the generic msecs_to_jiffies() and jiffies_to_msecs()
> that are added to include/linux/time.h
 
I was inplementing precisely the same diff while cleaning up the
select/poll timeout logic, when I came across this lkml post by George
Anziger,

   http://marc.theaimsgroup.com/?l=linux-kernel&m=107772721007761&w=4

the relevant part of which is:

   As to small drifts of ~170 PPM, they are caused by code (ps I would 
   guess)
   that assumes that jiffies is exactly 1/HZ whereas it is NOT in the 2.6.*
   kernel.  The size of the jiffie that the kernel uses is returned by:

   struct timespec tv;
   :
   :
   clock_res(CLOCK_REALTIME, &tv);

I inferred from the above that a generic msec_to_jiffies()/jiffies_to_msec()
ought to use TICK_NSEC, as with the other routines in time.h.  For 32-bit
platforms the scaled arithmetic is simple; one has to be more careful
when BITS_PER_LONG == 64.

After e-mailing George about it, he replied:

   You might want to look at the code in time.h that does the jiffies to
   timespec conversion.  We did a lot of work to get that right for both 32 and
   64 bit platforms.  I don't think you really need more precision than we get
   on the 32 bit platforms (if I recall correctly it is in the 100 PPB range,
   yeah that is parts per BILLION).

   Unless this conversion is done a lot, I would just start with the timespec
   conversions and convert from / to using simple math.  As it is all power of
   10 stuff it should not have a precision problem.

He also supplied me with his test harness.

I haven't gotten around to doing this properly yet.  It seems that the
only place where precision is actually important (due to the possibility
of very long timeouts) is in poll/epoll, so perhaps it is best
to just code up a special version for them, as the simple version
becomes a no-op everywhere else for the default HZ==1000.

Regards,

	Bill Rugolsky
-
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/