Re: [PATCH] more reliable system timer for SC1100 CPU
From: George Anzinger
Date: Fri Mar 11 2005 - 17:11:41 EST
Ted Phelps wrote:
First, procedure... patches should be *.patch and not compressed. If too long
they need to be broken up. Lately, folks have said they should be inline in the
email text, but watch out for your mailer doing UGLY things with white space.
Hello,
The attached patch is an attempt to work around the buggy timestamp
counter on the NatSemi SC1100 CPU by using the on-board 27MHz
high-resolution timer as an alternative time source. It should,
in theory, work with any of the SCx200 CPUs as well, though I have
been unable to test this. I have tested it fairly thoroughly with NTP
on an SC1100 and it seems to behave sanely.
That said, there are three things about it that I'm not entirely
comfortable with:
(1) The high-resolution timer is driven by a separate crystal than the
CPU's timer interrupt, and on the SC1100 I have access to, it's
consistently slower. I've found that it is necessary to
periodically *decrement* the jiffies_64 counter in mark_offset in
order to make gettimeofday produce anything reasonable. In
practice jiffies_64 is incremented again in do_timer before
anything else reads it, so the net effect is minimal.
I don't think this is what your seeing. As I read the code, if an interrupt
gets delayed and the next one is not, you will determine that you should
decrement jiffies. Interrupts DO get delayed. This counter is only being used
to cover the jiffie to jiffie time. I suspect that any systemic errors such as
different rocks are not really important (but drift needs to be accounted for,
see below).
The better thing to do here is to figure some arbitrary start time when a
jiffies edge is "close" to the actually interrupt time and use the counter time
at that time as the "base" time. Each jiffie you then bump this by the counts
per jiffie. (By the way, this should be calculated using TICK_NSEC (nsecs per
tick) and NOT HZ. TICK_NSEC accounts for the fact that the PIT does not produce
exactly 1/HZ ticks.)
In addition to this, at each interrupt, to account for drift, I have been using
code that, on each interrupt, checks if it is early (i.e.:
base + ticks_per_jiffy > now) if so adjust base to make it on time. If it is
late, I keep the minimum amount it is late for several ticks and then adjust
base to make it on time. This ends up making small changes in "base" to account
for any drift. It also ends up ignoring occasional late times caused by normal
interrupt latency. If it is late by over a tick, jiffies is adjusted for the
lost tick. (All this code is in the high-res-timers patch, see signature.)
Do note this assumes (and IMHO rightly so) that the PIT is the system time gold
standard.
George
(2) The 27MHz timer is accessed via the PCI bus, which is not
available when the system clock is initialized. To work around
this, I've written the init function to always fail so that
loops_per_jiffy is computed using another timer (the TSC in my
case). Once the high-resolution timer is accessible, the kernel
will switch to using it to compute gettimeofday and the monotonic
clock, but still use the original timer's delay function. This
is somewhat kludgy, but I can't see a cleaner way.
(3) The timer depends on CONFIG_SCx200, which appears later in the
configuration hierarchy to the timers, and in an entirely
different part. For now I've kept its Kconfig with the other
timers, but I'm not entirely happy with this choice.
The patch is against linux-2.6.11-mm2 as it relies on the
'determine-scx200-cb-address-at-run-time.patch' patch which has not
made it into in the mainline.
Please CC me if you reply as I'm not subscribed to LKML.
Cheers,
-Ted
--
George Anzinger george@xxxxxxxxxx
High-res-timers: http://sourceforge.net/projects/high-res-timers/
-
To unsubscribe from this list: send the line "unsubscribe linux-kernel" in
the body of a message to majordomo@xxxxxxxxxxxxxxx
More majordomo info at http://vger.kernel.org/majordomo-info.html
Please read the FAQ at http://www.tux.org/lkml/