tree cfa185f249c166f8302513a6b57432bfa11b53c8
parent 79f08d9ed217318b4f325b7fcf8f26dcd0e41689
author Uwe Kleine-König <u.kleine-koenig@pengutronix.de> 1391423479 +0100
committer Uwe Kleine-König <u.kleine-koenig@pengutronix.de> 1392754999 +0100

ARM: ixp4xx: fix timer latch calculation

In commit f0402f9b4711 ("ARM: ixp4xx: stop using <mach/timex.h>")
I didn't intend to implement a functional change, but as Olof noticed I
failed---at least a bit. Before this commit the following was used to
determine the latch value used:

	#define IXP4XX_TIMER_FREQ 66666000
	#define CLOCK_TICK_RATE \
		(((IXP4XX_TIMER_FREQ / HZ & ~IXP4XX_OST_RELOAD_MASK) + 1) * HZ)
	#define LATCH ((CLOCK_TICK_RATE + HZ/2) / HZ)

The complicated calculation was done "b/c the timer register ignores the
bottom 2 bits of the LATCH value." With HZ=100 CLOCK_TICK_RATE used to
calculate to 66666100 and so LATCH to 666661. In ixp4xx_set_mode the
term

	LATCH & ~IXP4XX_OST_RELOAD_MASK

was used to write to the relevant register (with IXP4XX_OST_RELOAD_MASK
being 3) and so effectively 666660 was used.

In commit f0402f9b4711 I translated that to:

	#define IXP4XX_TIMER_FREQ 66666000
	#define IXP4XX_LATCH DIV_ROUND_CLOSEST(IXP4XX_TIMER_FREQ, HZ)

which results in the same register writes, but still doesn't bear in
mind that the two least significant bits cannot be specified (which is
relevant only when HZ or IXP4XX_TIMER_FREQ are changed).

Instead of reverting back to the old approach use a more obvious and
also more correct way to calculate LATCH. (Regarding the more
correct claim: With IXP4XX_TIMER_FREQ == 66665999, the old code resulted
in LATCH = 666657 corresponding to a cycle time of 0.009999940149400597
seconds (error: -6.0e-8 s) while the new approach results in LATCH =
666660 and so a cycle time of 0.010000000150001503 seconds
(error: 1.5e-10 s).)

Fixes: f0402f9b4711 ("ARM: ixp4xx: stop using <mach/timex.h>")
Signed-off-by: Uwe Kleine-König <u.kleine-koenig@pengutronix.de>