[RFC PATCH 0/3] Add accurate boot timing to a Linux system

Hi Matthieu,

On Mon, Sep 26, 2011 at 12:53 AM, Matthieu CASTET
[off-list ref] wrote:

Hi,

Russell King - ARM Linux a ?crit :

quoted

On Fri, Sep 23, 2011 at 04:03:15PM -0700, Simon Glass wrote:

quoted

An accurate timer is required to make the numbers meaningful. Many
modern platforms have a microsecond timer. This patch set uses a
function called timer_get_us() to read the timer.

Not another 'get a time value' function. ?Why do we need soo many?
We have - at least:

ktime_get (and various flavours of it)
do_gettimeofday
getnstimeofday
sched_clock

Do we really need yet another one which will have to be multiplexed
amongst platforms, requiring scaling and so forth from whatever the
platform provides?

Remember that ARM timers are virtually all MMIO mapped, which means
they don't work during early kernel bringup when the MMU mappings for
the hardware have not been setup. ?(That's the reason stuff like
sched_clock for printk doesn't work early.)

Doesn't cortexA-8 (and A9 ?) have a cycle counter that can be read by
coprocessor 15 ?

Couldn't we use that counter for early stuff on those architectures ?

Yes we could, but we need to be careful that we use a 'time of day'
counter, not dependent on clock speed, CPU load and the like. See my
other email for what I see as the ideal timer. Obviously this will
have to be architecture-specific, but I hope for a simple fallback to
something available on all architectures, with perhaps the ability to
use better timers available on specific architectures which implement
them.

Regards,
Simon

Matthieu