CONFIG_PRINTK_TIME woes

[Luck, Tony]

It has been pointed out to me that ia64 doesn't boot
with CONFIG_PRINTK_TIME=y.  The issue is the call to
sched_clock() ... which on ia64 accesses some per-cpu
data to adjust for possible variations in processor
speed between different cpus.  Since the per-cpu page
is not set up for the first few printk() calls, we die.

Is this an issue on any other architecture?  Most versions
of sched_clock() seem to just scale jiffies into nanoseconds,
so I guess they don't.  s390, sparc64, x86 and x86_64 all
have more sophisticated versions but they don't appear to me
to have limitations on how early they might be called.

Possible solutions:

1) Fix ia64 version of sched_clock() to check whether
the per-cpu page hase been initialized before using it.

I don't like this solution as it will make sched_clock()
slower for its primary uses in kernel/sched.c ... none of
which can be called before the per-cpu area is initialized.


2) Add some test to the printk() path along the lines of:

  t = (can_call_sched_clock) ? sched_clock() : 0;

This is somewhat ugly ... perhaps too unpleasant for words in
that can_call_sched_clock is a per-cpu concept!

3) Make the printk() code get the time in some other way.

Using sched_clock() here seems wrong.  The ia64 version
has a big fat comment about not comparing the results of
two calls from different cpus ... but since printk() calls
can come from any cpu ... it seems likely that a user who
turns on PRINTK_TIME might subtract timestamps from two
lines to see how long it was between the messages.  This
could have significant error on a system that has been up
for a long time.

But I don't have a better suggestion from existing code.  I
assume that sched_clock() was picked as it is both fast and
lockless.

-Tony

-
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/