Re: CONFIG_PRINTK_TIME and initial value for jiffies?

[Sergei Shtylyov <sshtylyov@ru.mvista.com>]

Hello.

Herbert Valerio Riedel wrote:

> On Tue, 2006-05-16 at 01:35 +0400, Sergei Shtylyov wrote:

>>>>I just switched to 2.6.16.16 from 2.6.14 on a Au1550. I enabled
>>>>CONFIG_PRINTK_TIME, and for some reason jiffies doesn't start out near
>>>>zero like it does on x86. The first printk() always seems to have a
>>>>time of 4284667.296000.
>>
>>>>jiffies_64 and wall_jiffies gets initialized to INITIAL_JIFFIES, but
>>>>I'm not sure where jiffies is initialized. INITIAL_JIFFIES is -300*HZ
>>>>(with some weird casting)

>>    Yes, the casting is weird. I somewat doubt that:

>>#define INITIAL_JIFFIES ((unsigned long)(unsigned int) (-300*HZ))

>>u64 jiffies_64 = INITIAL_JIFFIES;

>>can do the trick of wrapping around 5 mins after boot on x86... :-/

> jfyi, starting with an offset of -300 seconds is done on purpose, to
> expose bugs in drivers which don't handle wrapping of the jiffies;

    Oh, thank you. I've read that in the source code. :-)

> and the trick to get printk to start at offset 0 is either define a
> arch-specific printk_clock() function (it's a weak symbol in
> kernel/printk.c) or like more drivers to it, to provide a sched_clock()
> (which is used by the default printk_clock() function) implementation
> which starts at offset 0...

    sched_clock() defined in arch/i386/kernel/timers/timer_tsc.c can hardly 
provide 0-based time if it's using TSC (at least I can't see where the TSC is 
cleared). Even if it's not using TSC, jiffies_64 is not 0-based as we saw, and 
neither it's set to -300 secs because of the double cast to ulong and then to 
u64 which should clear the high word. Probably something somewhere clears TSC 
but I can see the related code only in arch/i386/kernel/smpboot.c...

> regards,
> hvr

WBR, Sergei