[openbeosnetteam] Re: KDL visits...


> David wrote:
> > Do we think this might be due to dprintf???
> >
> > KERNEL PANIC: add_timer(001a6a60, 3000, 00129750, 2) already in
> > timer queue (tulip0_rx_thread, 00129777)!
> >
>
> And then added:
>
> > So, that one was due to dprintf...this one isn't...
> >
> > KERNEL PANIC: add_timer(001a6a60, 3000, 00129750, 2) already in
> > timer queue (idle thread 1, 00129777)!
>
> Seems to panic from dprintf() internal code and
> from idle thread code. Both (dprintf() and null_thread code) should use
> the kernel timers service, for serial port output support (dprintf) or
> simply snooze when idle...

Removing one of the dprintf's that was getting called a LOT cured the KDL
from dprintf. That was when the other one started.

>
> Anyway, AFAIK dprintf() is not interrupt-safe, only kprintf() is.
> The BeOS R5 kernel timers are interrupt-driven. Software interrupts, but
> still interrupts...
>
> The "Drivers" BeBook chapter state that calling cancel_timer() from
> the timer handler is not supported.
> I dunno if it was fixed in R5 kernel.
> See
http://bang.dhs.org/be/bebook/Drivers/KernelFunctions.html#cancel_timer()


>
> We ARE currently calling cancel_timer() from our arp timer
arp.c:arpq_run().

And we don't even build the arp code any more, let alone use it, so it won't
be that.

>
> [Oh, by the way, we never cancel the tcp timers tcp_slowtim and
tcp_fasttim.
> tcp.c:tcp_init()].

Correct. We don't need to until we stop using tcp.

> /me think that we could be hitting here some kernel timers internal
> table duplicate keys error: differents threads (00129777, always!)
> are callind add_timer() with same parameters! Strange.
> Looks like some cancel_timer()s are not really done
> (checking his return value could help make better diagnostic here).

That's what it says... As we never actually call cancel_timer not sure that
would help. I can't generate the error here on BONE so it'll have to wait
until I get home. I'll maybe try again in a while as BONE has more detail
when entering KDL than R5.

>
> Would it be difficult to use again the thread'based net_timers service
> contributed by Axel?
> They'll works in kernelland too, just a matter of calling
> spawn_kernel_thread() in right place...

It's possile, but seems like duplicating code Be has already provided...

david

Other related posts: