Messages in this thread | | | Date | Sun, 05 Jun 2011 22:15:27 +0200 | From | Arne Jansen <> | Subject | Re: [debug patch] printk: Add a printk killswitch to robustify NMI watchdog messages |
| |
On 05.06.2011 21:44, Ingo Molnar wrote: > > * Arne Jansen<lists@die-jansens.de> wrote: > >> From the timing I see I'd guess it has something to do with the >> scheduler kicking in during printk. I'm neither familiar with the >> printk code nor with the scheduler. > > Yeah, that's the well-known wake-up of klogd: > > void console_unlock(void) > { > ... > up(&console_sem); > > actually ... that's not the klogd wake-up at all (!). I so suck today > at bug analysis :-) > > It's the console lock()/unlock() sequence, and guess what does it: > > drivers/tty/tty_io.c: console_lock(); > drivers/tty/vt/selection.c: console_lock(); > > and the vt.c code in a dozen places. > > So maybe it's some sort of tty related memory corruption that was > made *visible* via the extra assert that the scheduler is doing? The > pi_list is embedded in task struct. > > This would explain why only printk() triggers it and other wakeup > patterns not. > > Now, i don't really like this theory either. Why is there no other > type of corruption? And exactly why did only the task_struct::pi_lock > field get corrupted while nearby fields not? Also, none of the fields > near pi_lock are even remotely tty related.
Can lockdep just get confused by the lockdep_off/on calls in printk while scheduling is allowed? There aren't many users of lockdep_off().
I'll can try again tomorrow to get a dump of all logs from the watchdog, but enough for today...
| |