Re: Fw: [Bug 4559] New: cfq scheduler lockup: NMI oops while runningltp - 20050207 on 2.6.12-rc2-mm3 with kdump enabled

[Oleg Nesterov <oleg@tv-sign.ru>]

Andrew Morton wrote:
> 
>  http://bugme.osdl.org/show_bug.cgi?id=4559
> 
> Timer bug, I guess.

Yes, the new timer code is racy. Example:

spinlock_t LOCK;

void void timer_func()
{
	spin_lock(&LOCK);
}

timer_list TIMER = TIMER_INITIALIZER(timer_func);

-------------------------------------------------------------------
CPU_0					CPU_1

					add_timer(&TIMER);
spin_lock(&LOCK);
					__run_timers:
						sets ->running_timer = &TIMER;
						calls timer_func()
							waits for &LOCK

__mod_timer(&TIMER);
	/* Ensure the timer is serialized. */
	retries while ->running_timer == &TIMER


Many thanks to Maneesh Soni for his excellent analysis in
http://bugme.osdl.org/show_bug.cgi?id=4559.

Note that del_timer_sync has this problem too, but this
situation is forbidden by synchronization rules.

At the moment I don't have a proper solution.

One option is to change __mod_timer() so that it would not
switch ->base when the timer is already running. But this
would be behavioural change: currently __mod_timer() guarantees
that the timer would be armed on the local cpu.

I'll try to find a solution, but perhaps it's better to drop
this patch for now.

Oleg.