Re: [PATCH] panic: fix incomplete panic log in panic()

[Andrew Morton]

On Wed, 17 Oct 2012 18:44:32 +0800
Qing Z <njumical@xxxxxxxxx> wrote:

>          In ./drivers/video/fbmem.c, codes below cause issues:
> 	
>                             case FBIOPAN_DISPLAY:
> 			...
> 			console_lock();
> 			ret = fb_pan_display(info, &var);
> 			console_unlock();
> 			...
> 			break;
> 
> issue case 1:
> 1. core 0 call console_lock();
> 2. panic;
> ...
> 4. panic process done.
> Result: all panic log won't be printed.
> 
> issue case 2:
> 1. core 0 panic;
> 2. core 1 call console_lock();
> 3. core 0 call smp_send_stop(), core1 stop;
> 4. core 0 panic process done.
> Result: only little top part of panic log will be printed.
> 
> My soluiton according to your suggestions:
> 
> As you said, the first priority is to get oops message reliably
> delivered. I think we needn't care about console_sem when panic, just
> make sure we print the log imediately, so add
> sema_init(&console_sem,1) in bust_spinlocks(0), just like zap_locks()
> do.  It is safer than console_unlock() or up().

hm, I see.

> We can't add sema_init(..) in bust_spinlocks(1) due to issue case2,
> although the condition is rare. About issue case 2: should we avoid
> call console_lock() when panic?

Well, I think we do have infrastructure to support that:

+	if (!oops_in_progress)
		console_lock();

I haven't looked to see how practical that approach would be.


It would be better if we were to do

	if (oops_in_progress)
		console_trylock();
	else
		console_lock();

where console_trylock() would *try* to do a console_lock() but would
bail out if it was unable to immediately take the lock.  This is better
because most of the time, the oopsing CPU *will* lock the console and
will prevent other code from getting into the console code and messing
things up.

A problem with this approach is that it is very hard to test - the
"console_trylock failed" case will be rare.


I think it would be acceptable to just skip over the console_lock() if
oops_in_progress is set.  And if we skipped the console_lock(), we
should also skip the console_unlock().  So something like:

	bool console_unlock_needed = true;

	if (unlikely(oops_in_progress))
		console_unlock_needed = false;
	else
		console_lock();

	...

	if (console_unlock_needed)
		console_unlock();


> If we init console_sem in panic, old text may be flushed too, but
> should be before panic oops message. Also we can fix it by updating
> con_start("con_start = log_end") once panic happen, only log after
> panic will be printed.
--
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/