On Fri 2018-06-01 13:40:50, Sergey Senozhatsky wrote:
> On (05/31/18 14:21), Petr Mladek wrote:
> > >
> > > Upstream printk has no printing kthread. And we also run
> > > printk()->console_unlock() with disabled preemption.
> >
> > Yes, the comment was wrong
>
> Yes, that was the only thing I meant.
> I really didn't have any time to look at the patch yesterday, just
> commented on the most obvious thing.
Fair enough.
> > but the problem is real.
>
> Yep, could be. But not exactly the way it is described in the commit
> messages and the patch does not fully address the problem.
>
> The patch assumes that all those events happen sequentially. While
> in reality they can happen in parallel on different CPUs.
>
> Example:
>
> CPU0 CPU1
>
> set console verbose
>
> dump_backtrace()
> {
> // for (;;) print frames
> printk("%pS\n", frame0);
> printk("%pS\n", frame1);
> printk("%pS\n", frame2);
> printk("%pS\n", frame3);
> ... console_loglevel =
> CONSOLE_LOGLEVEL_SILENT;
> printk("%pS\n", frame12);
> printk("%pS\n", frame13);
> }
>
> Part of backtrace or the entire backtrace will be missed, because
> we read the global console_loglevel. The problem is still there.
[...]
> So I'd say that most likely the following scenarios can suffer:
>
> - NMI comes in, sets loglevel to X, printk-s some data, restores the
> loglevel back to Y
> - IRQ comes in [like sysrq, etc] comes in and does the same thing
> - software exception comes in and does the same thing [e.g. bust_spinlocks()
> at arch/s390/mm/fault.c]
My view is:
The race with another printk() (console_lock owner) is much more
likely than a race between two CPUs manipulating console_loglevel.
The proposed patch seems to be in the right direction. It is supposed
to fix the most likely scenario. We could block it and request full
solution but I wonder if it is worth it.
I am personally fine with this partial solution for now. We could
always make it better if people meet the other scenarios.
Best Regards,
Petr
