Re: [PATCH] sysrq: Restore original console_loglevel when sysrq disabled

[Petr Mladek <pmladek@suse.com>]

On Fri 2019-01-11 22:07:29, Sergey Senozhatsky wrote:
> On (01/11/19 13:45), Petr Mladek wrote:
> > The sysrq header line is printed with an increased loglevel
> > to provide users some positive feedback.
> > 
> > The original loglevel is not restored when the sysrq operation
> > is disabled. This bug was introduced in 2.6.12 (pre-git-history)
> > by the commit ("Allow admin to enable only some of the Magic-Sysrq
> > functions").
> 
> 
> Good find, and the patch looks OK to me. A small comment below.
> FWIW,
> Reviewed-by: Sergey Senozhatsky <sergey.senozhatsky@gmail.com>

Thanks.

> A side note (nitpick, etc.); it's Friday night in here, I'm enjoying
> my beer; so maybe I'm wrong about the whole thing.
> 
> 
> > @@ -553,6 +553,7 @@ void __handle_sysrq(int key, bool check_mask)
> >  			op_p->handler(key);
> >  		} else {
> >  			pr_cont("This sysrq operation is disabled.\n");
> > +			console_loglevel = orig_log_level;
> >  		}
> 
> This looks a bit racy.
>
> When we do
> 
> 	printk("FOO\n");
> 	console_loglevel = XYZ;
> 
> We don't have any real guarantees that printk("FOO\n") will print
> anything straight ahead. It is possible that console_sem is already
> locked and the owner is preempted, so by the time the console_sem
> owner picks up that FOO\n messages, console_loglevel is back to
> orig_log_level and suppress_message_printing() will just tell us
> to skip the message.
> 
> Do we need pr_cont() there? Maybe we can just have a normal pr_err()
> which would always tell that "key" sysrq is disabled? (we also
> would need to change the error message a bit).

The same problem is with the sysrq header line. It uses the trick
with console_loglevel by intention. We want to show it but
it is not really an error message, see the commit
fb144adc517d9ebe8fd ("sysrq: add commentary on why we use
the console loglevel over using KERN_EMERG").

Best Regards,
Petr

PS: I am going to resend this patch as part of a patchset.
I was working on one more fix. It looked independent
and questionable. I wanted to send it as RFC separately
but there is a clash in the end...