On Wed 2019-11-13 17:28:45, Jonathan Richardson wrote:
> Adding printk maintainers.
> This commit seems to have introduced the error:
>
> commit f92b070f2dc89a8ff1a0cc8b608e20abef894c7d
> Author: Petr Mladek <pmladek@xxxxxxxx>
> Date: Thu Sep 13 14:34:06 2018 +0200
>
> printk: Do not miss new messages when replaying the log
>
> If I checkout to the commit before
> (a06b0c82a049d34d4dc273e8692ed0894458c468), the console output is
> normal when registering 2 bootconsoles and 2 normal consoles. I've
> added the log for 4.19.0-rc3 for comparison (previous version was
> 5.1.0). I don't think this commit took into account that more than one
> console could be registered. When the second console is registered,
> 'console_seq >= exclusive_console_stop_seq' is true (both are 0) and
> exclusive_console is always set to NULL resulting in the log being
> replayed again to the uart8250 console:
This race should not happen because Both exclusive_console and
exclusive_console_stop_seq are manipulated under console_lock.
And the log is replayed before console_lock is released.
> /* Output to all consoles once old messages replayed. */
> if (unlikely(exclusive_console &&
> console_seq >= exclusive_console_stop_seq)) {
> exclusive_console = NULL;
> }
>
> I'm looking into it but any input is helpful. Thanks.
IMHO, the problem is that the log should not be replayed at all.
See the following code in register_console():
/*
* If we have a bootconsole, and are switching to a real console,
* don't print everything out again, since when the boot console, and
* the real console are the same physical device, it's annoying to
* see the beginning boot messages twice
*/
if (bcon && ((newcon->flags & (CON_CONSDEV | CON_BOOT)) == CON_CONSDEV))
newcon->flags &= ~CON_PRINTBUFFER;
I already see two problems there:
1. CON_PRINTBUFFER is cleared only when the new console has
CON_CONSDEV flag set. It is set only for the console
that is defined as the last on the command line.
It is a so-called preferred console.
2. bcon is set to the first console in console_drivers list.
It is the first registered boot console.
Sigh, this works for simple configuration. But it fails badly when
more different consoles are configured.
We should clear CON_PRINTBUFFER flag when the real console
replacing an already registered boot console is registered.
BTW: Similar bug is also at the end of register_console().
The boot consoles are unregistered only when the preferred
console is installed.
For a proper solution we would need to match boot and real
consoles that write messages into the physical device.
But I am afraid that there is no support for this.
con->match() callback compares the name defined on
the command line. And it has side effects (the matching
console is prepared for registration).
To be honest I am not much familiar with the device interface.
I am not sure if there is a way to detect the two drivers
use the same physical hardware.
Sigh, it is a huge historical mess. It would needed a lot
of work to clean it up.
Best Regards,
Petr
