Re: [RFC PATCH v1 00/25] printk: new implementation
From: John Ogness
Date: Wed Feb 13 2019 - 09:15:59 EST
On 2019-02-13, Sergey Senozhatsky <sergey.senozhatsky.work@xxxxxxxxx> wrote:
>> - console_flush_on_panic() currently is a NOP. It is pretty clear how
>> this could be implemented if atomic_write was available. But if no
>> such console is registered, it is not clear what should be done. Is
>> this function really even needed?
> If you now rely on a fully preemptible printk kthread to flush
> pending logbuf messages, then console_flush_on_panic() is your
> only chance to see those pending logbuf messages on the serial
> console when the system dies.
Anything critical would have already been immediately print to the
emergency consoles. And if an emergency console was available,
console_flush_on_panic() could be a special case where _all_ unseen
messages (regardless of importance) are printed to the emergency
> Non-atomic consoles should become atomic once you call bust_spinlocks(1),
> this is what we currently have:
> bust_spinlocks(1) // sets oops_in_progress
> -> serial_driver_write()
> if (oops_in_progress)
> locked = spin_trylock_irqsave(&port->lock);
> if (locked)
I don't like bust_spinlocks() because drivers end up implementing
oops_in_progress with exactly that... ignoring their own locks. I prefer
consoles are provided with a locking mechanism that they can use to
support a separate NMI-safe write function. My series introduces
console_atomic_lock() for exactly this purpose.
But this doesn't help here. Here we are talking about a crashing system
that does _not_ have an emergency console. And in this case I would say
messages would be lost (just like they are now if all you have is a vt
console and it was busy).
I suppose we could keep the current bust_spinlocks() stuff for the
special case that there are no emergency consoles available. It's better
than nothing, but also not really reliable. Preferrably we figure out
how to implement write_atomic for all console drivers.