On Sat, Dec 08, 2018 at 12:12:49PM +0900, Sergey Senozhatsky wrote:

On (10/16/18 14:04), Sergey Senozhatsky wrote:
[..]

quoted

- The first entry point is console ->write() callback, which we call
  from printk(). A possible deadlock scenario there is:

  CPU0
	<NMI>
	spin_lock_irqsave(&port->lock, flags)      << deadlock
	serial_foo_write()
	call_console_drivers()
	console_unlock()
	console_flush_on_panic()
	panic()
	<NMI/>
	spin_lock_irqsave(&port->lock, flags)
	serial_foo_write()
	call_console_drivers()
	console_unlock()
	printk()
	...

[..]

quoted

- The rest (of entry points) requires a bit different handling.
  Let's take a look at the following backtrace:

  	CPU0
	<IRQ>
	spin_lock_irqsave(&port->lock, flags)      << deadlock
	serial_foo_write()
	call_console_drivers()
	console_unlock()
	printk()
	__queue_work()
	tty_flip_buffer_push()
	spin_lock_irqsave(&port->lock, flags)
	serial_foo_handle_IRQ()
	<IRQ/>

  Serial drivers invoke tons of core kernel functions - WQ, MM, etc. All
  of which may printk() in various cases. So we can't really just
  "remove those printk-s". The simples way to address this seems to be
  PRINTK_SAFE_CONTEXT_MASK.

serial/UART and printk guys, sorry to bother you, do you hate this
idea of removing console_driver->CORE KERNEL->printk->console_driver
deadlock path? Or is there any chance we can move forward?

If done in a sane manner, no objection from me.