Re: UART/TTY console deadlock

[Sergey Senozhatsky]

On (20/07/02 14:12), Sergey Senozhatsky wrote:
> Date: Thu, 2 Jul 2020 14:12:13 +0900
> From: Sergey Senozhatsky <sergey.senozhatsky@xxxxxxxxx>
> To: Tony Lindgren <tony@xxxxxxxxxxx>
> Cc: Sergey Senozhatsky <sergey.senozhatsky@xxxxxxxxx>, Petr Mladek
>  <pmladek@xxxxxxxx>, Andy Shevchenko <andy.shevchenko@xxxxxxxxx>, Raul
>  Rangel <rrangel@xxxxxxxxxx>, Sergey Senozhatsky
>  <sergey.senozhatsky.work@xxxxxxxxx>, linux-kernel
>  <linux-kernel@xxxxxxxxxxxxxxx>, Greg Kroah-Hartman
>  <gregkh@xxxxxxxxxxxxxxxxxxx>, Andy Shevchenko
>  <andriy.shevchenko@xxxxxxxxxxxxxxx>, kurt@xxxxxxxxxxxxx, "S, Shirish"
>  <Shirish.S@xxxxxxx>, Peter Zijlstra <peterz@xxxxxxxxxxxxx>, John Ogness
>  <john.ogness@xxxxxxxxxxxxx>, Steven Rostedt <rostedt@xxxxxxxxxxx>
> Subject: Re: UART/TTY console deadlock
> Message-ID: <20200702051213.GB3450@xxxxxxxxxxxxxxxxxxxxxxxx>
> 
> On (20/06/30 11:02), Tony Lindgren wrote:
> > This conditional disable for irq_shared does not look nice to me
> > from the other device point of view :)
> > 
> > Would it be possible to just set up te dummy interrupt handler
> > for the startup, then change it back afterwards? See for example
> > omap8250_no_handle_irq().
> 
> I think we can do it. serial8250_do_startup() and irq handler take
> port->lock, so they should be synchronized.

Hmm, hold on. Why does it disable IRQ in the first place? IRQ handlers
should grab the port->lock. So if there is already running IRQ, then
serial8250_do_startup() will wait until IRQ handler unlocks the port->lock.
If serial8250_do_startup() grabs the port->lock first, then IRQ will wait
for serial8250_do_startup() to unlock it. serial8250_do_startup() does
not release the port->unlock until its done:

		spin_lock_irqsave(&port->lock, flags);

		wait_for_xmitr(up, UART_LSR_THRE);
		serial_port_out_sync(port, UART_IER, UART_IER_THRI);
		udelay(1); /* allow THRE to set */
		iir1 = serial_port_in(port, UART_IIR);
		serial_port_out(port, UART_IER, 0);
		serial_port_out_sync(port, UART_IER, UART_IER_THRI);
		udelay(1); /* allow a working UART time to re-assert THRE */
		iir = serial_port_in(port, UART_IIR);
		serial_port_out(port, UART_IER, 0);

		spin_unlock_irqrestore(&port->lock, flags);

so IRQ will not see the inconsistent device state.

What exactly is the purpose of disable_irq_nosync()? Can we just remove
disable_irq_nosync()/enable_irq() instead? Are there any IRQ handlers
that don't acquire the port->lock?

	-ss