Re: [PATCH] tty: serial: men_z135_uart.c: Fix race between IRQ and set_termios()

[Johannes Thumshirn <jthumshirn@suse.de>]

Peter Hurley <peter@hurleysoftware.com> writes:

> On 08/04/2015 03:02 AM, Johannes Thumshirn wrote:
>> Peter Hurley <peter@hurleysoftware.com> writes:
>>> On 08/03/2015 09:58 AM, Johannes Thumshirn wrote:
>>>> Fix panic caused by a race between men_z135_intr() and men_z135_set_termios().
>>>>
>>>> men_z135_intr() and men_z135_set_termios() both hold the struct uart_port::lock
>>>> spinlock, but men_z135_intr() does a spin_lock_irqsave() and
>>>> men_z135_set_termios() does a normal spin_lock(), which can lead to a deadlock
>>>> when an interrupt is called while the lock is being helt by
>>>> men_z135_set_termios().
>>>
>>>
>>> The irq handler can and should use normal spin_lock()/unlock().
>> 
>> I always thought an irq handler _must_ use the irqsave versions. Good to
>> know that.
>
> Your irq handler does not need to protect itself from re-entrancy (by using
> the same irq handler for different irqs) and your serial driver doesn't support
> console (so can't be deadlocked by printk() usage either).
>

So once we have console support (I don't know if this is planned at
all), we must go back to the irqsave variant? But looking at the driver
again I have the feeling that the locking could be made more fine
grained (if this makes sense) and I saw two possible races, please
correct me if I'm wrong.

men_z135_intr() reads the status register and saves a copy in struct
men_z135_port::stat_reg. The men_z135_handle_lsr() and
men_z135_handle_modem_status() functions use this stat_reg value to get
to the LSR and MSR registers. But in the meanwhile the content of the
registers could have been changed by men_z135_intr() again. Is this a
problem or am I on the wrong track here?

>>> The set_termios() method should used spin_lock_irq(); there's no need to save the
>>> interrupt state because that method will never be called from interrupt context.
>>>
>>> So the 'flags' local can be dropped from the patch.
>> 
>> Given that the irqsave variant isn't needed that sounds reasonable.
>
> It's for a different reason; irqs will _always_ be on when your driver's
> set_termios() method is called. So you don't see to save the irq state, because
> you know it's always on. That's why you can use the spin_lock_irq()/spin_unlock_irq()
> version here.
>
>>> Also, the port lock is already initialized in uart_add_one_port() and should
>>> not be initialized by the probe() function.
>> 
>> OK, do you prefer (or better Greg and Jiri) prefer that change folded
>> into this patch or an extra patch?
>
> Separate patch please.

OK.

>
> I assume this deadlock fix will need to be pushed to -stable as well,
> yes?

I wasn't quite sure about this, I 1st had a CC stable for v4.0+ but
then removed it again before sending the patch. So I'll put it back in.

Thanks,
        Johannes