Re: [PATCH printk-rework 08/14] printk: add syslog_lock

[Helge Deller <deller@gmx.de>]

On 2/23/21 3:23 PM, Petr Mladek wrote:
> On Tue 2021-02-23 13:22:22, Helge Deller wrote:
>> On 2/22/21 5:28 PM, Petr Mladek wrote:
>>> On Sun 2021-02-21 22:39:42, Helge Deller wrote:
>>>> On 2/19/21 5:33 PM, John Ogness wrote:
>>>>> Added CC: linux-parisc@vger.kernel.org
>>>>>
>>>>> On 2021-02-19, John Ogness <john.ogness@linutronix.de> wrote:
>>>>>>>> diff --git a/kernel/printk/printk.c b/kernel/printk/printk.c
>>>>>>>> index 20c21a25143d..401df370832b 100644
>>>>>>>> --- a/kernel/printk/printk.c
>>>>>>>> +++ b/kernel/printk/printk.c
>>>>>>>> +/* Return a consistent copy of @syslog_seq. */
>>>>>>>> +static u64 read_syslog_seq_irq(void)
>>>>>>>> +{
>>>>>>>> +	u64 seq;
>>>>>>>> +
>>>>>>>> +	raw_spin_lock_irq(&syslog_lock);
>>>>>>>> +	seq = syslog_seq;
>>>>>>>> +	raw_spin_unlock_irq(&syslog_lock);
>>>>>>>
>>>>>>> Is there any particular reason to disable interrupts here?
>>>>>>>
>>>>> I found a possible call chain in interrupt context. From arch/parisc
>>>>> there is the interrupt handler:
>>>>>
>>>> Yes, handle_interruption() is the irq handler, running with irqs off.
>>>> HPMC is the crash handler - it's called when the kernel will stop
>>>> anyway. pdc_console is a very basic firmware console which prints
>>>> the last messages before the machine halts on fatal errors.
>>>> So, this code it's not the typical use case....
>>>
>>> Thanks for information.
>>>
>>> Is this code supposed to work only during early boot or anytime,
>>> please?
>>
>> No.
>> It's only called when the kernel completely crashes, when all
>> spinlocks should get busted and so on.
>> It's the emergency way to get some info out at least.
>
> OK.
>
>>> Note that it is not safe because register_console() takes
>>> console_lock() which is a sleeping lock.
>>
>> As I said, in that stage the plan is to bust all spinlocks.
>
> Just to be sure. Note that that register_console() does not bust
> console_lock in panic.

Ok.

> bust_spinlocks() just increments oops_in_progress counter. It has
> effect only when the caller checks this variable and use trylock
> when it is set. For example, see serial8250_console_write():
>
> void serial8250_console_write(struct uart_8250_port *up, const char *s,
> 			      unsigned int count)
> {
> 	int locked = 1;
>
> 	if (oops_in_progress)
> 		locked = spin_trylock_irqsave(&port->lock, flags);
> 	else
> 		spin_lock_irqsave(&port->lock, flags);
>
>
> 	...
>
>
> 	if (locked)
> 		spin_unlock_irqrestore(&port->lock, flags);
> }
>
> register_console() does not check oops_in_progress at the moment
> and might get blocked on console_sem.
>
> We could add the checks for oops_in_progress into register_console().
> But I am not sure if it is worth it.

It's not worth it just because of parisc.
I haven't seen any such crash in years, so the current implementation
is probably untested and outdated.

> It seems that you used this code for ages. The risk of the deadlock
> is small. It likely works most of the time. The upcoming printk rework
> should allow a cleaner solution.

Yes, it would be great if you can include such a "hard-panic/crash-dump-case"
in the rework.

Helge