Re: [RFC PATCH v1 19/25] printk: introduce emergency messages

[John Ogness]

On 2019-03-08, Petr Mladek <pmladek@xxxxxxxx> wrote:
>>> +static bool console_can_emergency(int level)
>>> +{
>>> +	struct console *con;
>>> +
>>> +	for_each_console(con) {
>>> +		if (!(con->flags & CON_ENABLED))
>>> +			continue;
>>> +		if (con->write_atomic && level < emergency_console_loglevel)
>>> +			return true;
>>> +		if (con->write && (con->flags & CON_BOOT))
>>> +			return true;
>>> +	}
>>> +	return false;
>>> +}
>>> +
>>> +static void call_emergency_console_drivers(int level, const char *text,
>>> +					   size_t text_len)
>>> +{
>>> +	struct console *con;
>>> +
>>> +	for_each_console(con) {
>>> +		if (!(con->flags & CON_ENABLED))
>>> +			continue;
>>> +		if (con->write_atomic && level < emergency_console_loglevel) {
>>> +			con->write_atomic(con, text, text_len);
>>> +			continue;
>>> +		}
>>> +		if (con->write && (con->flags & CON_BOOT)) {
>>> +			con->write(con, text, text_len);
>>> +			continue;
>>> +		}
>>> +	}
>>> +}
>>> +
>>> +static void printk_emergency(char *buffer, int level, u64 ts_nsec, u16 cpu,
>>> +			     char *text, u16 text_len)
>>> +{
>>> +	struct printk_log msg;
>>> +	size_t prefix_len;
>>> +
>>> +	if (!console_can_emergency(level))
>>> +		return;
>>> +
>>> +	msg.level = level;
>>> +	msg.ts_nsec = ts_nsec;
>>> +	msg.cpu = cpu;
>>> +	msg.facility = 0;
>>> +
>>> +	/* "text" must have PREFIX_MAX preceding bytes available */
>>> +
>>> +	prefix_len = print_prefix(&msg,
>>> +				  console_msg_format & MSG_FORMAT_SYSLOG,
>>> +				  printk_time, buffer);
>>> +	/* move the prefix forward to the beginning of the message text */
>>> +	text -= prefix_len;
>>> +	memmove(text, buffer, prefix_len);
>>> +	text_len += prefix_len;
>>> +
>>> +	text[text_len++] = '\n';
>>> +
>>> +	call_emergency_console_drivers(level, text, text_len);
>> 
>> So this iterates the console list and calls consoles' callbacks, but
>> what prevents console driver to be rmmod-ed under us?
>> 
>> CPU0                                  CPU1
>> 
>> printk_emergency()                    rmmod netcon
>>  call_emergency_console_drivers()
>>   con_foo->flags & CON_ENABLED == 1
>>                                       unregister_console(con_foo)
>>                                       con_foo->flags &= ~CON_ENABLED
>>                                       __exit // con_foo gone ?
>>   con_foo->write()
>> 
>> We use console_lock()/console_trylock() in order to protect the list
>> and console drivers; but this brings scheduler to the picture, with
>> all its locks.
>
> Great catch!

Yes, thanks!

> I think that it is doable to guard the list using RCU.

I think it would be enough to take the prb_cpulock when modifying the
console linked list. That will keep printk_emergency() out until the
list has been updated. (registering/unregistering consoles is not
something that happens often.)

John Ogness