Re: [RFC] printk/sysrq: Don't play with console_loglevel

[Sergey Senozhatsky]

On (05/28/19 01:24), Dmitry Safonov wrote:
[..]
> While handling sysrq the console_loglevel is bumped to default to print
> sysrq headers. It's done to print sysrq messages with WARNING level for
> consumers of /proc/kmsg, though it sucks by the following reasons:
> - changing console_loglevel may produce tons of messages (especially on
>   bloated with debug/info prints systems)
> - it doesn't guarantee that the message will be printed as printk may
>   deffer the actual console output from buffer (see the comment near
>   printk() in kernel/printk/printk.c)
> 
> Provide KERN_UNSUPPRESSED printk() annotation for such legacy places.
> Make sysrq print the headers unsuppressed instead of changing
> console_loglevel.

I've been thinking about this a while ago... So what I thought back
then was that affected paths are atomic: sysrq, irqs, NMI, etc. Well
at leasted it seemed to be so. Hence we can use per-CPU flag to tell
printk that whatever comes from this-CPU is important and printk should
eventually print it (next time it hits console_unlock()). One candidate
for such per-CPU flag was this_cpu(printk_context). We can steal high
bit (next to NMI printk_safe bit). So the intended use case was something
like this

	sysrq/etc  /* atomic context */
	{
		printk_blah_enter();

		for (...)
			printk();
		...
		dump_bar();

		prinkt_blah_exit();
	}

printk_blah_enter() would set that special - printk_safe_mask_blah - bit,
and prinkt_blah_exit() would clear it. Whenever prinkt->vprintk_store()
would see printk_safe_mask_blah bit set it would mark the log_stored message
as "important, always print!", and console_unlock() would always print those
"important" messages.

	-ss