Re: [PATCH 1/2] printk: Add an option to flush all messages on panic

[Feng Tang]

Hi Sergey,

Thanks for the review.

On Tue, Apr 02, 2019 at 11:14:19AM +0900, Sergey Senozhatsky wrote:
> On (04/01/19 18:48), Feng Tang wrote:
> > diff --git a/arch/powerpc/kernel/traps.c b/arch/powerpc/kernel/traps.c
> > index 1fd45a8..58d9580 100644
> > --- a/arch/powerpc/kernel/traps.c
> > +++ b/arch/powerpc/kernel/traps.c
> > @@ -179,7 +179,7 @@ extern void panic_flush_kmsg_end(void)
> >  	kmsg_dump(KMSG_DUMP_PANIC);
> >  	bust_spinlocks(0);
> >  	debug_locks_off();
> > -	console_flush_on_panic();
> > +	console_flush_on_panic(false);
> >  }
> [..]
> > @@ -277,7 +277,7 @@ void panic(const char *fmt, ...)
> >  	 * panic() is not being callled from OOPS.
> >  	 */
> >  	debug_locks_off();
> > -	console_flush_on_panic();
> > +	console_flush_on_panic(false);
> [..]
> > -void console_flush_on_panic(void)
> > +void console_flush_on_panic(bool flush_all)
> >  {
> >  	/*
> >  	 * If someone else is holding the console lock, trylock will fail
> > @@ -2539,6 +2540,12 @@ void console_flush_on_panic(void)
> >  	 */
> >  	console_trylock();
> >  	console_may_schedule = 0;
> > +
> > +	/* User may want to see all the printk messages on panic */
> > +	if (flush_all) {
> > +		console_seq = log_first_seq;
> > +		console_idx = log_first_idx;
> > +	}
> >  	console_unlock();
> 
> So my first thought was - let's not add a `bool flag', but instead add
> an `enum' with clear flag names, e.g. DUMP_ALL/DUMP_PENDING, etc. Something
> similar to what ftrace_dump(DUMP_ALL) does. And we already have panic_print
> bit-mask and panic_print_sys_info(), which controls what we print when in
> panic. So we can move console_flush_on_panic() to panic_print_sys_info()
> and handle different types of console_flush_on_panic() there:
> 
> ---
> 
> diff --git a/kernel/panic.c b/kernel/panic.c
> index 0ae0d7332f12..e142faaebbcd 100644
> --- a/kernel/panic.c
> +++ b/kernel/panic.c
> @@ -134,6 +134,11 @@ EXPORT_SYMBOL(nmi_panic);
>  
>  static void panic_print_sys_info(void)
>  {
> +	if (panic_print & PANIC_PRINT_REPLAY_LOGBUF)
> +		console_flush_on_panic(DUMP_ALL);
> +	else
> +		console_flush_on_panic(DUMP_PENDING);
> +
>  	if (panic_print & PANIC_PRINT_TASK_INFO)
>  		show_state();
>  
> @@ -277,7 +282,6 @@ void panic(const char *fmt, ...)
>  	 * panic() is not being callled from OOPS.
>  	 */
>  	debug_locks_off();
> -	console_flush_on_panic();
>  
>  	panic_print_sys_info();
>  

I guess this is what my 2/2 patch exactly does :)


> ---
> 
> After looking at this more - DUMP_ALL/DUMP_PENDING clearly depend
> on panic_print. So... May be we can move panic_print_sys_info() to
> printk.c, and for PANIC_PRINT_REPLAY_LOGBUF case just reset console
> seq/idx (console seq/idx are internal to printk) before we
> flush_on_panic. This way we would not need to modify
> console_flush_on_panic() prototype at all.

My thought is panic_print_sys_info() may be more related to panic, but
I'm fine with either way.

Thanks,
Feng

> 
> Let's hear from people.
> 
> 	-ss