Re: [PATCH 1/2] printk delay for each line break instead of callback

[Andrew Morton <akpm@linux-foundation.org>]

On Sat, 6 Feb 2010 21:34:25 +0800
Dave Young <hidave.darkstar@gmail.com> wrote:

> printk delay for every callback does not make sense, change to delay every line
> 
> Signed-off-by: Dave Young <hidave.darkstar@gmail.com>
> ---
> 
>  kernel/printk.c |    2 +-
>  1 file changed, 1 insertion(+), 1 deletion(-)
> 
> --- linux-2.6.orig/kernel/printk.c	2010-02-02 13:38:47.646659531 +0800
> +++ linux-2.6/kernel/printk.c	2010-02-02 13:39:19.446657319 +0800
> @@ -678,7 +678,6 @@ asmlinkage int vprintk(const char *fmt, 
>  	char *p;
>  
>  	boot_delay_msec();
> -	printk_delay();
>  
>  	preempt_disable();
>  	/* This stops the holder of console_sem just where we want him */
> @@ -746,6 +745,7 @@ asmlinkage int vprintk(const char *fmt, 
>  	 */
>  	for ( ; *p; p++) {
>  		if (new_text_line) {
> +			printk_delay();
>  			/* Always output the token */
>  			emit_log_char('<');
>  			emit_log_char(current_log_level + '0');

This moves the printk_delay() so that it is now inside
spin_lock_irqsave(logbuf_lock).

This fixes the race I mentioned in the previous email, but it seems a
bad idea.  If the delay is long enough, it could even cause other CPUs
to get hit by the NMI watchdog when trying to acquire logbuf_lock.

A better approach would be to perform the calculation of "how long must
I delay" at this site, but perform the actual delay later, after the
raw_local_irq_restore().  This means that if the printk contains
"a\nb\nc\n" then we won't delay until the final \n, but that seems a
fairly small problem.