Re: [RFC][PATCH] printk: Fixup the nmi printk mess

[Peter Zijlstra <peterz@infradead.org>]

On Wed, Jun 10, 2015 at 09:23:04PM +0200, Peter Zijlstra wrote:
> +static int vprintk_nmi(const char *fmt, va_list args)
> +{
> +	struct nmi_seq_buf *s = this_cpu_ptr(&nmi_print_seq);
> +	int add = -1, len;
> +
> +	irq_work_queue(&s->work);
> +
> +	/*
> +	 * head > tail; indicates someone is writing.
> +	 * avoids the buffer being truncated.
> +	 *
> +	 * we rely on the memory barrier to ensure the increment
> +	 * is visible before we start writing to the buffer.
> +	 */
> +	len = atomic_inc_return(&s->head);

That wants to be:

	len = atomic_inc_return(&s->head) - 1;

we need the value before the increment, not after it.

> +
> +	if (len < sizeof(s->buffer)) {
> +		add = vsnprintf(s->buffer + len, sizeof(s->buffer) - len, fmt, args);
> +		/*
> +		 * Fwd head to the right location; NMIs do not nest, therefore
> +		 * we can use regular stores.
> +		 */
> +		atomic_set(&s->head, len + add);
> +		/*
> +		 * Once the data is written and head is correct, update tail to
> +		 * match; indicating we're complete.
> +		 */
> +		smp_wmb();
> +		atomic_set(&s->tail, len + add);
> +	}
> +
> +	return add;
> +}