Re: [RFC PATCH 1/5] printk: implement pr_cont_t

[Sergey Senozhatsky]

On (20/08/20 01:32), John Ogness wrote:
> +#define CONT_LINE_MAX LOG_LINE_MAX
> +#define CONT_BUF_COUNT 10
> +static char cont_buf[CONT_BUF_COUNT][CONT_LINE_MAX];
> +static DECLARE_BITMAP(cont_buf_map, CONT_BUF_COUNT);
> +
> +static int get_cont_buf(void)
> +{
> +	int bit;
> +
> +	do {
> +		bit = find_first_zero_bit(cont_buf_map, CONT_BUF_COUNT);
> +		if (bit == CONT_BUF_COUNT)
> +			break;
> +	} while (test_and_set_bit(bit, cont_buf_map));
> +
> +	return bit;
> +}
> +static void put_cont_buf(int index)
> +{
> +	if (WARN_ON(index >= CONT_BUF_COUNT))
> +		return;
> +	if (WARN_ON(!test_bit(index, cont_buf_map)))
> +		return;

Doesn't WARN_ON() do pr_cont() print outs as well? I'm a bit worried by
the path that re-enters pr_cont() from "broken" pr_cont() path.

Just ideas, to keep the discussion alive:

Overall, I wonder if pr_cont buffers should hold more info, e.g. owner
context. If the same context does "normal" printk() while still owning
the unfinished cont buffer then we should flush cont buffer. I may be
in minority here, I don't see a very strong reason to keep the global
order of the messages - e.g. if pid 234 does printk on CPU1 then we
probably don't need to flush pid's 8889 cont buffer (which runs on CPU42,
for instance), but keeping the order of the messages within the context
is somehow much more important. That is, if pid 8889 starts pr_cont buffer
and then triggers a pr_warn() or pr_alert() or any other printk() then I
think we need to flush the cont buffer. Just 5 cents.

	-ss