On Wed 2026-11-11 19:25:17, John Ogness wrote:
> Prepare for per-console atomic flushing policies by introducing
> enum pr_flush_reason to classify various reasons for performing a
> printk backlog flush. The reasons are specified for
> printk_trigger_flush() and console_flush_on_panic().
>
> For atomic printing from printk_trigger_flush(), propagate the
> reason down to __nbcon_atomic_flush_pending_con(), where the actual
> atomic flushing is executed.
>
> For legacy printing, use a new global variable @legacy_flush_reason
> to track the reason in case nbcon consoles are using legacy
> printing.
>
> Update all printk_trigger_flush() and console_flush_on_panic()
> call sites as well as any explicit lower-level atomic flush calls
> to specify the reason.
>
> Note that PR_FLUSH_RUNTIME is used for cases where the reason for
> the flush is not clearly defined.
> --- a/include/linux/printk.h
> +++ b/include/linux/printk.h
> @@ -145,6 +145,24 @@ void early_printk(const char *s, ...) { }
>
> struct dev_printk_info;
>
> +/**
> + * enum pr_flush_reason - Reason for a flush
> + * @PR_FLUSH_RUNTIME: A flush while the kernel is running
> + * @PR_FLUSH_RESTART: The kernel is ending, but another instance
> should
> + * follow (reboot, kexec)
> + * @PR_FLUSH_HANG: The kernel is ending and will enter an infinite loop
> + * @PR_FLUSH_KDB: Printing while KDB is active
> + *
> + * Flush reasons are used to determine if unsafe atomic flushing should be
> + * performed, based on individual console policies.
> + */
> +enum pr_flush_reason {
> + PR_FLUSH_RUNTIME = 0,
> + PR_FLUSH_RESTART,
> + PR_FLUSH_HANG,
> + PR_FLUSH_KDB,
> +};
This entire patchset adds a lot of complexity just to enable all
4 new policy flags for netconsole in the last patch.
There already have few dimensions which affect console flushing:
+ console_loglevel
+ nbcon_prio
This patchset adds 8 new terms in pr_flush_reason and
nbcon_unsafe_flush_flags just to explicitly set all
flush_flags in netconsole.
Let's start with pr_flush_reason. I am not sure if it is worth it:
+ PR_FLUSH_RUNTIME is the default
+ PR_FLUSH_RESTART can be detected by system_state > SYSTEM_RUNNING
+ PR_FLUSH_HANG has only one caller currently distinguished by
the global panic_nbcon_allow_unsafe_takeover
+ PR_FLUSH_KDB can be detected by kdb_trap_printk and
kdb_printf_cpu, see vprintk() implementation.
IMHO, it makes things worse because nbcon_atomic_flush_pending() has
to use PR_FLUSH_RUNTIME in vprintk_emit(). But I think that we
would like to flush it the same way as an explicit flush called
in the given context.
Best Regards,
Petr
