Re: [RFC][PATCHv2 4/8] pm: switch to printk.emergency mode in unsafe places

[Petr Mladek]

On Sun 2017-04-09 12:59:18, Andreas Mohr wrote:
> On Thu, Apr 06, 2017 at 07:20:52PM +0200, Pavel Machek wrote:
> > On Wed 2017-03-29 18:25:07, Sergey Senozhatsky wrote:
> > > It's not always possible/safe to wake_up() printk kernel
> > > thread. For example, late suspend/early resume may printk()
> > > while timekeeping is not initialized yet, so calling into the
> > > scheduler may result in recursive warnings.
> > > 
> > > Another thing to notice is the fact PM at some point
> > > freezes user space and kernel threads: freeze_processes()
> > > and freeze_kernel_threads(), correspondingly. Thus we need
> > > printk() to operate in old mode there and attempt to
> > > immediately flush pending kernel message to the console.
> > > 
> Sergey has mentioned it already:
> "at some point freezes user space and kernel threads".
> Well, this is the action which is *itself* causing thoroughly disrupting consequences,
> which I'd think thus ought to be responsible to
> ensure *itself* that all resulting consequences actually can be dealt with properly,
> rather than having
> weird *completely-unrelated-dependency* crap
> ("there happens to be some functionality called printk, and we need to bend it,
> since we need to bend it, since otherwise it would not be bent" - ahem...)
> leak into ("layer violation" keyword)
> pm handling implementation specifics.
> IOW, I would think that for any relevant kthread use in API user code,
> such code ought to be able to
> register kthread-API-provided callbacks (observer pattern, or whatever)
> where the (back to current case:) printk kthread would then be able to
> *implicitly*/*invisibly* switch the entire printk operation interface
> (e.g. via a global interface struct) to
> the "dumb"/"safe" fallback variant.
> Potential interface: kthread_notify(callback_func, kthread_notification_type);

Interesting idea. The power management area probably can be solved
by the existing notifiers framework, see register_pm_notifier().

I haven't checked it but if the notifiers are called on right
locations, it would be cleaner than adding the calls into
the pm code.


> That way it could (hopefully) be ensured that
> people could use a consistent "printk" *interface* universally regardless of which
> "special" conditions happen to be in place at the moment.
> (IOW, keep interface behaviour which is required/expected at user code
> definitely isolated from
> awkward "implementation aspects" necessity which is currently poisoning user code implementation).

> Put differently,
> handling preferrably ought to get consistently adapted (i.e., switched) *centrally*,
> rather than
> requiring weird helpers (printk_emergency_X()) at all user code sites.

Note that there already all many printk/console related "hacks"
in sensitive code paths. For example, see the use of
pm_prepare_console(), suspend_console(), console_level.

Best Regards,
Petr