Re: [PATCH] net/skbuff: silence warnings under memory pressure

On (09/06/19 17:32), Petr Mladek wrote:

quoted

[..]

quoted

I mean, really, do we need to keep calling wake up if it
probably never even executed?

I guess ratelimiting you are talking about ("if it probably never even
executed") would be to check if we have already called wake up on the
log_wait ->head. For that we need to, at least, take log_wait spin_lock
and check that ->head is still in TASK_INTERRUPTIBLE; which is (quite,
but not exactly) close to what wake_up_interruptible() does - it doesn't
wake up the same task twice, it bails out on `p->state & state' check.

I have just realized that only sleeping tasks are in the waitqueue.
It is already handled by waitqueue_active() check.

Yes.

I am afraid that we could not ratelimit the wakeups. The userspace
loggers might then miss the last lines for a long.

That's my concern as well.

We could move wake_up_klogd() back to console_unlock(). But it might
end up with a back-and-forth games according to who is currently
complaining.

We still don't need irq_work, tho.

If we can do
	printk()->console_unlock()->up()->try_to_wake_up()
then we can also do
	printk()           ->             try_to_wake_up()

It's LOGLEVEL_SCHED which tells us if we can try_to_wake_up()
or cannot.

Sigh, I still suggest to ratelimit the warning about failed
allocation.

Hard to imagine how many printk()-s we will have to ratelimit.
To imagine NET maintainers being OK with this is even harder.

	-ss