printk() vs tty_io

[Peter Zijlstra <peterz@infradead.org>]

Hi tty folks,

I've been poking at reducing the constraints on printk(), like make it
work under rq->lock etc..

Aside from a fwd port of the patch that abuses the console_sem.lock:
https://lkml.org/lkml/2011/6/9/298 and a few other not so very pretty
patches, I ran into the following lockdep splat (using a not so very
pretty lockdep early_printk() patch):

watchdog/0/10 is trying to acquire lock:
 ((console_sem).lock){-.-...}, at: 
but task is already holding lock:
 (&rt_rq->rt_runtime_lock){-.-...}, at: 
which lock already depends on the new lock.

the existing dependency chain (in reverse order) is:

-> #5 (&rt_rq->rt_runtime_lock){-.-...}:
-> #4 (&rq->lock){-.-.-.}:
-> #3 (&p->pi_lock){-.-.-.}:
-> #2 (&tty->write_wait){-.-...}:
-> #1 (&port_lock_key){-.-...}:
-> #0 ((console_sem).lock){-.-...}:

It turns out that writing to a console does wakeups due to tty_io.c.

My question is basically, is there a feasible way around doing these
wakeups from the console::write() path? Everything I thought of was
really quite horrible... and very likely would break stuff since I'm not
that well versed in the whole tty thing.