Re: [RFC][PATCH v4 1/2] printk: Make printk() completely async

From: Sergey Senozhatsky
Date: Wed Mar 16 2016 - 03:56:56 EST


On (03/16/16 16:30), Byungchul Park wrote:
>
> Do you mean the wake_up_process() in console_unlock?

no, I meant wake_up_process(printk_kthread), the newly added one.


-- if we are going to have wake_up_process() in wake_up_klogd_work_func(),
then we need `in_sched' message to potentially trigger a recursion chain

wake_up_klogd_work_func()->wake_up_process()->printk()->wake_up_process()->printk()...

to break this printk()->wake_up_process()->printk(), we need wake_up_process() to
be under the logbuf lock; so vprintk_emit()'s if (logbuf_cpu == this_cpu) will act.


-- if we are going to have wake_up_process() in console_unlock(), then

console_unlock()->{up(), wake_up_process()}->printk()->{console_lock(), console_unlock()}->{up(), wake_up_process()}->printk()...

is undetectable... by the time console_unlock() calls wake_up_process() there
are no printk() locks that this CPU owns.


> I said they should be kept *out of* the critical section. :-)
> Otherwise, it can recurse us forever.

can you explain?

-ss