Re: [RFC PATCH v1 00/25] printk: new implementation

[Sergey Senozhatsky]

On (03/13/19 09:19), John Ogness wrote:
> >> Yes. I will post a series that only implements the ringbuffer using
> >> your simplified API. That will be enough to remove printk_safe and
> >> actually does most of the work of updating devkmsg, kmsg_dump, and
> >> syslog.
> >
> > This may _not_ be enough to remove printk_safe. One of the reasons
> > printk_safe "condom" came into existence was console_sem (which
> > is a bit too important to ignore it):
> >
> > 	printk()
> > 	 console_trylock()
> > 	  console_unlock()
> > 	   up()
> > 	    raw_spin_lock_irqsave(&sem->lock, flags)
> > 	     __up()
> > 	      wake_up_process()
> > 	       WARN/etc
> > 	        printk()
> > 		 console_trylock()
> > 		  down_trylock()
> > 		   raw_spin_lock_irqsave(&sem->lock, flags)  << deadlock
> >
> > Back then we were looking at
> >
> > 	printk->console_sem->lock->printk->console_sem->lock
> >
> > deadlock report from LG, if I'm not mistaken.
> 
> The main drawback of printk_safe is the safe buffers, which, aside from
> bogus timestamping, may never make it back to the printk log buffer.
>
> With the new ring buffer the safe buffers are not needed, even in the
> recursive situation. As you are pointing out, the notification/wake
> component of printk_safe will still be needed. I will leave that (small)
> part in printk_safe.c.

Yeah, all I'm saying is that as it stands new printk() is missing a huge
and necessary part - console semaphore. And things can get very different
once you add that missing part. It brings a lot of stuff back to printk.

logbuf and logbuf_lock are not really huge printk problems. scheduler,
timekeeping locks, etc. are much bigger ones. Those dependencies don't
come from logbuf/logbuf_lock.

	-ss