Re: [PATCH printk v3 14/15] printk: extend console_lock for proper kthread support

[Petr Mladek <pmladek@suse.com>]

On Fri 2022-04-22 23:31:11, John Ogness wrote:
> On 2022-04-22, Petr Mladek <pmladek@suse.com> wrote:
> > Another problem is that the ordering is not stable. The console
> > might come and go.
> 
> The console list is protected by @console_sem, so it wouldn't be an
> actual problem. The real issue is that lockdep would not like it. A new
> lockdep class would need to be setup for each register_console().

Yeah. I did not mention it explicitely but I meant it as a problem
with lockdep.

> >> Anyway, I will first look into the nested locking solution. That
> >> seems more promising to me and it would go a long way to simplify the
> >> locking hierarchy.
> >
> > Please, do not spend too much time on this. The solution must be
> > simple in principle. If it gets complicated than it will likely
> > be worse than the current code.
> 
> Sure. The goal is to simplify. The only complexity will be doing in a
> way that allow lockdep to understand it.

I am not sure how to distinguish intentional and non-intentional
ordering change.


> > Alternative solution would be to reduce the number of variables
> > affected by the race. I mean:
> >
> >    + replace CON_THB_BLOCKED flag with con->blocked to avoid
> >      the needed of READ_ONCE()/WRITE_ONCE().
> >
> >    + check con->blocked right after taking con->lock in
> >      printk_kthread_func() so that all the other accesses are
> >      safe.
> 
> Honestly, I would prefer this to what v4 is doing. The only reason
> CON_THD_BLOCKED is a flag is to save space. But we are only talking
> about a few bytes being saved. There aren't that many consoles.
> 
> It would be a very simple change. Literally just replacing the 3 lines
> that set/clear CON_THD_BLOCKED and replacing/reordering the 2 lines that
> check the flag. Then all the READ_ONCE/WRITE_ONCE to @flags could be
> removed.

I agree that it sounds like the easiest solution for now. If you
prepare v5 with this change then I push it into linux-next instead
of v4.

Well, I think that we need to make con->lock safe to use in the long
term. The above workaround in printk_kthread_func() is good enough
for now because this is the only location where con->lock is taken without
console_sem. But I am sure that we/people will want to do more
console-specific operations without console_sem in the future.

IMHO, the only sane approach is to follow the proposed rules:

    + console_lock() will synchronize both global and per-console
      stuff.

    + con->lock will synchronize per-console stuff.

    + con->lock could not be taken alone when the big console_lock()
      is taken.


I currently know only about two solutions:

    1. The nested locking. console_lock() will take console_sem
       and all con->lock's and will keep them locked.

       It is rather trivial in principle. The problem is lockdep
       and possible ABBA deadlocks caused by unstable ordering.


    2. Create the wrappers around con->lock that will check
       whether console_sem is taken (con->locked flag).

       It will require additional per-console waitqueue. But all
       the magic will be hidden in the wrappers.


I personally prefer 2nd approach for the long term solution. It might
look more complicated but it will not break lockdep.

Best Regards,
Petr