Re: [PATCH] mm/page_alloc: Use write_seqlock_irqsave() instead write_seqlock() + local_irq_save().

[Sebastian Andrzej Siewior <bigeasy@linutronix.de>]

On 2023-06-21 20:33:35 [+0900], Tetsuo Handa wrote:
> On 2023/06/21 19:40, Sebastian Andrzej Siewior wrote:
> > printk_deferred_enter() has to be invoked in non-migrate-able context to
> > ensure that deferred printing is enabled and disabled on the same CPU.
> 
> I can't catch. local_irq_save(flags); makes non-migrate-able context
> because sleeping is not allowed while IRQ is disabled, doesn't it?

That is correct. The problem is the local_irq_save() which remains on
PREEMPT_RT and this is problematic during write_seqlock() which acquires
spinlock_t which becomes a sleeping lock. write_seqlock_irqsave()
substitutes everything properly so on RT there is no local_irq_save().

> > This is the case after zonelist_update_seq has been acquired.
> > 
> > Acquire zonelist_update_seq with write_seqlock_irqsave() and then defer
> > printk output.
> 
> What guarantees that write_seqlock_irqsave() never calls printk()
> (e.g. lockdep warning) before printk_deferred_enter() takes effect?

Actually nothing but this could fixed by making the lockdep annotation
first followed by the actual increment in
do_write_seqcount_begin_nested(). I don't know this is the other way
around since we do annotation followed by locking in order to see the
splat before it happens.

Let me do a patch for that since it looks like the right thing to do.

If this "rare" issue is a blocker I don't see a sane way around it. Be
aware that tty_insert_flip_string_and_push_buffer() has atomic/ KASAN
annotation which _could_ detect something which then leads to pr_err()
invocations while tty_port::lock is acquired. Looking over
tty_port_close_start() there are pr_warn() invocation with the lock held
so again, possibilities.

Petr, is printing via print threads and direct printing via explicit
driver callback (which would avoid this problem) on the agenda?

Sebastian