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

[Tetsuo Handa <penguin-kernel@I-love.SAKURA.ne.jp>]

On 2023/06/23 19:11, Sebastian Andrzej Siewior wrote:
> | unsigned __seqprop_spinlock_sequence(const seqcount_spinlock_t *s)
> | { 
> | 	unsigned seq = READ_ONCE(s->seqcount.sequence);
> | 
> | 	if (unlikely(seq & 1)) {
> | 		spin_lock(s->lock); 
> | 		spin_unlock(s->lock);
> | 		seq = READ_ONCE(s->seqcount.sequence);
> | 	}
> | 	return seq;
> | }

OK. I understood that read_seqbegin() implies spin_lock()/spin_unlock() if RT.
What a deep macro. Thank you for explanation.



Well,

/*
 * Zonelists may change due to hotplug during allocation. Detect when zonelists
 * have been rebuilt so allocation retries. Reader side does not lock and
 * retries the allocation if zonelist changes. Writer side is protected by the
 * embedded spin_lock.
 */

is not accurate. Something like below?

  If !RT, reader side does not lock and retries the allocation if zonelist changes.
  If RT, reader side grabs and releases the embedded spin_lock in order to wait
  for zonelist change operations to complete.



Hmm, I feel worried that kmalloc(GFP_ATOMIC) from hard IRQ context might sleep if RT...