Re: [PATCH] percpu_ida: Use _irqsave() instead of local_irq_save() + spin_lock

From: Matthew Wilcox <willy@infradead.org>
To: Jens Axboe <axboe@kernel.dk>
Cc: Andrew Morton <akpm@linux-foundation.org>,
	Sebastian Andrzej Siewior <bigeasy@linutronix.de>,
	linux-kernel@vger.kernel.org, tglx@linutronix.de,
	Nicholas Bellinger <nab@linux-iscsi.org>,
	Shaohua Li <shli@fb.com>,
	Kent Overstreet <kent.overstreet@gmail.com>
Subject: Re: [PATCH] percpu_ida: Use _irqsave() instead of local_irq_save() + spin_lock
Date: Mon, 7 May 2018 06:47:31 -0700	[thread overview]
Message-ID: <20180507134731.GA28974@bombadil.infradead.org> (raw)
In-Reply-To: <20180505155202.GA29992@bombadil.infradead.org>

On Sat, May 05, 2018 at 08:52:02AM -0700, Matthew Wilcox wrote:
> init and destroy seem to map to sbitmap_queue_init_node and
> sbitmap_queue_free.  percpu_ida_free maps to sbitmap_queue_clear.

Hmm.

void sbitmap_queue_clear(struct sbitmap_queue *sbq, unsigned int nr,
                         unsigned int cpu)
{
        sbitmap_clear_bit_unlock(&sbq->sb, nr);
        sbq_wake_up(sbq);
        if (likely(!sbq->round_robin && nr < sbq->sb.depth))
                *per_cpu_ptr(sbq->alloc_hint, cpu) = nr;
}
EXPORT_SYMBOL_GPL(sbitmap_queue_clear);

If we free a tag on a CPU other than the one it's allocated on, that seems
like it's going to guarantee a cacheline pingpong.  Is the alloc_hint
really that valuable?  I'd be tempted to maintain the alloc_hint (if it's
at all valuable) as being just a hint for which word to look at first,
and only update it on allocation, rather than updating it on free.
Then we can drop the 'cpu' argument to sbitmap_queue_clear(), which
would help this conversion because the percpu_ida users don't know what
CPU their tag was allocated on.