Re: [PATCH -next RFC v2 8/8] sbitmap: wake up the number of threads based on required tags

[Bart Van Assche <bvanassche@acm.org>]

On 4/8/22 00:39, Yu Kuai wrote:
> Always wake up 'wake_batch' threads will intensify competition and
> split io won't be issued continuously. Now that how many tags is required
> is recorded for huge io, it's safe to wake up baed on required tags.
> 
> Signed-off-by: Yu Kuai <yukuai3@huawei.com>
> ---
>   lib/sbitmap.c | 22 +++++++++++++++++++++-
>   1 file changed, 21 insertions(+), 1 deletion(-)
> 
> diff --git a/lib/sbitmap.c b/lib/sbitmap.c
> index 8d01e02ea4b1..eac9fa5c2b4d 100644
> --- a/lib/sbitmap.c
> +++ b/lib/sbitmap.c
> @@ -614,6 +614,26 @@ static inline void sbq_update_preemption(struct sbitmap_queue *sbq,
>   	WRITE_ONCE(sbq->force_tag_preemption, force);
>   }
>   
> +static unsigned int get_wake_nr(struct sbq_wait_state *ws, unsigned int nr_tags)

Consider renaming "get_wake_nr()" into "nr_to_wake_up()".

> +{
> +	struct sbq_wait *wait;
> +	struct wait_queue_entry *entry;
> +	unsigned int nr = 1;
> +
> +	spin_lock_irq(&ws->wait.lock);
> +	list_for_each_entry(entry, &ws->wait.head, entry) {
> +		wait = container_of(entry, struct sbq_wait, wait);
> +		if (nr_tags <= wait->nr_tags)
> +			break;
> +
> +		nr++;
> +		nr_tags -= wait->nr_tags;
> +	}
> +	spin_unlock_irq(&ws->wait.lock);
> +
> +	return nr;
> +}
> +
>   static bool __sbq_wake_up(struct sbitmap_queue *sbq)
>   {
>   	struct sbq_wait_state *ws;
> @@ -648,7 +668,7 @@ static bool __sbq_wake_up(struct sbitmap_queue *sbq)
>   	smp_mb__before_atomic();
>   	atomic_set(&ws->wait_cnt, wake_batch);
>   	sbq_update_preemption(sbq, wake_batch);
> -	wake_up_nr(&ws->wait, wake_batch);
> +	wake_up_nr(&ws->wait, get_wake_nr(ws, wake_batch));
>   
>   	return true;
>   }

ws->wait.lock is unlocked after the number of threads to wake up has 
been computed and is locked again by wake_up_nr(). The ws->wait.head 
list may be modified after get_wake_nr() returns and before wake_up_nr() 
is called. Isn't that a race condition?

Thanks,

Bart.