Re: [PATCH] md/raid5: serialize plug list add with device_lock

[yu kuai]

Hi,

在 2026/9/2 17:53, Li Youhong 写道:
> From: Li Youhong <liyouhong@xxxxxxxxxx>
>
> raid5_unplug() can spin forever under conf->device_lock when
> raid5_plug_cb.list still points at a stripe whose sh->lru has
> already been reinitialized (self-looped). That disables IRQs on
> the holder CPU and causes multi-CPU hard lockups on waiters of
> the same lock.
>
> This happens because release_stripe_plug() sets
> STRIPE_ON_UNPLUG_LIST and list_add_tail(sh->lru) without
> device_lock, while do_release_stripe() may concurrently move the
> same lru onto handle/inactive when the last reference drops.
>
> Note: a 2020 proposal tried extra refs / checking
> STRIPE_ON_UNPLUG_LIST in do_release_stripe() without serializing
> the plug enqueue:
> https://lore.kernel.org/linux-raid/20200108163023.9301-1-guoqing.jiang@xxxxxxxxxxxxxxx/
> That still leaves a TOCTOU window where the bit is clear during
> the check and set afterwards, allowing two list_add on the same
> lru. It was not merged.
>
> Serialize the bit update and list_add with device_lock. If
> do_release_stripe() still sees STRIPE_ON_UNPLUG_LIST, restore the
> reference and let raid5_unplug() own the final release.
>
> Observed on production 9-disk NVMe RAID5 under MySQL AIO
> (io_submit -> blk_finish_plug -> raid5_unplug).
>
> Fixes: 8811b5968f62 ("raid5: make_request use batch stripe release")
> Cc: stable@xxxxxxxxxxxxxxx
> Signed-off-by: Li Youhong <liyouhong@xxxxxxxxxx>
> ---
>   drivers/md/raid5.c | 33 ++++++++++++++++++++++++++++++---
>   1 file changed, 30 insertions(+), 3 deletions(-)
>
> diff --git a/drivers/md/raid5.c b/drivers/md/raid5.c
> index b91545ce090d..c9197b63b1c4 100644
> --- a/drivers/md/raid5.c
> +++ b/drivers/md/raid5.c
> @@ -229,6 +229,17 @@ static void do_release_stripe(struct r5conf *conf, struct stripe_head *sh,
>   	int i;
>   	int injournal = 0;	/* number of date pages with R5_InJournal */
>   
> +	/*
> +	 * Stripe is owned by release_stripe_plug()'s cb->list. A concurrent
> +	 * last-ref release can reach here after the stripe was queued for
> +	 * unplug (lru may already be non-empty). Do not re-add lru elsewhere;
> +	 * restore the reference and let raid5_unplug() finish the release.
> +	 */
> +	if (test_bit(STRIPE_ON_UNPLUG_LIST, &sh->state)) {
> +		atomic_inc(&sh->count);
> +		return;
> +	}
> +
>   	BUG_ON(!list_empty(&sh->lru));
>   	BUG_ON(atomic_read(&conf->active_stripes)==0);
>   
> @@ -5760,6 +5771,9 @@ static void release_stripe_plug(struct mddev *mddev,
>   		raid5_unplug, mddev,
>   		sizeof(struct raid5_plug_cb));
>   	struct raid5_plug_cb *cb;
> +	struct r5conf *conf = mddev->private;
> +	unsigned long flags;
> +	bool queued = false;
>   
>   	if (!blk_cb) {
>   		raid5_release_stripe(sh);
> @@ -5775,9 +5789,22 @@ static void release_stripe_plug(struct mddev *mddev,
>   			INIT_LIST_HEAD(cb->temp_inactive_list + i);
>   	}
>   
> -	if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state))
> -		list_add_tail(&sh->lru, &cb->list);
> -	else
> +	/*
> +	 * Serialize with do_release_stripe() on device_lock so sh->lru cannot
> +	 * be added to handle/inactive and cb->list at the same time.
> +	 */
> +	spin_lock_irqsave(&conf->device_lock, flags);
> +	if (!test_and_set_bit(STRIPE_ON_UNPLUG_LIST, &sh->state)) {
> +		if (unlikely(!list_empty(&sh->lru))) {
> +			clear_bit(STRIPE_ON_UNPLUG_LIST, &sh->state);
> +		} else {
> +			list_add_tail(&sh->lru, &cb->list);
> +			queued = true;
> +		}
> +	}
> +	spin_unlock_irqrestore(&conf->device_lock, flags);

Do you run some performance test on this patch? I believe this is from IO hot path,
a spinlock is not acceptable. I know there is already some spinlock for raid5, but
I'd like not to introduce new lock contention.

Can this problem be fixed by a new llist in stripe_head?

> +
> +	if (!queued)
>   		raid5_release_stripe(sh);
>   }
>   

-- 
Thanks,
Kuai