Re: [RFC PATCH] blk-cgroup: prevent rcu_sched detected stalls warnings in blkg_destroy_all()

[yukuai (C)]

On 2020/11/25 20:32, Tejun Heo wrote:
> Hello,
>
> Thanks for the fix. A couple comments below.
>
> On Sat, Nov 21, 2020 at 04:34:20PM +0800, Yu Kuai wrote:
> > +#define BLKG_DESTROY_BATH 4096
>
> I think you meant BLKG_DESTROY_BATCH.
>
> >   static void blkg_destroy_all(struct request_queue *q)
> >   {
> >   	struct blkcg_gq *blkg, *n;
> > +	int count = BLKG_DESTROY_BATH;
>
> But might as well just write 4096 here.
>
> >   	spin_lock_irq(&q->queue_lock);
> >   	list_for_each_entry_safe(blkg, n, &q->blkg_list, q_node) {
> >   		struct blkcg *blkcg = blkg->blkcg;
> >   
> > +		/*
> > +		 * If the list is too long, the loop can took a long time,
> > +		 * thus relese the lock for a while when a batch of blkcg
> > +		 * were destroyed.
> > +		 */
> > +		if (!(--count)) {
> > +			count = BLKG_DESTROY_BATH;
> > +			spin_unlock_irq(&q->queue_lock);
> > +			cond_resched();
> > +			spin_lock_irq(&q->queue_lock);
>
> You can't continue iteration after dropping both locks. You'd have to jump
> out of loop and start list_for_each_entry_safe() again.

Thanks for your review, it's right. On the other hand
blkcg_activate_policy() and blkcg_deactivate_policy() might have the
same issue. My idea is that inserting a bookmark to the list, and
restard from here.

By the way, I found that blk_throtl_update_limit_valid() is called from
throtl_pd_offline(). If CONFIG_BLK_DEV_THROTTLING_LOW is off, lower
limit will always be zero, therefor a lot of time will be wasted to
iterate descendants to find a nonzero lower limit.

Do you think it's ok to do such modification:

diff --git a/block/blk-throttle.c b/block/blk-throttle.c
index b771c4299982..d52cac9f3a7c 100644
--- a/block/blk-throttle.c
+++ b/block/blk-throttle.c
@@ -587,6 +587,7 @@ static void throtl_pd_online(struct blkg_policy_data 
*pd)
        tg_update_has_rules(tg);
 }

+#ifdef CONFIG_BLK_DEV_THROTTLING_LOW
 static void blk_throtl_update_limit_valid(struct throtl_data *td)
 {
        struct cgroup_subsys_state *pos_css;
@@ -607,6 +608,11 @@ static void blk_throtl_update_limit_valid(struct 
throtl_data *td)

        td->limit_valid[LIMIT_LOW] = low_valid;
 }
+#else
+static inline void blk_throtl_update_limit_valid(struct throtl_data *td)
+{
+}
+#endif

 static void throtl_upgrade_state(struct throtl_data *td);
 static void throtl_pd_offline(struct blkg_policy_data *pd)

Thanks!
Yu Kuai