Re: [PATCH v3 -rcu] workqueue: Convert for_each_wq to use built-in list check

From: Joel Fernandes
Date: Thu Aug 15 2019 - 11:25:02 EST


On Thu, Aug 15, 2019 at 07:57:49AM -0700, Matthew Wilcox wrote:
> On Thu, Aug 15, 2019 at 10:18:42AM -0400, Joel Fernandes (Google) wrote:
> > list_for_each_entry_rcu now has support to check for RCU reader sections
> > as well as lock. Just use the support in it, instead of explicitly
> > checking in the caller.
>
> ...
>
> > #define assert_rcu_or_wq_mutex_or_pool_mutex(wq) \
> > RCU_LOCKDEP_WARN(!rcu_read_lock_held() && \
> > !lockdep_is_held(&wq->mutex) && \
>
> Can't you also get rid of this macro?

Could be. But that should be a different patch. I am only cleaning up the RCU
list lockdep checking in this series since the series introduces that
concept). Please feel free to send a patch for the same.

Arguably, keeping the macro around also can be beneficial in the future.

> It's used in one place:
>
> static struct pool_workqueue *unbound_pwq_by_node(struct workqueue_struct *wq,
> int node)
> {
> assert_rcu_or_wq_mutex_or_pool_mutex(wq);
>
> /*
> * XXX: @node can be NUMA_NO_NODE if CPU goes offline while a
> * delayed item is pending. The plan is to keep CPU -> NODE
> * mapping valid and stable across CPU on/offlines. Once that
> * happens, this workaround can be removed.
> */
> if (unlikely(node == NUMA_NO_NODE))
> return wq->dfl_pwq;
>
> return rcu_dereference_raw(wq->numa_pwq_tbl[node]);
> }
>
> Shouldn't we delete that assert and use
>
> + return rcu_dereference_check(wq->numa_pwq_tbl[node],
> + lockdep_is_held(&wq->mutex) ||
> + lockdep_is_held(&wq_pool_mutex));

Makes sense. This API also does sparse checking. Also hopefully no sparse
issues show up because rcu_dereference_check() but anyone such issues should
be fixed as well.

thanks,

- Joel

>