Received-SPF: pass (google.com: domain of linux-kernel+bounces-46852-linux.lists.archive=gmail.com@vger.kernel.org designates 147.75.48.161 as permitted sender) client-ip=147.75.48.161;
Sender: Tejun Heo <htejun@gmail.com>
Date: Wed, 31 Jan 2024 07:00:03 -1000
From: Tejun Heo <tj@kernel.org>
To: Waiman Long <longman@redhat.com>
Cc: Lai Jiangshan <jiangshanlai@gmail.com>, linux-kernel@vger.kernel.org,
	Juri Lelli <juri.lelli@redhat.com>,
	Cestmir Kalina <ckalina@redhat.com>,
	Alex Gladkov <agladkov@redhat.com>
Subject: Re: [RFC PATCH 3/3] workqueue: Enable unbound cpumask update on
 ordered workqueues
Message-ID: <Zbp8k1mbKuujC94q@slm.duckdns.org>
References: <20240130183336.511948-1-longman@redhat.com>
 <20240130183336.511948-4-longman@redhat.com>
Precedence: bulk
MIME-Version: 1.0
Content-Type: text/plain; charset=us-ascii
Content-Disposition: inline
In-Reply-To: <20240130183336.511948-4-longman@redhat.com>

Hello,

On Tue, Jan 30, 2024 at 01:33:36PM -0500, Waiman Long wrote:
> +/* requeue the work items stored in wq->o_list */
> +static void requeue_ordered_works(struct workqueue_struct *wq)
> +{
> +	LIST_HEAD(head);
> +	struct work_struct *work, *next;
> +
> +	raw_spin_lock_irq(&wq->o_lock);
> +	if (list_empty(&wq->o_list))
> +		goto unlock_out;	/* No requeuing is needed */
> +
> +	list_splice_init(&wq->o_list, &head);
> +	raw_spin_unlock_irq(&wq->o_lock);
> +
> +	/*
> +	 * Requeue the first batch of work items. Since it may take a while
> +	 * to drain the old pwq and update the workqueue attributes, there
> +	 * may be a rather long list of work items to process. So we allow
> +	 * queue_work() callers to continue putting their work items in o_list.
> +	 */
> +	list_for_each_entry_safe(work, next, &head, entry) {
> +		list_del_init(&work->entry);
> +		local_irq_disable();
> +		__queue_work_rcu_locked(WORK_CPU_UNBOUND, wq, work);
> +		local_irq_enable();
> +	}
> +
> +	/*
> +	 * Now check if there are more work items queued, if so set ORD_WAIT
> +	 * and force incoming queue_work() callers to busy wait until the 2nd
> +	 * batch of work items have been properly requeued. It is assumed
> +	 * that the 2nd batch should be much smaller.
> +	 */
> +	raw_spin_lock_irq(&wq->o_lock);
> +	if (list_empty(&wq->o_list))
> +		goto unlock_out;
> +	WRITE_ONCE(wq->o_state, ORD_WAIT);
> +	list_splice_init(&wq->o_list, &head);
> +	raw_spin_unlock(&wq->o_lock);	/* Leave interrupt disabled */
> +	list_for_each_entry_safe(work, next, &head, entry) {
> +		list_del_init(&work->entry);
> +		__queue_work_rcu_locked(WORK_CPU_UNBOUND, wq, work);
> +	}
> +	WRITE_ONCE(wq->o_state, ORD_NORMAL);
> +	local_irq_enable();
> +	return;
> +
> +unlock_out:
> +	WRITE_ONCE(wq->o_state, ORD_NORMAL);
> +	raw_spin_unlock_irq(&wq->o_lock);
> +}

I'm not a big fan of this approach. It's a rather big departure from how
things are usually done in workqueue. I'd much prefer sth like the
following:

- Add the ability to mark an unbound pwq plugged. If plugged,
   pwq_tryinc_nr_active() always fails.

- When cpumasks need updating, set max_active of all ordered workqueues to
  zero and flush them. Note that if you set all max_actives to zero (note
  that this can be another "plug" flag on the workqueue) first, all the
  ordered workqueues would already be draining, so calling flush_workqueue()
  on them sequentially shouldn't take too long.

- Do the normal pwq allocation and linking but make sure that all new
  ordered pwqs start plugged.

- When update is done, restore the max_actives on all ordered workqueues.

- New work items will now get queued to the newest dfl_pwq which is plugged
  and we know that wq->pwqs list contain pwqs in reverse creation order. So,
  from pwq_release_workfn(), if the pwq being released is for an ordered
  workqueue and not plugged, unplug the pwq right in front.

This hopefully should be less invasive.

Thanks.

-- 
tejun