Re: [PATCH] io_uring/io-wq: avoid repeated task_work scans during teardown

[Gabriel Krisman Bertazi <krisman@suse.de>]

"Fengnan Chang" <changfengnan@bytedance.com> writes:

> We hit hard-lockup reports from iou-wrk threads stuck in

It seems like a soft-lockup instead no?  From your description,
eventually it solves itself, the task is just uninterruptible while
contending on the spinlock.

> + */
> +struct callback_head *
> +task_work_cancel_match_all(struct task_struct *task,
> +			   bool (*match)(struct callback_head *, void *data),
> +			   void *data)
> +{
> +	struct callback_head **pprev = &task->task_works;
> +	struct callback_head *work, *next;
> +	struct callback_head *head = NULL, **tail = &head;
> +	unsigned long flags;
> +
> +	if (likely(!task_work_pending(task)))
> +		return NULL;
> +
> +	raw_spin_lock_irqsave(&task->pi_lock, flags);
> +	work = READ_ONCE(*pprev);
> +	while (work && work != &work_exited) {
> +		next = READ_ONCE(work->next);
> +		if (!match(work, data)) {
> +			pprev = &work->next;
> +			work = next;
> +			continue;
> +		}
> +
> +		if (!try_cmpxchg(pprev, &work, next))
> +			continue;


IIUC, you could ignore the cmpxchg here because the following loop
iteration on the caller would catch it and retry.  In this case, it no
retry in io_wq_cancel_tw_create, which looks weird.  Did I miss something?

> +
> +		work->next = NULL;
> +		*tail = work;
> +		tail = &work->next;
> +		work = next;
> +	}
> +	raw_spin_unlock_irqrestore(&task->pi_lock, flags);
> +
> +	return head;
> +}
> +
>  static bool task_work_func_match(struct callback_head *cb, void *data)
>  {
>  	return cb->func == data;

-- 
Gabriel Krisman Bertazi