Re: [PATCH 6/6] io_uring: batch completion in prior_task_list

[Pavel Begunkov <asml.silence@gmail.com>]

On 10/29/21 13:22, Hao Xu wrote:
> In previous patches, we have already gathered some tw with
> io_req_task_complete() as callback in prior_task_list, let's complete
> them in batch regardless uring lock. For instance, we are doing simple
> direct read, most task work will be io_req_task_complete(), with this
> patch we don't need to hold uring lock there for long time.
> 
> Signed-off-by: Hao Xu <haoxu@linux.alibaba.com>
> ---
>   fs/io_uring.c | 52 ++++++++++++++++++++++++++++++++++++++++++---------
>   1 file changed, 43 insertions(+), 9 deletions(-)
> 
> diff --git a/fs/io_uring.c b/fs/io_uring.c
> index 694195c086f3..565cd0b34f18 100644
> --- a/fs/io_uring.c
> +++ b/fs/io_uring.c
> @@ -2166,6 +2166,37 @@ static inline unsigned int io_put_rw_kbuf(struct io_kiocb *req)
>   	return io_put_kbuf(req, req->kbuf);
>   }
>   
> +static void handle_prior_tw_list(struct io_wq_work_node *node)
> +{
> +	struct io_ring_ctx *ctx = NULL;
> +
> +	do {
> +		struct io_wq_work_node *next = node->next;
> +		struct io_kiocb *req = container_of(node, struct io_kiocb,
> +						    io_task_work.node);
> +		if (req->ctx != ctx) {
> +			if (ctx) {
> +				io_commit_cqring(ctx);
> +				spin_unlock(&ctx->completion_lock);
> +				io_cqring_ev_posted(ctx);
> +				percpu_ref_put(&ctx->refs);
> +			}
> +			ctx = req->ctx;
> +			percpu_ref_get(&ctx->refs);
> +			spin_lock(&ctx->completion_lock);
> +		}
> +		__io_req_complete_post(req, req->result, io_put_rw_kbuf(req));
> +		node = next;
> +	} while (node);
> +
> +	if (ctx) {
> +		io_commit_cqring(ctx);
> +		spin_unlock(&ctx->completion_lock);
> +		io_cqring_ev_posted(ctx);
> +		percpu_ref_put(&ctx->refs);
> +	}
> +}
> +
>   static void handle_tw_list(struct io_wq_work_node *node, struct io_ring_ctx **ctx, bool *locked)
>   {
>   	do {
> @@ -2193,25 +2224,28 @@ static void tctx_task_work(struct callback_head *cb)
>   						  task_work);
>   
>   	while (1) {
> -		struct io_wq_work_node *node;
> -		struct io_wq_work_list *merged_list;
> +		struct io_wq_work_node *node1, *node2;
>   
> -		if (!tctx->prior_task_list.first &&
> -		    !tctx->task_list.first && locked)
> +		if (!tctx->task_list.first &&
> +		    !tctx->prior_task_list.first && locked)
>   			io_submit_flush_completions(ctx);
>   
>   		spin_lock_irq(&tctx->task_lock);
> -		merged_list = wq_list_merge(&tctx->prior_task_list, &tctx->task_list);
> -		node = merged_list->first;
> +		node1 = tctx->prior_task_list.first;
> +		node2 = tctx->task_list.first;
>   		INIT_WQ_LIST(&tctx->task_list);
>   		INIT_WQ_LIST(&tctx->prior_task_list);
> -		if (!node)
> +		if (!node2 && !node1)
>   			tctx->task_running = false;
>   		spin_unlock_irq(&tctx->task_lock);
> -		if (!node)
> +		if (!node2 && !node1)
>   			break;
>   
> -		handle_tw_list(node, &ctx, &locked);
> +		if (node1)
> +			handle_prior_tw_list(node1);

IIUC, it moves all IRQ rw completions to this new path even when we already
have the lock. One concern is that io_submit_flush_completions() is better
optimised. Should probably be visible for one threaded apps and a bunch of
other cases.

How about a combined scheme? if we can grab the lock, go through the old
path, otherwise handle_prior_tw_list(). The rest looks good, will formally
review once we deal with this one.

> +
> +		if (node2)
> +			handle_tw_list(node2, &ctx, &locked);
>   		cond_resched();
>   	}
>   
> 

-- 
Pavel Begunkov