Re: [PATCH 5/5] io_uring: implement multishot mode for accept

[Hao Xu <haoxu.linux@gmail.com>]

在 2022/5/6 下午10:42, Jens Axboe 写道:
> On 5/6/22 1:01 AM, Hao Xu wrote:
>> diff --git a/fs/io_uring.c b/fs/io_uring.c
>> index 0a83ecc457d1..9febe7774dc3 100644
>> --- a/fs/io_uring.c
>> +++ b/fs/io_uring.c
>> @@ -1254,6 +1254,7 @@ static int io_close_fixed(struct io_kiocb *req, unsigned int issue_flags);
>>   static enum hrtimer_restart io_link_timeout_fn(struct hrtimer *timer);
>>   static void io_eventfd_signal(struct io_ring_ctx *ctx);
>>   static void io_req_tw_post_queue(struct io_kiocb *req, s32 res, u32 cflags);
>> +static void io_poll_remove_entries(struct io_kiocb *req);
>>   
>>   static struct kmem_cache *req_cachep;
>>   
>> @@ -5690,24 +5691,29 @@ static int io_recv(struct io_kiocb *req, unsigned int issue_flags)
>>   static int io_accept_prep(struct io_kiocb *req, const struct io_uring_sqe *sqe)
>>   {
>>   	struct io_accept *accept = &req->accept;
>> +	bool multishot;
>>   
>>   	if (unlikely(req->ctx->flags & IORING_SETUP_IOPOLL))
>>   		return -EINVAL;
>> -	if (sqe->ioprio || sqe->len || sqe->buf_index)
>> +	if (sqe->len || sqe->buf_index)
>>   		return -EINVAL;
>>   
>>   	accept->addr = u64_to_user_ptr(READ_ONCE(sqe->addr));
>>   	accept->addr_len = u64_to_user_ptr(READ_ONCE(sqe->addr2));
>>   	accept->flags = READ_ONCE(sqe->accept_flags);
>>   	accept->nofile = rlimit(RLIMIT_NOFILE);
>> +	multishot = !!(READ_ONCE(sqe->ioprio) & IORING_ACCEPT_MULTISHOT);
> 
> I tend to like:
> 
> 	multishot = READ_ONCE(sqe->ioprio) & IORING_ACCEPT_MULTISHOT) != 0;
> 
> as I think it's more readable. But I think we really want it ala:
> 
> 	u16 poll_flags;
> 
> 	poll_flags = READ_ONCE(sqe->ioprio);
> 	if (poll_flags & ~IORING_ACCEPT_MULTISHOT)
> 		return -EINVAL;
> 
> 	...
> 
> to ensure that we can add more flags later, hence only accepting this
> single flag right now.
> 
> Do we need REQ_F_APOLL_MULTI_POLLED, or can we just store whether this
> is a multishot request in struct io_accept?
I think we can do it in this way, but it may be a bit inconvenient if we
add other multishot OPCODE. With REQ_F_APOLL_MULTI_POLLED we can just
check req->flags in the poll arming path, which keeps it op unrelated.
> 
>> @@ -5760,7 +5774,35 @@ static int io_accept(struct io_kiocb *req, unsigned int issue_flags)
>>   		ret = io_install_fixed_file(req, file, issue_flags,
>>   					    accept->file_slot - 1);
>>   	}
>> -	__io_req_complete(req, issue_flags, ret, 0);
>> +
>> +	if (req->flags & REQ_F_APOLL_MULTISHOT) {
>> +		if (ret >= 0) {
>> +			bool filled;
>> +
>> +			spin_lock(&ctx->completion_lock);
>> +			filled = io_fill_cqe_aux(ctx, req->cqe.user_data, ret,
>> +						 IORING_CQE_F_MORE);
>> +			io_commit_cqring(ctx);
>> +			spin_unlock(&ctx->completion_lock);
>> +			if (unlikely(!filled)) {
>> +				io_poll_clean(req);
>> +				return -ECANCELED;
>> +			}
>> +			io_cqring_ev_posted(ctx);
>> +			goto retry;
>> +		} else {
>> +			/*
>> +			 * the apoll multishot req should handle poll
>> +			 * cancellation by itself since the upper layer
>> +			 * who called io_queue_sqe() cannot get errors
>> +			 * happened here.
>> +			 */
>> +			io_poll_clean(req);
>> +			return ret;
>> +		}
>> +	} else {
>> +		__io_req_complete(req, issue_flags, ret, 0);
>> +	}
>>   	return 0;
>>   }
> 
> I'd probably just make that:
> 
> 	if (!(req->flags & REQ_F_APOLL_MULTISHOT)) {
> 		__io_req_complete(req, issue_flags, ret, 0);
> 		return 0;
> 	}
> 	if (ret >= 0) {
> 		bool filled;
> 
> 		spin_lock(&ctx->completion_lock);
> 		filled = io_fill_cqe_aux(ctx, req->cqe.user_data, ret,
> 					 IORING_CQE_F_MORE);
> 		io_commit_cqring(ctx);
> 		spin_unlock(&ctx->completion_lock);
> 		if (filled) {
> 			io_cqring_ev_posted(ctx);
> 			goto retry;
> 		}
> 		/* fall through to error case */
> 		ret = -ECANCELED;
> 	}
> 
> 	/*
> 	 * the apoll multishot req should handle poll
> 	 * cancellation by itself since the upper layer
> 	 * who called io_queue_sqe() cannot get errors
> 	 * happened here.
> 	 */
> 	io_poll_clean(req);
> 	return ret;
> 
> which I think is a lot easier to read and keeps the indentation at a
> manageable level and reduces duplicate code.
Great, thanks, it's better.
>