nvmet-tcp: set MSG_MORE only if we actually have more to send

When we send PDU data, we want to optimize the tcp stack
operation if we have more data to send. So when we set MSG_MORE
when:
- We have more fragments coming in the batch, or
- We have a more data to send in this PDU
- We don't have a data digest trailer
- We optimize with the SUCCESS flag and omit the NVMe completion
  (used if sq_head pointer update is disabled)

This addresses a regression in QD=1 with SUCCESS flag optimization
as we unconditionally set MSG_MORE when we didn't actually have
more data to send.

Fixes: 7058329 ("nvmet-tcp: implement C2HData SUCCESS optimization")
Reported-by: Mark Wunderlich <[email protected]>
Tested-by: Mark Wunderlich <[email protected]>
Signed-off-by: Sagi Grimberg <[email protected]>
Signed-off-by: Keith Busch <[email protected]>

Author: sagigrimberg

drivers/nvme/target/tcp.c

@@ -515,17 +515,23 @@ static int nvmet_try_send_data_pdu(struct nvmet_tcp_cmd *cmd)
 	return 1;
 }
 
-static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd)
+static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
 {
 	struct nvmet_tcp_queue *queue = cmd->queue;
 	int ret;
 
 	while (cmd->cur_sg) {
 		struct page *page = sg_page(cmd->cur_sg);
 		u32 left = cmd->cur_sg->length - cmd->offset;
+		int flags = MSG_DONTWAIT;
+
+		if ((!last_in_batch && cmd->queue->send_list_len) ||
+		    cmd->wbytes_done + left < cmd->req.transfer_len ||
+		    queue->data_digest || !queue->nvme_sq.sqhd_disabled)
+			flags |= MSG_MORE;
 
 		ret = kernel_sendpage(cmd->queue->sock, page, cmd->offset,
-					left, MSG_DONTWAIT | MSG_MORE);
+					left, flags);
 		if (ret <= 0)
 			return ret;
 
@@ -660,7 +666,7 @@ static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
 	}
 
 	if (cmd->state == NVMET_TCP_SEND_DATA) {
-		ret = nvmet_try_send_data(cmd);
+		ret = nvmet_try_send_data(cmd, last_in_batch);
 		if (ret <= 0)
 			goto done_send;
 	}