Skip to contentSkip to conversation

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

PATCH In review Sagi Grimberg 2 participants 4 messages Mar 12, 2020 – Mar 13, 2020

Subsystems (2)
lore
Share & tools
lore

Sagi Grimberg

author maintainer
#0 lore raw

PATCH v2 1/2 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 70583295388a ("nvmet-tcp: implement C2HData SUCCESS optimization")

---
- No changes from v1

 drivers/nvme/target/tcp.c | 12 +++++++++---
 1 file changed, 9 insertions(+), 3 deletions(-)
drivers/nvme/target/tcp.c +9 -3
index 1f4322ee7dbe..e5d000f12059 100644
@@ -525,7 +525,7 @@ 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;
@@ -533,9 +533,15 @@ static int nvmet_try_send_data(struct nvmet_tcp_cmd *cmd)
 	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;
 
@@ -666,7 +672,7 @@ 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;
 	}
Signature · 7 lines
2.20.1


_______________________________________________
linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme
author maintainer
#1 lore raw

PATCH v2 2/2 nvmet-tcp: optimize tcp stack TX when data digest is used

If we have a 4-byte data digest to send to the wire, but we have more data to send, set MSG_MORE to tell the stack that more is coming.

---
Changes from v1:
- fix compilation error

 drivers/nvme/target/tcp.c | 7 +++++--
 1 file changed, 5 insertions(+), 2 deletions(-)
drivers/nvme/target/tcp.c +5 -2
index e5d000f12059..f4de7b075284 100644
@@ -628,7 +628,7 @@ static int nvmet_try_send_r2t(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
 	return 1;
 }
 
-static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd)
+static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd, bool last_in_batch)
 {
 	struct nvmet_tcp_queue *queue = cmd->queue;
 	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
@@ -638,6 +638,9 @@ static int nvmet_try_send_ddgst(struct nvmet_tcp_cmd *cmd)
 	};
 	int ret;
 
+	if (!last_in_batch && cmd->queue->send_list_len)
+		msg.msg_flags |= MSG_MORE;
+
 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
 	if (unlikely(ret <= 0))
 		return ret;
@@ -678,7 +681,7 @@ static int nvmet_tcp_try_send_one(struct nvmet_tcp_queue *queue,
 	}
 
 	if (cmd->state == NVMET_TCP_SEND_DDGST) {
-		ret = nvmet_try_send_ddgst(cmd);
+		ret = nvmet_try_send_ddgst(cmd, last_in_batch);
 		if (ret <= 0)
 			goto done_send;
 	}
Signature · 7 lines
2.20.1


_______________________________________________
linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme
Reviewed-by on v2 2/2
#2 lore raw

RE: [PATCH v2 2/2] nvmet-tcp: optimize tcp stack TX when data digest is used

Quoted 8 lines

Subject: [PATCH v2 2/2] nvmet-tcp: optimize tcp stack TX when data digest is used

If we have a 4-byte data digest to send to the wire, but we have more data to send, set MSG_MORE to tell the stack that more is coming.

---
Changes from v1:
- fix compilation error

Looks good now.

_______________________________________________
linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme

Tested-by on v2 1/2
#3 lore raw
In reply to Sagi Grimberg
Quoted 8 lines

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)

Quoted 6 lines

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 70583295388a ("nvmet-tcp: implement C2HData SUCCESS optimization")

---

_______________________________________________
linux-nvme mailing list
[email protected]
http://lists.infradead.org/mailman/listinfo/linux-nvme