Skip to contentSkip to conversation

nvme-tcp: strict pdu pacing to avoid send stalls on TLS

PATCH Superseded Hannes Reinecke 3 participants 3 messages Apr 17, 2024 – Apr 18, 2024 NVMe

v1 of 3 Superseded Latest is v3 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 nvme-tcp: strict pdu pacing to avoid send stalls on TLS

TLS requires a strict pdu pacing via MSG_EOR to signal the end of a record and subsequent encryption. If we do not set MSG_EOR at the end of a sequence the record won't be closed, encryption doesn't start, and we end up with a send stall as the message will never be passed on to the TCP layer.
So do not check for the queue status when figuring out whether MSG_MORE should be set but rather make it dependent on the current command only.

---
 drivers/nvme/host/tcp.c | 11 +++--------
 1 file changed, 3 insertions(+), 8 deletions(-)
drivers/nvme/host/tcp.c +3 -8
index 2b821cbbdf1f..b460ebf72a1a 100644
@@ -1049,7 +1049,7 @@ static int nvme_tcp_try_send_data(struct nvme_tcp_request *req)
 		int req_data_sent = req->data_sent;
 		int ret;
 
-		if (last && !queue->data_digest && !nvme_tcp_queue_more(queue))
+		if (last && !queue->data_digest)
 			msg.msg_flags |= MSG_EOR;
 		else
 			msg.msg_flags |= MSG_MORE;
@@ -1105,7 +1105,7 @@ static int nvme_tcp_try_send_cmd_pdu(struct nvme_tcp_request *req)
 	int len = sizeof(*pdu) + hdgst - req->offset;
 	int ret;
 
-	if (inline_data || nvme_tcp_queue_more(queue))
+	if (inline_data)
 		msg.msg_flags |= MSG_MORE;
 	else
 		msg.msg_flags |= MSG_EOR;
@@ -1175,17 +1175,12 @@ static int nvme_tcp_try_send_ddgst(struct nvme_tcp_request *req)
 	size_t offset = req->offset;
 	u32 h2cdata_left = req->h2cdata_left;
 	int ret;
-	struct msghdr msg = { .msg_flags = MSG_DONTWAIT };
+	struct msghdr msg = { .msg_flags = MSG_DONTWAIT | MSG_EOR };
 	struct kvec iov = {
 		.iov_base = (u8 *)&req->ddgst + req->offset,
 		.iov_len = NVME_TCP_DIGEST_LENGTH - req->offset
 	};
 
-	if (nvme_tcp_queue_more(queue))
-		msg.msg_flags |= MSG_MORE;
-	else
-		msg.msg_flags |= MSG_EOR;
-
 	ret = kernel_sendmsg(queue->sock, &msg, &iov, 1, iov.iov_len);
 	if (unlikely(ret <= 0))
 		return ret;
Signature · 1 line
2.35.3
maintainer
#1 lore raw
Quoted 8 lines · Hannes Reinecke

TLS requires a strict pdu pacing via MSG_EOR to signal the end of a record and subsequent encryption. If we do not set MSG_EOR at the end of a sequence the record won't be closed, encryption doesn't start, and we end up with a send stall as the message will never be passed on to the TCP layer.
So do not check for the queue status when figuring out whether MSG_MORE should be set but rather make it dependent on the current command only.

How about making nvme_tcp_queue_more take into account nvme_tcp_tls()? so we preserve the behavior without tls.

i.e. something like:
--

drivers/nvme/host/tcp.c +0 -0
index 0ba62fc647b3..bbffc67f8a1e 100644
@@ -360,12 +360,18 @@ static inline void nvme_tcp_send_all(struct nvme_tcp_queue *queue)

        } while (ret > 0);
 }

-static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue) +static inline bool nvme_tcp_queue_has_pending(struct nvme_tcp_queue *queue) {
        return !list_empty(&queue->send_list) || !llist_empty(&queue->req_list);
 }

+static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
+{
+       return !nvme_tcp_tls(queue->ctrl) &&
+               nvme_tcp_queue_has_pending(queue);
+}
+
 static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req, bool sync, bool last)
 {
@@ -386,7 +392,7 @@ static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req, mutex_unlock(&queue->send_mutex);
        }

-       if (last && nvme_tcp_queue_more(queue))
+       if (last && nvme_tcp_queue_has_pending(queue)) queue_work_on(queue->io_cpu, nvme_tcp_wq, &queue->io_work); }
--

#2 lore raw
Quoted 52 lines · Sagi Grimberg
Quoted 8 lines · Hannes Reinecke

TLS requires a strict pdu pacing via MSG_EOR to signal the end of a record and subsequent encryption. If we do not set MSG_EOR at the end of a sequence the record won't be closed, encryption doesn't start, and we end up with a send stall as the message will never be passed on to the TCP layer.
So do not check for the queue status when figuring out whether MSG_MORE should be set but rather make it dependent on the current command only.

How about making nvme_tcp_queue_more take into account nvme_tcp_tls()? so we preserve the behavior without tls.

i.e. something like:

drivers/nvme/host/tcp.c +7 -1
index 0ba62fc647b3..bbffc67f8a1e 100644
@@ -360,12 +360,18 @@ static inline void nvme_tcp_send_all(struct 
 nvme_tcp_queue *queue)
         } while (ret > 0);
  }
 
-static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
+static inline bool nvme_tcp_queue_has_pending(struct nvme_tcp_queue 
 *queue)
  {
         return !list_empty(&queue->send_list) ||
                 !llist_empty(&queue->req_list);
  }
 
+static inline bool nvme_tcp_queue_more(struct nvme_tcp_queue *queue)
+{
+       return !nvme_tcp_tls(queue->ctrl) &&
+               nvme_tcp_queue_has_pending(queue);
+}
+
  static inline void nvme_tcp_queue_request(struct nvme_tcp_request *req,

                bool sync, bool last)
 {

drivers/nvme/host/tcp.c +1 -1
@@ -386,7 +392,7 @@ static inline void nvme_tcp_queue_request(struct 
 nvme_tcp_request *req,
                 mutex_unlock(&queue->send_mutex);
         }
 
-       if (last && nvme_tcp_queue_more(queue))
+       if (last && nvme_tcp_queue_has_pending(queue))
                 queue_work_on(queue->io_cpu, nvme_tcp_wq, 
 &queue->io_work);
Quoted 3 lines

 }

Would work as well, I guess.
I'll give it a go.

Cheers,

Hannes