Skip to contentSkip to conversation

nvmet-tcp: complete command on H2C protocol errors

PATCH New Xixin Liu 1 participant 2 messages Jun 4

Subsystems (2)
lore
Share & tools
lore

Xixin Liu

author
#0 lore raw

cover nvmet-tcp: complete command on H2C protocol errors

Hi,

This series fixes the H2C Data PDU error path in nvmet-tcp.

When nvmet_tcp_handle_h2c_data_pdu() hits err_proto it used to return -EPROTO without completing the associated command (there is even a FIXME in that path). The host often only sees the connection drop. Complete commands that already have a valid transfer tag with nvmet_req_complete() before returning -EPROTO. Skip completion when req->sq is unset (pre-connect placeholder).

Tested: make M=drivers/nvme/target on linux-next; nvmet TCP target with null-blk backend; nvme connect -t tcp loopback read smoke test; malformed H2C Data PDU before Fabrics Connect and on an active Write without kernel oops.

Thanks,
Xixin Liu

---

Xixin Liu (1):
nvmet-tcp: complete command on H2C protocol errors

 drivers/nvme/target/tcp.c | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)
Signature · 1 line
2.43.0
author
#1 lore raw

PATCH 1/1 nvmet-tcp: complete command on H2C protocol errors

When nvmet_tcp_handle_h2c_data_pdu() detects a malformed H2C PDU it returned -EPROTO without completing the associated NVMe command. The host only saw a transport teardown. Complete commands that are already associated with a valid transfer tag before returning -EPROTO.

Use nvmet_req_complete() so the host receives a failed completion.

Pick status codes by failure class: CMD_SEQ_ERROR for offset/sequence and command-state problems; DATA_XFER_ERROR for length/plen mismatch. Both are fatal at the NVMe/TCP transport layer, but the command should still be failed explicitly. Use DNR because retrying the same malformed PDU will not help.

Only call nvmet_req_complete() when req->sq is set (the command went through nvmet_req_init()). The pre-connect placeholder has no sq and must not be completed.

When the transfer tag is out of range there is no command to complete; only reset the receiver. Initialize cmd to NULL so err_proto never dereferences an uninitialized pointer.

---
 drivers/nvme/target/tcp.c | 18 ++++++++++++++++--
 1 file changed, 16 insertions(+), 2 deletions(-)
drivers/nvme/target/tcp.c +16 -2
index 20f150d17a96..9a9d3a354566 100644
@@ -977,10 +977,21 @@ static int nvmet_tcp_handle_req_failure(struct nvmet_tcp_queue *queue,
 	return ret;
 }
 
+static void nvmet_tcp_fail_h2c_cmd(struct nvmet_tcp_cmd *cmd, u16 status)
+{
+	if (!cmd->req.sq)
+		return;
+
+	if (!(cmd->flags & NVMET_TCP_F_INIT_FAILED))
+		nvmet_req_complete(&cmd->req, status | NVME_STATUS_DNR);
+	nvmet_tcp_free_cmd_buffers(cmd);
+}
+
 static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
 {
 	struct nvme_tcp_data_pdu *data = &queue->pdu.data;
-	struct nvmet_tcp_cmd *cmd;
+	struct nvmet_tcp_cmd *cmd = NULL;
+	u16 err_status = NVME_SC_CMD_SEQ_ERROR;
 	unsigned int exp_data_len;
 
 	if (likely(queue->nr_cmds)) {
@@ -1011,6 +1022,7 @@ static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
 		     cmd->pdu_len == 0 ||
 		     cmd->pdu_len > NVMET_TCP_MAXH2CDATA)) {
 		pr_err("H2CData PDU len %u is invalid\n", cmd->pdu_len);
+		err_status = NVME_SC_DATA_XFER_ERROR;
 		goto err_proto;
 	}
        /*
@@ -1036,7 +1048,9 @@ static int nvmet_tcp_handle_h2c_data_pdu(struct nvmet_tcp_queue *queue)
 	return 0;
 
 err_proto:
-	/* FIXME: use proper transport errors */
+	if (cmd)
+		nvmet_tcp_fail_h2c_cmd(cmd, err_status);
+	nvmet_prepare_receive_pdu(queue);
 	return -EPROTO;
 }
 
Signature · 1 line
2.43.0