Skip to contentSkip to conversation

nvme-tcp: Do not reset transport on data digest errors

PATCH Superseded Daniel Wagner 3 participants 8 messages Aug 5, 2021 – Aug 23, 2021 NVMe

v1 of 6 Superseded Latest is v6 →

lore
Share & tools
lore
#0 lore raw

PATCH 1/1 nvme-tcp: Do not reset transport on data digest errors

The spec says

7.4.6.1 Digest Error handling

When a host detects a data digest error in a C2HData PDU, that host shall continue processing C2HData PDUs associated with the command and when the command processing has completed, if a successful status was returned by the controller, the host shall fail the command with a non-fatal transport error.

Currently the transport is reseted when a data digest error is detected. To fix this, keep track of the final status in the queue object and use it when completing the request.

The new member can be placed adjacent to the receive related members and fits in the cacheline as there is a 4 byte hole.

---

Hi,

I've tested this by modifying the receive path. Via the fault_inject interface I injecting wrong hash values. The request would then be completed with status != 0 and nvme_decide_disposition decices to retry the request. So this seems be in more sync with what the spec says on this topic.

Daniel

 drivers/nvme/host/tcp.c | 11 +++++++----
 1 file changed, 7 insertions(+), 4 deletions(-)
drivers/nvme/host/tcp.c +7 -4
index 097f7dd10ed3..5253147df4c7 100644
@@ -89,6 +89,7 @@ struct nvme_tcp_queue {
 	size_t			data_remaining;
 	size_t			ddgst_remaining;
 	unsigned int		nr_cqe;
+	u16			status;
 
 	/* send state */
 	struct nvme_tcp_request *request;
@@ -496,7 +497,8 @@ static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
 		return -EINVAL;
 	}
 
-	if (!nvme_try_complete_req(rq, cqe->status, cqe->result))
+	if (!nvme_try_complete_req(rq, queue->status ?
+			queue->status : cqe->status, cqe->result))
 		nvme_complete_rq(rq);
 	queue->nr_cqe++;
 
@@ -676,6 +678,7 @@ static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 
 	switch (hdr->type) {
 	case nvme_tcp_c2h_data:
+		queue->status = NVME_SC_SUCCESS;
 		return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
 	case nvme_tcp_rsp:
 		nvme_tcp_init_recv_ctx(queue);
@@ -758,7 +761,7 @@ static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 			queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
 		} else {
 			if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
-				nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
+				nvme_tcp_end_request(rq, queue->status);
 				queue->nr_cqe++;
 			}
 			nvme_tcp_init_recv_ctx(queue);
@@ -792,14 +795,14 @@ static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
 			"data digest error: recv %#x expected %#x\n",
 			le32_to_cpu(queue->recv_ddgst),
 			le32_to_cpu(queue->exp_ddgst));
-		return -EIO;
+		queue->status = NVME_SC_DATA_XFER_ERROR;
 	}
 
 	if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
 		struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
 					pdu->command_id);
 
-		nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
+		nvme_tcp_end_request(rq, queue->status);
 		queue->nr_cqe++;
 	}
 
Signature · 1 line
2.29.2
maintainer
#1 lore raw
Quoted 39 lines

The spec says

7.4.6.1 Digest Error handling

When a host detects a data digest error in a C2HData PDU, that host shall continue processing C2HData PDUs associated with the command and when the command processing has completed, if a successful status was returned by the controller, the host shall fail the command with a non-fatal transport error.

Currently the transport is reseted when a data digest error is detected. To fix this, keep track of the final status in the queue object and use it when completing the request.

The new member can be placed adjacent to the receive related members and fits in the cacheline as there is a 4 byte hole.

---

Hi,

I've tested this by modifying the receive path. Via the fault_inject interface I injecting wrong hash values. The request would then be completed with status != 0 and nvme_decide_disposition decices to retry the request. So this seems be in more sync with what the spec says on this topic.

Daniel

  drivers/nvme/host/tcp.c | 11 +++++++----
  1 file changed, 7 insertions(+), 4 deletions(-)
drivers/nvme/host/tcp.c +0 -0
index 097f7dd10ed3..5253147df4c7 100644

drivers/nvme/host/tcp.c struct nvme_tcp_queue { · line 92

Quoted 4 lines
drivers/nvme/host/tcp.c +1 -0
@@ -89,6 +89,7 @@ struct nvme_tcp_queue {
  	size_t			data_remaining;
  	size_t			ddgst_remaining;
  	unsigned int		nr_cqe;
+	u16			status;

Why is this a queue member and not a request member?

Quoted 48 lines
drivers/nvme/host/tcp.c +6 -4
  	/* send state */
  	struct nvme_tcp_request *request;
@@ -496,7 +497,8 @@ static int nvme_tcp_process_nvme_cqe(struct nvme_tcp_queue *queue,
  		return -EINVAL;
  	}
 
-	if (!nvme_try_complete_req(rq, cqe->status, cqe->result))
+	if (!nvme_try_complete_req(rq, queue->status ?
+			queue->status : cqe->status, cqe->result))
  		nvme_complete_rq(rq);
  	queue->nr_cqe++;
 
@@ -676,6 +678,7 @@ static int nvme_tcp_recv_pdu(struct nvme_tcp_queue *queue, struct sk_buff *skb,
 
  	switch (hdr->type) {
  	case nvme_tcp_c2h_data:
+		queue->status = NVME_SC_SUCCESS;
  		return nvme_tcp_handle_c2h_data(queue, (void *)queue->pdu);
  	case nvme_tcp_rsp:
  		nvme_tcp_init_recv_ctx(queue);
@@ -758,7 +761,7 @@ static int nvme_tcp_recv_data(struct nvme_tcp_queue *queue, struct sk_buff *skb,
  			queue->ddgst_remaining = NVME_TCP_DIGEST_LENGTH;
  		} else {
  			if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
-				nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
+				nvme_tcp_end_request(rq, queue->status);
  				queue->nr_cqe++;
  			}
  			nvme_tcp_init_recv_ctx(queue);
@@ -792,14 +795,14 @@ static int nvme_tcp_recv_ddgst(struct nvme_tcp_queue *queue,
  			"data digest error: recv %#x expected %#x\n",
  			le32_to_cpu(queue->recv_ddgst),
  			le32_to_cpu(queue->exp_ddgst));
-		return -EIO;
+		queue->status = NVME_SC_DATA_XFER_ERROR;
  	}
 
  	if (pdu->hdr.flags & NVME_TCP_F_DATA_SUCCESS) {
  		struct request *rq = nvme_cid_to_rq(nvme_tcp_tagset(queue),
  					pdu->command_id);
 
-		nvme_tcp_end_request(rq, NVME_SC_SUCCESS);
+		nvme_tcp_end_request(rq, queue->status);
  		queue->nr_cqe++;
  	}
 
#2 lore raw

Hi Sagi,

Quoted 7 lines · Sagi Grimberg

drivers/nvme/host/tcp.c struct nvme_tcp_queue { · line 92

Quoted 5 lines
drivers/nvme/host/tcp.c +1 -0
@@ -89,6 +89,7 @@ struct nvme_tcp_queue {
  	size_t			data_remaining;
  	size_t			ddgst_remaining;
  	unsigned int		nr_cqe;
+	u16			status;

Why is this a queue member and not a request member?

I was not sure if the TCP transport specific error handling should impact all other transport (size of struct request). Also I tried to avoid accessing cachelines which are not already in use. Except this I don't see there should be no problem to put this to struct request.

Daniel

maintainer
#3 lore raw
Quoted 12 lines

Hi Sagi,

Quoted 7 lines · Sagi Grimberg

drivers/nvme/host/tcp.c struct nvme_tcp_queue { · line 92

Quoted 5 lines
drivers/nvme/host/tcp.c +1 -0
@@ -89,6 +89,7 @@ struct nvme_tcp_queue {
   	size_t			data_remaining;
   	size_t			ddgst_remaining;
   	unsigned int		nr_cqe;
+	u16			status;

Why is this a queue member and not a request member?

Quoted 4 lines

I was not sure if the TCP transport specific error handling should impact all other transport (size of struct request). Also I tried to avoid accessing cachelines which are not already in use. Except this I don't see there should be no problem to put this to struct request.

It is the correct place, lets see that it doesn't increase the struct.

#4 lore raw
Quoted 12 lines · Sagi Grimberg
Quoted 12 lines

Hi Sagi,

Quoted 7 lines · Sagi Grimberg

drivers/nvme/host/tcp.c struct nvme_tcp_queue { · line 92

Quoted 5 lines
drivers/nvme/host/tcp.c +1 -0
@@ -89,6 +89,7 @@ struct nvme_tcp_queue {
   	size_t			data_remaining;
   	size_t			ddgst_remaining;
   	unsigned int		nr_cqe;
+	u16			status;

Why is this a queue member and not a request member?

Quoted 6 lines
Quoted 4 lines

I was not sure if the TCP transport specific error handling should impact all other transport (size of struct request). Also I tried to avoid accessing cachelines which are not already in use. Except this I don't see there should be no problem to put this to struct request.

It is the correct place, lets see that it doesn't increase the struct.

It should not according pahole:

/* XXX 7 bytes hole, try to pack */

void * pdu; /* 144 8 */
int pdu_remaining; /* 152 4 */
int pdu_offset; /* 156 4 */
size_t data_remaining; /* 160 8 */
size_t ddgst_remaining; /* 168 8 */
unsigned int nr_cqe; /* 176 4 */
u16 status; /* 180 2 */

/* XXX 2 bytes hole, try to pack */

Daniel

#5 lore raw
Quoted 3 lines · Daniel Wagner
Quoted 1 line

It is the correct place, lets see that it doesn't increase the struct.

It should not according pahole:

Anything I need to do to get this patch accepted?

maintainer
#6 lore raw
Quoted 6 lines
Quoted 4 lines
Quoted 2 lines

It is the correct place, lets see that it doesn't increase the struct.

It should not according pahole:

Anything I need to do to get this patch accepted?

Yes, move the status to nvme_tcp_request, I think its a more appropriate place for it.

Reviewed-by on v1 1/1
#7 lore raw
In reply to Daniel Wagner
Quoted 14 lines · Daniel Wagner

The spec says

7.4.6.1 Digest Error handling

When a host detects a data digest error in a C2HData PDU, that host shall continue processing C2HData PDUs associated with the command and when the command processing has completed, if a successful status was returned by the controller, the host shall fail the command with a non-fatal transport error.

Currently the transport is reseted when a data digest error is detected. To fix this, keep track of the final status in the queue object and use it when completing the request.

Quoted 6 lines

The new member can be placed adjacent to the receive related members and fits in the cacheline as there is a 4 byte hole.

---

Cheers,

Hannes

Signature · 4 lines
Dr. Hannes Reinecke                Kernel Storage Architect
[email protected]                              +49 911 74053 688
SUSE Software Solutions GmbH, Maxfeldstr. 5, 90409 Nürnberg
HRB 36809 (AG Nürnberg), Geschäftsführer: Felix Imendörffer