[dpdk-dev] [PATCH 0/4] net/szedata2: fixes or improvements

[dpdk-dev] [PATCH 0/4] net/szedata2: fixes or improvements

[Matej Vido]

Matej Vido (4):
  net/szedata2: fix total stats
  net/szedata2: use dynamically allocated queues
  net/szedata2: add stat of mbuf allocation failures
  net/szedata2: fix format string for pci address

 drivers/net/szedata2/rte_eth_szedata2.c | 171 ++++++++++++++++++++++----------
 1 file changed, 121 insertions(+), 50 deletions(-)

-- 
1.8.3.1

[dpdk-dev] [PATCH 1/4] net/szedata2: fix total stats

[Matej Vido]

Counters from all queues have to be summed up for total stats
even though the number of queue stats counters is not sufficient.

Fixes: 83556fd2c0fc ("szedata2: change to physical device type")
Cc: stable@dpdk.org

Signed-off-by: Matej Vido <vido@cesnet.cz>
---
 drivers/net/szedata2/rte_eth_szedata2.c | 33 ++++++++++++++++++++-------------
 1 file changed, 20 insertions(+), 13 deletions(-)

diff --git a/drivers/net/szedata2/rte_eth_szedata2.c b/drivers/net/szedata2/rte_eth_szedata2.c
index 3cfe388..fc11d68 100644
--- a/drivers/net/szedata2/rte_eth_szedata2.c
+++ b/drivers/net/szedata2/rte_eth_szedata2.c
@@ -1058,22 +1058,29 @@ struct pmd_internals {
 	uint64_t tx_err_total = 0;
 	uint64_t rx_total_bytes = 0;
 	uint64_t tx_total_bytes = 0;
-	const struct pmd_internals *internals = dev->data->dev_private;
 
-	for (i = 0; i < RTE_ETHDEV_QUEUE_STAT_CNTRS && i < nb_rx; i++) {
-		stats->q_ipackets[i] = internals->rx_queue[i].rx_pkts;
-		stats->q_ibytes[i] = internals->rx_queue[i].rx_bytes;
-		rx_total += stats->q_ipackets[i];
-		rx_total_bytes += stats->q_ibytes[i];
+	for (i = 0; i < nb_rx; i++) {
+		struct szedata2_rx_queue *rxq = dev->data->rx_queues[i];
+
+		if (i < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
+			stats->q_ipackets[i] = rxq->rx_pkts;
+			stats->q_ibytes[i] = rxq->rx_bytes;
+		}
+		rx_total += rxq->rx_pkts;
+		rx_total_bytes += rxq->rx_bytes;
 	}
 
-	for (i = 0; i < RTE_ETHDEV_QUEUE_STAT_CNTRS && i < nb_tx; i++) {
-		stats->q_opackets[i] = internals->tx_queue[i].tx_pkts;
-		stats->q_obytes[i] = internals->tx_queue[i].tx_bytes;
-		stats->q_errors[i] = internals->tx_queue[i].err_pkts;
-		tx_total += stats->q_opackets[i];
-		tx_total_bytes += stats->q_obytes[i];
-		tx_err_total += stats->q_errors[i];
+	for (i = 0; i < nb_tx; i++) {
+		struct szedata2_tx_queue *txq = dev->data->tx_queues[i];
+
+		if (i < RTE_ETHDEV_QUEUE_STAT_CNTRS) {
+			stats->q_opackets[i] = txq->tx_pkts;
+			stats->q_obytes[i] = txq->tx_bytes;
+			stats->q_errors[i] = txq->err_pkts;
+		}
+		tx_total += txq->tx_pkts;
+		tx_total_bytes += txq->tx_bytes;
+		tx_err_total += txq->err_pkts;
 	}
 
 	stats->ipackets = rx_total;
-- 
1.8.3.1

[dpdk-dev] [PATCH 2/4] net/szedata2: use dynamically allocated queues

[Matej Vido]

Previously the queues were the part of private data structure of the
Ethernet device.
Now the queues are allocated at setup thus numa-aware allocation is
possible.

Signed-off-by: Matej Vido <vido@cesnet.cz>
---
 drivers/net/szedata2/rte_eth_szedata2.c | 99 ++++++++++++++++++++++++---------
 1 file changed, 74 insertions(+), 25 deletions(-)

diff --git a/drivers/net/szedata2/rte_eth_szedata2.c b/drivers/net/szedata2/rte_eth_szedata2.c
index fc11d68..f41716d 100644
--- a/drivers/net/szedata2/rte_eth_szedata2.c
+++ b/drivers/net/szedata2/rte_eth_szedata2.c
@@ -86,8 +86,6 @@ struct szedata2_tx_queue {
 };
 
 struct pmd_internals {
-	struct szedata2_rx_queue rx_queue[RTE_ETH_SZEDATA2_MAX_RX_QUEUES];
-	struct szedata2_tx_queue tx_queue[RTE_ETH_SZEDATA2_MAX_TX_QUEUES];
 	uint16_t max_rx_queues;
 	uint16_t max_tx_queues;
 	char sze_dev[PATH_MAX];
@@ -1098,17 +1096,18 @@ struct pmd_internals {
 	uint16_t i;
 	uint16_t nb_rx = dev->data->nb_rx_queues;
 	uint16_t nb_tx = dev->data->nb_tx_queues;
-	struct pmd_internals *internals = dev->data->dev_private;
 
 	for (i = 0; i < nb_rx; i++) {
-		internals->rx_queue[i].rx_pkts = 0;
-		internals->rx_queue[i].rx_bytes = 0;
-		internals->rx_queue[i].err_pkts = 0;
+		struct szedata2_rx_queue *rxq = dev->data->rx_queues[i];
+		rxq->rx_pkts = 0;
+		rxq->rx_bytes = 0;
+		rxq->err_pkts = 0;
 	}
 	for (i = 0; i < nb_tx; i++) {
-		internals->tx_queue[i].tx_pkts = 0;
-		internals->tx_queue[i].tx_bytes = 0;
-		internals->tx_queue[i].err_pkts = 0;
+		struct szedata2_tx_queue *txq = dev->data->tx_queues[i];
+		txq->tx_pkts = 0;
+		txq->tx_bytes = 0;
+		txq->err_pkts = 0;
 	}
 }
 
@@ -1116,9 +1115,11 @@ struct pmd_internals {
 eth_rx_queue_release(void *q)
 {
 	struct szedata2_rx_queue *rxq = (struct szedata2_rx_queue *)q;
-	if (rxq->sze != NULL) {
-		szedata_close(rxq->sze);
-		rxq->sze = NULL;
+
+	if (rxq != NULL) {
+		if (rxq->sze != NULL)
+			szedata_close(rxq->sze);
+		rte_free(rxq);
 	}
 }
 
@@ -1126,9 +1127,11 @@ struct pmd_internals {
 eth_tx_queue_release(void *q)
 {
 	struct szedata2_tx_queue *txq = (struct szedata2_tx_queue *)q;
-	if (txq->sze != NULL) {
-		szedata_close(txq->sze);
-		txq->sze = NULL;
+
+	if (txq != NULL) {
+		if (txq->sze != NULL)
+			szedata_close(txq->sze);
+		rte_free(txq);
 	}
 }
 
@@ -1263,23 +1266,42 @@ struct pmd_internals {
 eth_rx_queue_setup(struct rte_eth_dev *dev,
 		uint16_t rx_queue_id,
 		uint16_t nb_rx_desc __rte_unused,
-		unsigned int socket_id __rte_unused,
+		unsigned int socket_id,
 		const struct rte_eth_rxconf *rx_conf __rte_unused,
 		struct rte_mempool *mb_pool)
 {
 	struct pmd_internals *internals = dev->data->dev_private;
-	struct szedata2_rx_queue *rxq = &internals->rx_queue[rx_queue_id];
+	struct szedata2_rx_queue *rxq;
 	int ret;
 	uint32_t rx = 1 << rx_queue_id;
 	uint32_t tx = 0;
 
+	if (dev->data->rx_queues[rx_queue_id] != NULL) {
+		eth_rx_queue_release(dev->data->rx_queues[rx_queue_id]);
+		dev->data->rx_queues[rx_queue_id] = NULL;
+	}
+
+	rxq = rte_zmalloc_socket("szedata2 rx queue",
+			sizeof(struct szedata2_rx_queue),
+			RTE_CACHE_LINE_SIZE, socket_id);
+	if (rxq == NULL) {
+		RTE_LOG(ERR, PMD, "rte_zmalloc_socket() failed for rx queue id "
+				"%" PRIu16 "!\n", rx_queue_id);
+		return -ENOMEM;
+	}
+
 	rxq->sze = szedata_open(internals->sze_dev);
-	if (rxq->sze == NULL)
+	if (rxq->sze == NULL) {
+		RTE_LOG(ERR, PMD, "szedata_open() failed for rx queue id "
+				"%" PRIu16 "!\n", rx_queue_id);
+		eth_rx_queue_release(rxq);
 		return -EINVAL;
+	}
 	ret = szedata_subscribe3(rxq->sze, &rx, &tx);
 	if (ret != 0 || rx == 0) {
-		szedata_close(rxq->sze);
-		rxq->sze = NULL;
+		RTE_LOG(ERR, PMD, "szedata_subscribe3() failed for rx queue id "
+				"%" PRIu16 "!\n", rx_queue_id);
+		eth_rx_queue_release(rxq);
 		return -EINVAL;
 	}
 	rxq->rx_channel = rx_queue_id;
@@ -1290,6 +1312,10 @@ struct pmd_internals {
 	rxq->err_pkts = 0;
 
 	dev->data->rx_queues[rx_queue_id] = rxq;
+
+	RTE_LOG(DEBUG, PMD, "Configured rx queue id %" PRIu16 " on socket "
+			"%u.\n", rx_queue_id, socket_id);
+
 	return 0;
 }
 
@@ -1297,22 +1323,41 @@ struct pmd_internals {
 eth_tx_queue_setup(struct rte_eth_dev *dev,
 		uint16_t tx_queue_id,
 		uint16_t nb_tx_desc __rte_unused,
-		unsigned int socket_id __rte_unused,
+		unsigned int socket_id,
 		const struct rte_eth_txconf *tx_conf __rte_unused)
 {
 	struct pmd_internals *internals = dev->data->dev_private;
-	struct szedata2_tx_queue *txq = &internals->tx_queue[tx_queue_id];
+	struct szedata2_tx_queue *txq;
 	int ret;
 	uint32_t rx = 0;
 	uint32_t tx = 1 << tx_queue_id;
 
+	if (dev->data->tx_queues[tx_queue_id] != NULL) {
+		eth_tx_queue_release(dev->data->tx_queues[tx_queue_id]);
+		dev->data->tx_queues[tx_queue_id] = NULL;
+	}
+
+	txq = rte_zmalloc_socket("szedata2 tx queue",
+			sizeof(struct szedata2_tx_queue),
+			RTE_CACHE_LINE_SIZE, socket_id);
+	if (txq == NULL) {
+		RTE_LOG(ERR, PMD, "rte_zmalloc_socket() failed for tx queue id "
+				"%" PRIu16 "!\n", tx_queue_id);
+		return -ENOMEM;
+	}
+
 	txq->sze = szedata_open(internals->sze_dev);
-	if (txq->sze == NULL)
+	if (txq->sze == NULL) {
+		RTE_LOG(ERR, PMD, "szedata_open() failed for tx queue id "
+				"%" PRIu16 "!\n", tx_queue_id);
+		eth_tx_queue_release(txq);
 		return -EINVAL;
+	}
 	ret = szedata_subscribe3(txq->sze, &rx, &tx);
 	if (ret != 0 || tx == 0) {
-		szedata_close(txq->sze);
-		txq->sze = NULL;
+		RTE_LOG(ERR, PMD, "szedata_subscribe3() failed for tx queue id "
+				"%" PRIu16 "!\n", tx_queue_id);
+		eth_tx_queue_release(txq);
 		return -EINVAL;
 	}
 	txq->tx_channel = tx_queue_id;
@@ -1321,6 +1366,10 @@ struct pmd_internals {
 	txq->err_pkts = 0;
 
 	dev->data->tx_queues[tx_queue_id] = txq;
+
+	RTE_LOG(DEBUG, PMD, "Configured tx queue id %" PRIu16 " on socket "
+			"%u.\n", tx_queue_id, socket_id);
+
 	return 0;
 }
 
-- 
1.8.3.1

[dpdk-dev] [PATCH 3/4] net/szedata2: add stat of mbuf allocation failures

[Matej Vido]

Signed-off-by: Matej Vido <vido@cesnet.cz>
---
 drivers/net/szedata2/rte_eth_szedata2.c | 31 +++++++++++++++++++++++--------
 1 file changed, 23 insertions(+), 8 deletions(-)

diff --git a/drivers/net/szedata2/rte_eth_szedata2.c b/drivers/net/szedata2/rte_eth_szedata2.c
index f41716d..8278780 100644
--- a/drivers/net/szedata2/rte_eth_szedata2.c
+++ b/drivers/net/szedata2/rte_eth_szedata2.c
@@ -67,7 +67,16 @@
 
 #define SZEDATA2_DEV_PATH_FMT "/dev/szedataII%u"
 
+struct pmd_internals {
+	struct rte_eth_dev *dev;
+	uint16_t max_rx_queues;
+	uint16_t max_tx_queues;
+	char sze_dev[PATH_MAX];
+	struct rte_mem_resource *pci_rsc;
+};
+
 struct szedata2_rx_queue {
+	struct pmd_internals *priv;
 	struct szedata *sze;
 	uint8_t rx_channel;
 	uint16_t in_port;
@@ -78,6 +87,7 @@ struct szedata2_rx_queue {
 };
 
 struct szedata2_tx_queue {
+	struct pmd_internals *priv;
 	struct szedata *sze;
 	uint8_t tx_channel;
 	volatile uint64_t tx_pkts;
@@ -85,13 +95,6 @@ struct szedata2_tx_queue {
 	volatile uint64_t err_pkts;
 };
 
-struct pmd_internals {
-	uint16_t max_rx_queues;
-	uint16_t max_tx_queues;
-	char sze_dev[PATH_MAX];
-	struct rte_mem_resource *pci_rsc;
-};
-
 static struct ether_addr eth_addr = {
 	.addr_bytes = { 0x00, 0x11, 0x17, 0x00, 0x00, 0x00 }
 };
@@ -130,8 +133,10 @@ struct pmd_internals {
 	for (i = 0; i < nb_pkts; i++) {
 		mbuf = rte_pktmbuf_alloc(sze_q->mb_pool);
 
-		if (unlikely(mbuf == NULL))
+		if (unlikely(mbuf == NULL)) {
+			sze_q->priv->dev->data->rx_mbuf_alloc_failed++;
 			break;
+		}
 
 		/* get the next sze packet */
 		if (sze->ct_rx_lck != NULL && !sze->ct_rx_rem_bytes &&
@@ -351,6 +356,8 @@ struct pmd_internals {
 	uint16_t packet_len1 = 0;
 	uint16_t packet_len2 = 0;
 	uint16_t hw_data_align;
+	uint64_t *mbuf_failed_ptr =
+		&sze_q->priv->dev->data->rx_mbuf_alloc_failed;
 
 	if (unlikely(sze_q->sze == NULL || nb_pkts == 0))
 		return 0;
@@ -538,6 +545,7 @@ struct pmd_internals {
 			sze->ct_rx_lck = ct_rx_lck_backup;
 			sze->ct_rx_rem_bytes = ct_rx_rem_bytes_backup;
 			sze->ct_rx_cur_ptr = ct_rx_cur_ptr_backup;
+			sze_q->priv->dev->data->rx_mbuf_alloc_failed++;
 			break;
 		}
 
@@ -587,6 +595,7 @@ struct pmd_internals {
 						ct_rx_rem_bytes_backup;
 					sze->ct_rx_cur_ptr =
 						ct_rx_cur_ptr_backup;
+					(*mbuf_failed_ptr)++;
 					goto finish;
 				}
 
@@ -630,6 +639,7 @@ struct pmd_internals {
 							ct_rx_rem_bytes_backup;
 						sze->ct_rx_cur_ptr =
 							ct_rx_cur_ptr_backup;
+						(*mbuf_failed_ptr)++;
 						goto finish;
 					}
 
@@ -1086,6 +1096,7 @@ struct pmd_internals {
 	stats->ibytes = rx_total_bytes;
 	stats->obytes = tx_total_bytes;
 	stats->oerrors = tx_err_total;
+	stats->rx_nombuf = dev->data->rx_mbuf_alloc_failed;
 
 	return 0;
 }
@@ -1290,6 +1301,7 @@ struct pmd_internals {
 		return -ENOMEM;
 	}
 
+	rxq->priv = internals;
 	rxq->sze = szedata_open(internals->sze_dev);
 	if (rxq->sze == NULL) {
 		RTE_LOG(ERR, PMD, "szedata_open() failed for rx queue id "
@@ -1346,6 +1358,7 @@ struct pmd_internals {
 		return -ENOMEM;
 	}
 
+	txq->priv = internals;
 	txq->sze = szedata_open(internals->sze_dev);
 	if (txq->sze == NULL) {
 		RTE_LOG(ERR, PMD, "szedata_open() failed for tx queue id "
@@ -1543,6 +1556,8 @@ struct pmd_internals {
 			pci_addr->domain, pci_addr->bus, pci_addr->devid,
 			pci_addr->function);
 
+	internals->dev = dev;
+
 	/* Get index of szedata2 device file and create path to device file */
 	ret = get_szedata2_index(pci_addr, &szedata2_index);
 	if (ret != 0) {
-- 
1.8.3.1

[dpdk-dev] [PATCH 4/4] net/szedata2: fix format string for pci address

[Matej Vido]

For fscanf() function SCN macros should be used but PRI macros were
wrongly used.
Also use correct sizes of variables for read values.

Fixes: 83556fd2c0fc ("szedata2: change to physical device type")
Cc: stable@dpdk.org

Signed-off-by: Matej Vido <vido@cesnet.cz>
---
 drivers/net/szedata2/rte_eth_szedata2.c | 8 ++++----
 1 file changed, 4 insertions(+), 4 deletions(-)

diff --git a/drivers/net/szedata2/rte_eth_szedata2.c b/drivers/net/szedata2/rte_eth_szedata2.c
index 8278780..04dc8bf 100644
--- a/drivers/net/szedata2/rte_eth_szedata2.c
+++ b/drivers/net/szedata2/rte_eth_szedata2.c
@@ -1488,9 +1488,9 @@ struct szedata2_tx_queue {
 	FILE *fd;
 	char pcislot_path[PATH_MAX];
 	uint32_t domain;
-	uint32_t bus;
-	uint32_t devid;
-	uint32_t function;
+	uint8_t bus;
+	uint8_t devid;
+	uint8_t function;
 
 	dir = opendir("/sys/class/combo");
 	if (dir == NULL)
@@ -1515,7 +1515,7 @@ struct szedata2_tx_queue {
 		if (fd == NULL)
 			continue;
 
-		ret = fscanf(fd, "%4" PRIx16 ":%2" PRIx8 ":%2" PRIx8 ".%" PRIx8,
+		ret = fscanf(fd, "%8" SCNx32 ":%2" SCNx8 ":%2" SCNx8 ".%" SCNx8,
 				&domain, &bus, &devid, &function);
 		fclose(fd);
 		if (ret != 4)
-- 
1.8.3.1

Re: [dpdk-dev] [PATCH 2/4] net/szedata2: use dynamically allocated queues

[Ferruh Yigit]

On 4/4/2018 2:42 PM, Matej Vido wrote:
> Previously the queues were the part of private data structure of the
> Ethernet device.
> Now the queues are allocated at setup thus numa-aware allocation is
> possible.

Hi Matej,

Yes by default [rt]x_queues are allocated via rte_zmalloc, which uses SOCKET_ID_ANY.

And in burst functions, we do:
 nb_rx = (*dev->rx_pkt_burst)(dev->data->rx_queues[queue_id], rx_pkts, nb_pkts);

So there is an access to rx_queues in each rte_eth_rx_burst() call.

I wonder if you observe any performance difference with this update?
And what about moving to the ethdev layer instead of keeping local to the PMD?

> 
> Signed-off-by: Matej Vido <vido@cesnet.cz>

<...>

Re: [dpdk-dev] [PATCH 2/4] net/szedata2: use dynamically allocated queues

[Ferruh Yigit]

On 4/6/2018 2:20 PM, Ferruh Yigit wrote:
> On 4/4/2018 2:42 PM, Matej Vido wrote:
>> Previously the queues were the part of private data structure of the
>> Ethernet device.
>> Now the queues are allocated at setup thus numa-aware allocation is
>> possible.
> 
> Hi Matej,
> 
> Yes by default [rt]x_queues are allocated via rte_zmalloc, which uses SOCKET_ID_ANY.
> 
> And in burst functions, we do:
>  nb_rx = (*dev->rx_pkt_burst)(dev->data->rx_queues[queue_id], rx_pkts, nb_pkts);
> 
> So there is an access to rx_queues in each rte_eth_rx_burst() call.
> 
> I wonder if you observe any performance difference with this update?
> And what about moving to the ethdev layer instead of keeping local to the PMD?

Forget about it, I thought you are allocating data->[rt]x_queues in specific
socket, but this just allocating queues in specific socket, which is OK.

Still I would like to hear comments if allocating data->[rt]x_queues arrays in
specific socket helps for performance.

> 
>>
>> Signed-off-by: Matej Vido <vido@cesnet.cz>
> 
> <...>
> 

Re: [dpdk-dev] [PATCH 0/4] net/szedata2: fixes or improvements

[Ferruh Yigit]

On 4/4/2018 2:42 PM, Matej Vido wrote:
> Matej Vido (4):
>   net/szedata2: fix total stats
>   net/szedata2: use dynamically allocated queues
>   net/szedata2: add stat of mbuf allocation failures
>   net/szedata2: fix format string for pci address

Series applied to dpdk-next-net/master, thanks.