net/netvsc: allow setting rx and tx copy break

PATCH Superseded Long Li 3 participants 4 messages Oct 23, 2020 – Oct 31, 2020 Network drivers

lore

Share & tools

lore

Long Li

author

2020-10-23 21:54 UTC #0 lore raw

PATCH v2 1/2 net/netvsc: allow setting rx and tx copy break

From: Stephen Hemminger <[email protected]>

The values for Rx and Tx copy break should be tunable rather than hard coded constants.

The rx_copybreak sets the threshold where the driver uses an external mbuf to avoid having to copy data. Setting 0 for copybreak will cause driver to always create an external mbuf. Setting a value greater than the MTU would prevent it from ever making an external mbuf and always copy. The default value is 256 (bytes).

Likewise the tx_copybreak sets the threshold where the driver aggregates multiple small packets into one request. If tx_copybreak is 0 then each packet goes as a VMBus request (no copying). If tx_copybreak is set larger than the MTU, then all packets smaller than the chunk size of the VMBus send buffer will be copied; larger packets always have to go as a single direct request. The default value is 512 (bytes).

Signed-off-by Stephen Hemminger <[email protected]>

Signed-off-by Long Li <[email protected]>

---
 doc/guides/nics/netvsc.rst     | 17 ++++++++++++
 drivers/net/netvsc/hn_ethdev.c | 48 +++++++++++++++++++++++++---------
 drivers/net/netvsc/hn_rxtx.c   |  8 +++---
 drivers/net/netvsc/hn_var.h    |  5 ++++
 4 files changed, 60 insertions(+), 18 deletions(-)

doc/guides/nics/netvsc.rst +17 -0

index 6dbb9a5513..5a68ffa8a3 100644
@@ -116,3 +116,20 @@ The user can specify below argument in devargs.
     values save CPU cycles. This parameter is in microseconds.
     If the value is too large or too small it will be
     ignored by the host. (Default: 50)
+
+#.  ``rx_copybreak``:
+
+    The rx_copybreak sets the threshold where the driver uses an external
+    mbuf to avoid having to copy data. Setting 0 for copybreak will cause
+    driver to always create an external mbuf. Setting a value greater than
+    the MTU would prevent it from ever making an external mbuf and always
+    copy. The default value is 256 (bytes).
+
+#.  ``tx_copybreak``:
+
+    The tx_copybreak sets the threshold where the driver aggregates
+    multiple small packets into one request. If tx_copybreak is 0 then
+    each packet goes as a VMBus request (no copying). If tx_copybreak is
+    set larger than the MTU, then all packets smaller than the chunk size
+    of the VMBus send buffer will be copied; larger packets always have to
+    go as a single direct request. The default value is 512 (bytes).

drivers/net/netvsc/hn_ethdev.c +35 ~~-13~~

92 lines not on this page · raw message

drivers/net/netvsc/hn_rxtx.c +3 -5

index 5d59db513c..ce75b4d092 100644
@@ -40,9 +40,6 @@
 	(sizeof(struct vmbus_chanpkt_hdr) + sizeof(struct hn_nvs_rndis))
 
 #define HN_TXD_CACHE_SIZE	32 /* per cpu tx_descriptor pool cache */
-#define HN_TXCOPY_THRESHOLD	512
-
-#define HN_RXCOPY_THRESHOLD	256
 #define HN_RXQ_EVENT_DEFAULT	2048
 
 struct hn_rxinfo {

@@ -568,7 +565,7 @@ static void hn_rxpkt(struct hn_rx_queue *rxq, struct hn_rx_bufinfo *rxb,
 	 * For large packets, avoid copy if possible but need to keep
 	 * some space available in receive area for later packets.
 	 */
-	if (dlen >= HN_RXCOPY_THRESHOLD &&
+	if (dlen > hv->rx_copybreak &&
 	    (uint32_t)rte_atomic32_read(&rxq->rxbuf_outstanding) <
 			hv->rxbuf_section_cnt / 2) {
 		struct rte_mbuf_ext_shared_info *shinfo;

@@ -1516,7 +1513,8 @@ hn_xmit_pkts(void *ptxq, struct rte_mbuf **tx_pkts, uint16_t nb_pkts)
 			break;
 
 		/* For small packets aggregate them in chimney buffer */
-		if (m->pkt_len < HN_TXCOPY_THRESHOLD && pkt_size <= txq->agg_szmax) {
+		if (m->pkt_len <= hv->tx_copybreak &&
+		    pkt_size <= txq->agg_szmax) {
 			/* If this packet will not fit, then flush  */
 			if (txq->agg_pktleft == 0 ||
 			    RTE_ALIGN(pkt_size, txq->agg_align) > txq->agg_szleft) {

drivers/net/netvsc/hn_var.h +5 -0

index 56bfa972d1..c1bdafb413 100644
@@ -23,6 +23,9 @@
 /* Host monitor interval */
 #define HN_CHAN_LATENCY_NS	50000
 
+#define HN_TXCOPY_THRESHOLD	512
+#define HN_RXCOPY_THRESHOLD	256
+
 /* Buffers need to be aligned */
 #ifndef PAGE_SIZE
 #define PAGE_SIZE 4096

@@ -114,6 +117,7 @@ struct hn_data {
 
 	struct rte_mem_resource *rxbuf_res;	/* UIO resource for Rx */
 	uint32_t	rxbuf_section_cnt;	/* # of Rx sections */
+	uint32_t	rx_copybreak;
 	uint16_t	max_queues;		/* Max available queues */
 	uint16_t	num_queues;
 	uint64_t	rss_offloads;

@@ -122,6 +126,7 @@ struct hn_data {
 	struct rte_mem_resource *chim_res;	/* UIO resource for Tx */
 	struct rte_bitmap *chim_bmap;		/* Send buffer map */
 	void		*chim_bmem;
+	uint32_t	tx_copybreak;
 	uint32_t	chim_szmax;		/* Max size per buffer */
 	uint32_t	chim_cnt;		/* Max packets per buffer */

Signature · 1 line

2.25.1

Long Li

author

2020-10-23 21:54 UTC #1 lore raw

PATCH v2 2/2 net/netvsc: introduce driver parameter to control the use of external mbuf on receiving data

From: Long Li <[email protected]>

When receiving packets, netvsp puts data in a buffer mapped through UIO. Depending on packet size, netvsc may attach the buffer as an external mbuf. This is not a problem if this mbuf is consumed in the application, and the application can correctly read data out of an external mbuf.

However, there are two problems with data in an external mbuf.
1. Due to the limitation of the kernel UIO implementation, physical address of this external buffer is not exposed to the user-mode. If this mbuf is passed to another driver, the other driver is unable to map this buffer to iova.
2. Some DPDK applications are not aware of external mbuf, and may bug when they receive an mbuf with external buffer attached.

Introduce a driver parameter "rx_extmbuf_enable" to control if netvsc should use external mbuf for receiving packets. The default value is 0. (netvsc doesn't use external mbuf, it always allocates mbuf and copy data to mbuf) A non-zero value tells netvsc to attach external buffers to mbuf on receiving packets, thus avoid copying memory.

Signed-off-by Long Li <[email protected]>

---
 doc/guides/nics/netvsc.rst     | 8 ++++++++
 drivers/net/netvsc/hn_ethdev.c | 7 +++++++
 drivers/net/netvsc/hn_rxtx.c   | 2 +-
 drivers/net/netvsc/hn_var.h    | 3 +++
 4 files changed, 19 insertions(+), 1 deletion(-)

doc/guides/nics/netvsc.rst +8 -0

index 5a68ffa8a3..19f9940fe6 100644
@@ -133,3 +133,11 @@ The user can specify below argument in devargs.
     set larger than the MTU, then all packets smaller than the chunk size
     of the VMBus send buffer will be copied; larger packets always have to
     go as a single direct request. The default value is 512 (bytes).
+
+#.  ``rx_extmbuf_enable``:
+    The rx_extmbuf_enable is used to control if netvsc should use external
+    mbuf for receiving packets. The default value is 0. (netvsc doesn't use
+    external mbuf, it always allocates mbuf and copy received data to mbuf)
+    A non-zero value tells netvsc to attach external buffers to mbuf on
+    receiving packets, thus avoid copying memory. Use of external buffers
+    requires the application is able to read data from external mbuf.

drivers/net/netvsc/hn_ethdev.c +7 -0

index 2de951bae9..2d78a0118c 100644
@@ -48,6 +48,7 @@
 #define NETVSC_ARG_LATENCY "latency"
 #define NETVSC_ARG_RXBREAK "rx_copybreak"
 #define NETVSC_ARG_TXBREAK "tx_copybreak"
+#define NETVSC_ARG_RX_EXTMBUF_ENABLE "rx_extmbuf_enable"
 
 struct hn_xstats_name_off {
 	char name[RTE_ETH_XSTATS_NAME_SIZE];

@@ -164,6 +165,10 @@ static int hn_set_parameter(const char *key, const char *value, void *opaque)
 		hv->tx_copybreak = v;
 		PMD_DRV_LOG(DEBUG, "tx copy break set to %u",
 			    hv->tx_copybreak);
+	} else if (!strcmp(key, NETVSC_ARG_RX_EXTMBUF_ENABLE)) {
+		hv->rx_extmbuf_enable = v;
+		PMD_DRV_LOG(DEBUG, "rx extmbuf enable set to %u",
+			    hv->rx_extmbuf_enable);
 	}
 
 	return 0;

@@ -178,6 +183,7 @@ static int hn_parse_args(const struct rte_eth_dev *dev)
 		NETVSC_ARG_LATENCY,
 		NETVSC_ARG_RXBREAK,
 		NETVSC_ARG_TXBREAK,
+		NETVSC_ARG_RX_EXTMBUF_ENABLE,
 		NULL
 	};
 	struct rte_kvargs *kvlist;

@@ -987,6 +993,7 @@ eth_hn_dev_init(struct rte_eth_dev *eth_dev)
 	hv->latency = HN_CHAN_LATENCY_NS;
 	hv->rx_copybreak = HN_RXCOPY_THRESHOLD;
 	hv->tx_copybreak = HN_TXCOPY_THRESHOLD;
+	hv->rx_extmbuf_enable = HN_RX_EXTMBUF_ENABLE;
 	hv->max_queues = 1;
 
 	rte_rwlock_init(&hv->vf_lock);

drivers/net/netvsc/hn_rxtx.c +1 -1

index ce75b4d092..35c046b1a7 100644
@@ -565,7 +565,7 @@ static void hn_rxpkt(struct hn_rx_queue *rxq, struct hn_rx_bufinfo *rxb,
 	 * For large packets, avoid copy if possible but need to keep
 	 * some space available in receive area for later packets.
 	 */
-	if (dlen > hv->rx_copybreak &&
+	if (hv->rx_extmbuf_enable && dlen > hv->rx_copybreak &&
 	    (uint32_t)rte_atomic32_read(&rxq->rxbuf_outstanding) <
 			hv->rxbuf_section_cnt / 2) {
 		struct rte_mbuf_ext_shared_info *shinfo;

drivers/net/netvsc/hn_var.h +3 -0

index c1bdafb413..9dcb669f7a 100644
@@ -26,6 +26,8 @@
 #define HN_TXCOPY_THRESHOLD	512
 #define HN_RXCOPY_THRESHOLD	256
 
+#define HN_RX_EXTMBUF_ENABLE	0
+
 /* Buffers need to be aligned */
 #ifndef PAGE_SIZE
 #define PAGE_SIZE 4096

@@ -118,6 +120,7 @@ struct hn_data {
 	struct rte_mem_resource *rxbuf_res;	/* UIO resource for Rx */
 	uint32_t	rxbuf_section_cnt;	/* # of Rx sections */
 	uint32_t	rx_copybreak;
+	uint32_t	rx_extmbuf_enable;
 	uint16_t	max_queues;		/* Max available queues */
 	uint16_t	num_queues;
 	uint64_t	rss_offloads;

Signature · 1 line

2.25.1

Ferruh Yigit

2020-10-29 20:00 UTC #2 lore raw

In reply to Long Li

Quoted 21 lines · Long Li

From: Stephen Hemminger <[email protected]>
The values for Rx and Tx copy break should be tunable rather than hard coded constants.
The rx_copybreak sets the threshold where the driver uses an external mbuf to avoid having to copy data. Setting 0 for copybreak will cause driver to always create an external mbuf. Setting a value greater than the MTU would prevent it from ever making an external mbuf and always copy. The default value is 256 (bytes).
Likewise the tx_copybreak sets the threshold where the driver aggregates multiple small packets into one request. If tx_copybreak is 0 then each packet goes as a VMBus request (no copying). If tx_copybreak is set larger than the MTU, then all packets smaller than the chunk size of the VMBus send buffer will be copied; larger packets always have to go as a single direct request. The default value is 512 (bytes).
Signed-off-by Stephen Hemminger <[email protected]>
Signed-off-by Long Li <[email protected]>

<...>

drivers/net/netvsc/hn_ethdev.c static int hn_parse_args(const struct rte_eth_dev *dev) · line 202

Quoted 12 lines

drivers/net/netvsc/hn_ethdev.c +5 -3

@@ -181,9 +195,14 @@ static int hn_parse_args(const struct rte_eth_dev *dev)
  		return -EINVAL;
  	}
 
-	ret = rte_kvargs_process(kvlist, "latency", hn_set_latency, hv);
-	if (ret)
-		PMD_DRV_LOG(ERR, "Unable to process latency arg\n");
+	for (i = 0; i < sizeof(valid_keys) / sizeof(valid_keys[0]) - 1; i++) {
+		ret = rte_kvargs_process(kvlist, valid_keys[i],
+					 hn_set_parameter, hv);
+		if (ret) {
+			PMD_DRV_LOG(ERR, "Unable to process latency arg\n");

Need to update the log, it is not only 'latency' anymore.

drivers/net/netvsc/hn_ethdev.c static int hn_parse_args(const struct rte_eth_dev *dev) · line 205

Quoted 3 lines

drivers/net/netvsc/hn_ethdev.c +3 -0
+			break;
+		}
+	}

Since there is single callback for all args, and there is already a 'key' check in the callback function, why not call the 'rte_kvargs_process()' with NULL argument and drop the for loop:
ret = rte_kvargs_process(kvlist, NULL, hn_set_parameter, hv);

Can you also register the devargs in .c file:
'RTE_PMD_REGISTER_PARAM_STRING'
This is to get PMD supported devargs from .so, using "./usertools/dpdk-pmdinfo.py".

Thanks,
ferruh

Long Li

2020-10-31 05:59 UTC #3 lore raw

Quoted 54 lines

Subject: Re: [dpdk-dev] [PATCH V2 1/2] net/netvsc: allow setting rx and tx copy break
Quoted 21 lines · Long Li
From: Stephen Hemminger <[email protected]>
The values for Rx and Tx copy break should be tunable rather than hard coded constants.
The rx_copybreak sets the threshold where the driver uses an external mbuf to avoid having to copy data. Setting 0 for copybreak will cause driver to always create an external mbuf. Setting a value greater than the MTU would prevent it from ever making an external mbuf and always copy. The default value is 256 (bytes).
Likewise the tx_copybreak sets the threshold where the driver aggregates multiple small packets into one request. If tx_copybreak is 0 then each packet goes as a VMBus request (no copying).
If tx_copybreak is set larger than the MTU, then all packets smaller than the chunk size of the VMBus send buffer will be copied; larger packets always have to go as a single direct request. The default value is 512 (bytes).
Signed-off-by Stephen Hemminger <[email protected]>
Signed-off-by Long Li <[email protected]>
<...>
Review static int hn_parse_args(const struct rte_eth_dev
Quoted 1 line
(quoted diff) +0 -0
@@ -181,9 +195,14 @@ static int hn_parse_args(const struct rte_eth_dev
*dev)
drivers/net/netvsc/hn_ethdev.c static int hn_parse_args(const struct rte_eth_dev *dev)
Quoted 11 lines
drivers/net/netvsc/hn_ethdev.c +5 -3
  		return -EINVAL;
  	}
 
-	ret = rte_kvargs_process(kvlist, "latency", hn_set_latency, hv);
-	if (ret)
-		PMD_DRV_LOG(ERR, "Unable to process latency arg\n");
+	for (i = 0; i < sizeof(valid_keys) / sizeof(valid_keys[0]) - 1; i++) {
+		ret = rte_kvargs_process(kvlist, valid_keys[i],
+					 hn_set_parameter, hv);
+		if (ret) {
+			PMD_DRV_LOG(ERR, "Unable to process latency
arg\n");
Need to update the log, it is not only 'latency' anymore.
drivers/net/netvsc/hn_ethdev.c static int hn_parse_args(const struct rte_eth_dev *dev) · line 205
Quoted 3 lines
drivers/net/netvsc/hn_ethdev.c +3 -0
+			break;
+		}
+	}
Since there is single callback for all args, and there is already a 'key' check in the callback function, why not call the 'rte_kvargs_process()' with NULL argument and drop the for loop:
ret = rte_kvargs_process(kvlist, NULL, hn_set_parameter, hv);

Quoted 4 lines

Can you also register the devargs in .c file:
'RTE_PMD_REGISTER_PARAM_STRING'
This is to get PMD supported devargs from .so, using "./usertools/dpdk- pmdinfo.py".

I have sent V3 to address comments.

Thanks!

Long

Quoted 3 lines

Thanks,
ferruh