* Re: [dpdk-dev] [PATCH v2] net/vhost: merge vhost stats loop in vhost Tx/Rx
2021-09-28 1:43 ` Gaoxiang Liu
@ 2021-10-06 10:42 ` Gaoxiang Liu
2021-10-15 12:16 ` Maxime Coquelin
2021-10-17 23:19 ` [dpdk-dev] [PATCH v3] " Gaoxiang Liu
2 siblings, 0 replies; 11+ messages in thread
From: Gaoxiang Liu @ 2021-10-06 10:42 UTC (permalink / raw)
To: maxime.coquelin, chenbo.xia; +Cc: dev, liugaoxiang, Gaoxiang Liu
Hi, Maxime, Chenbo
Please help to review the patch.
Thanks.
Gaoxiang
On 09/28/2021 09:43, Gaoxiang Liu wrote:
To improve performance in vhost Tx/Rx, merge vhost stats loop.
eth_vhost_tx has 2 loop of send num iteraion.
It can be merge into one.
eth_vhost_rx has the same issue as Tx.
Fixes: 4d6cf2ac93dc ("net/vhost: add extended statistics")
Signed-off-by: Gaoxiang Liu <gaoxiangliu0@163.com>
---
v2:
* Fix coding style issues.
---
drivers/net/vhost/rte_eth_vhost.c | 62 ++++++++++++++-----------------
1 file changed, 28 insertions(+), 34 deletions(-)
diff --git a/drivers/net/vhost/rte_eth_vhost.c b/drivers/net/vhost/rte_eth_vhost.c
index a202931e9a..a4129980f2 100644
--- a/drivers/net/vhost/rte_eth_vhost.c
+++ b/drivers/net/vhost/rte_eth_vhost.c
@@ -336,38 +336,29 @@ vhost_count_xcast_packets(struct vhost_queue *vq,
}
static void
-vhost_update_packet_xstats(struct vhost_queue *vq, struct rte_mbuf **bufs,
- uint16_t count, uint64_t nb_bytes,
- uint64_t nb_missed)
+vhost_update_single_packet_xstats(struct vhost_queue *vq, struct rte_mbuf *buf)
{
uint32_t pkt_len = 0;
- uint64_t i = 0;
uint64_t index;
struct vhost_stats *pstats = &vq->stats;
- pstats->xstats[VHOST_BYTE] += nb_bytes;
- pstats->xstats[VHOST_MISSED_PKT] += nb_missed;
- pstats->xstats[VHOST_UNICAST_PKT] += nb_missed;
-
- for (i = 0; i < count ; i++) {
- pstats->xstats[VHOST_PKT]++;
- pkt_len = bufs[i]->pkt_len;
- if (pkt_len == 64) {
- pstats->xstats[VHOST_64_PKT]++;
- } else if (pkt_len > 64 && pkt_len < 1024) {
- index = (sizeof(pkt_len) * 8)
- - __builtin_clz(pkt_len) - 5;
- pstats->xstats[index]++;
- } else {
- if (pkt_len < 64)
- pstats->xstats[VHOST_UNDERSIZE_PKT]++;
- else if (pkt_len <= 1522)
- pstats->xstats[VHOST_1024_TO_1522_PKT]++;
- else if (pkt_len > 1522)
- pstats->xstats[VHOST_1523_TO_MAX_PKT]++;
- }
- vhost_count_xcast_packets(vq, bufs[i]);
+ pstats->xstats[VHOST_PKT]++;
+ pkt_len = buf->pkt_len;
+ if (pkt_len == 64) {
+ pstats->xstats[VHOST_64_PKT]++;
+ } else if (pkt_len > 64 && pkt_len < 1024) {
+ index = (sizeof(pkt_len) * 8)
+ - __builtin_clz(pkt_len) - 5;
+ pstats->xstats[index]++;
+ } else {
+ if (pkt_len < 64)
+ pstats->xstats[VHOST_UNDERSIZE_PKT]++;
+ else if (pkt_len <= 1522)
+ pstats->xstats[VHOST_1024_TO_1522_PKT]++;
+ else if (pkt_len > 1522)
+ pstats->xstats[VHOST_1523_TO_MAX_PKT]++;
}
+ vhost_count_xcast_packets(vq, buf);
}
static uint16_t
@@ -376,7 +367,6 @@ eth_vhost_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
struct vhost_queue *r = q;
uint16_t i, nb_rx = 0;
uint16_t nb_receive = nb_bufs;
- uint64_t nb_bytes = 0;
if (unlikely(rte_atomic32_read(&r->allow_queuing) == 0))
return 0;
@@ -411,11 +401,11 @@ eth_vhost_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
if (r->internal->vlan_strip)
rte_vlan_strip(bufs[i]);
- nb_bytes += bufs[i]->pkt_len;
- }
+ r->stats.bytes += bufs[i]->pkt_len;
+ r->stats.xstats[VHOST_BYTE] += bufs[i]->pkt_len;
- r->stats.bytes += nb_bytes;
- vhost_update_packet_xstats(r, bufs, nb_rx, nb_bytes, 0);
+ vhost_update_single_packet_xstats(r, bufs[i]);
+ }
out:
rte_atomic32_set(&r->while_queuing, 0);
@@ -471,16 +461,20 @@ eth_vhost_tx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
break;
}
- for (i = 0; likely(i < nb_tx); i++)
+ for (i = 0; likely(i < nb_tx); i++) {
nb_bytes += bufs[i]->pkt_len;
+ vhost_update_single_packet_xstats(r, bufs[i]);
+ }
nb_missed = nb_bufs - nb_tx;
r->stats.pkts += nb_tx;
r->stats.bytes += nb_bytes;
- r->stats.missed_pkts += nb_bufs - nb_tx;
+ r->stats.missed_pkts += nb_missed;
- vhost_update_packet_xstats(r, bufs, nb_tx, nb_bytes, nb_missed);
+ r->stats.xstats[VHOST_BYTE] += nb_bytes;
+ r->stats.xstats[VHOST_MISSED_PKT] += nb_missed;
+ r->stats.xstats[VHOST_UNICAST_PKT] += nb_missed;
/* According to RFC2863, ifHCOutUcastPkts, ifHCOutMulticastPkts and
* ifHCOutBroadcastPkts counters are increased when packets are not
--
2.32.0
^ permalink raw reply [flat|nested] 11+ messages in thread* Re: [dpdk-dev] [PATCH v2] net/vhost: merge vhost stats loop in vhost Tx/Rx
2021-09-28 1:43 ` Gaoxiang Liu
2021-10-06 10:42 ` Gaoxiang Liu
@ 2021-10-15 12:16 ` Maxime Coquelin
2021-10-16 8:59 ` Gaoxiang Liu
2021-10-17 23:19 ` [dpdk-dev] [PATCH v3] " Gaoxiang Liu
2 siblings, 1 reply; 11+ messages in thread
From: Maxime Coquelin @ 2021-10-15 12:16 UTC (permalink / raw)
To: Gaoxiang Liu, chenbo.xia; +Cc: dev, liugaoxiang
Hi,
On 9/28/21 03:43, Gaoxiang Liu wrote:
> To improve performance in vhost Tx/Rx, merge vhost stats loop.
> eth_vhost_tx has 2 loop of send num iteraion.
> It can be merge into one.
> eth_vhost_rx has the same issue as Tx.
>
> Fixes: 4d6cf2ac93dc ("net/vhost: add extended statistics")
Please remove the Fixes tag, this is an optimization, not a fix.
>
> Signed-off-by: Gaoxiang Liu <gaoxiangliu0@163.com>
> ---
>
> v2:
> * Fix coding style issues.
> ---
> drivers/net/vhost/rte_eth_vhost.c | 62 ++++++++++++++-----------------
> 1 file changed, 28 insertions(+), 34 deletions(-)
>
> diff --git a/drivers/net/vhost/rte_eth_vhost.c b/drivers/net/vhost/rte_eth_vhost.c
> index a202931e9a..a4129980f2 100644
> --- a/drivers/net/vhost/rte_eth_vhost.c
> +++ b/drivers/net/vhost/rte_eth_vhost.c
> @@ -336,38 +336,29 @@ vhost_count_xcast_packets(struct vhost_queue *vq,
> }
>
> static void
> -vhost_update_packet_xstats(struct vhost_queue *vq, struct rte_mbuf **bufs,
> - uint16_t count, uint64_t nb_bytes,
> - uint64_t nb_missed)
> +vhost_update_single_packet_xstats(struct vhost_queue *vq, struct rte_mbuf *buf)
I tried to build without and with your patch, and I think that what can
explain most of the performance difference is that without your patch
the function is not inlined, whereas it is implicitely inlined with your
patch applied.
I agree with your patch, but I think we might add __rte_always_inline to
this function to make it explicit. What do you think?
Other than that:
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Thanks,
Maxime
> {
> uint32_t pkt_len = 0;
> - uint64_t i = 0;
> uint64_t index;
> struct vhost_stats *pstats = &vq->stats;
>
> - pstats->xstats[VHOST_BYTE] += nb_bytes;
> - pstats->xstats[VHOST_MISSED_PKT] += nb_missed;
> - pstats->xstats[VHOST_UNICAST_PKT] += nb_missed;
> -
> - for (i = 0; i < count ; i++) {
> - pstats->xstats[VHOST_PKT]++;
> - pkt_len = bufs[i]->pkt_len;
> - if (pkt_len == 64) {
> - pstats->xstats[VHOST_64_PKT]++;
> - } else if (pkt_len > 64 && pkt_len < 1024) {
> - index = (sizeof(pkt_len) * 8)
> - - __builtin_clz(pkt_len) - 5;
> - pstats->xstats[index]++;
> - } else {
> - if (pkt_len < 64)
> - pstats->xstats[VHOST_UNDERSIZE_PKT]++;
> - else if (pkt_len <= 1522)
> - pstats->xstats[VHOST_1024_TO_1522_PKT]++;
> - else if (pkt_len > 1522)
> - pstats->xstats[VHOST_1523_TO_MAX_PKT]++;
> - }
> - vhost_count_xcast_packets(vq, bufs[i]);
> + pstats->xstats[VHOST_PKT]++;
> + pkt_len = buf->pkt_len;
> + if (pkt_len == 64) {
> + pstats->xstats[VHOST_64_PKT]++;
> + } else if (pkt_len > 64 && pkt_len < 1024) {
> + index = (sizeof(pkt_len) * 8)
> + - __builtin_clz(pkt_len) - 5;
> + pstats->xstats[index]++;
> + } else {
> + if (pkt_len < 64)
> + pstats->xstats[VHOST_UNDERSIZE_PKT]++;
> + else if (pkt_len <= 1522)
> + pstats->xstats[VHOST_1024_TO_1522_PKT]++;
> + else if (pkt_len > 1522)
> + pstats->xstats[VHOST_1523_TO_MAX_PKT]++;
> }
> + vhost_count_xcast_packets(vq, buf);
> }
>
> static uint16_t
> @@ -376,7 +367,6 @@ eth_vhost_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
> struct vhost_queue *r = q;
> uint16_t i, nb_rx = 0;
> uint16_t nb_receive = nb_bufs;
> - uint64_t nb_bytes = 0;
>
> if (unlikely(rte_atomic32_read(&r->allow_queuing) == 0))
> return 0;
> @@ -411,11 +401,11 @@ eth_vhost_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
> if (r->internal->vlan_strip)
> rte_vlan_strip(bufs[i]);
>
> - nb_bytes += bufs[i]->pkt_len;
> - }
> + r->stats.bytes += bufs[i]->pkt_len;
> + r->stats.xstats[VHOST_BYTE] += bufs[i]->pkt_len;
>
> - r->stats.bytes += nb_bytes;
> - vhost_update_packet_xstats(r, bufs, nb_rx, nb_bytes, 0);
> + vhost_update_single_packet_xstats(r, bufs[i]);
> + }
>
> out:
> rte_atomic32_set(&r->while_queuing, 0);
> @@ -471,16 +461,20 @@ eth_vhost_tx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
> break;
> }
>
> - for (i = 0; likely(i < nb_tx); i++)
> + for (i = 0; likely(i < nb_tx); i++) {
> nb_bytes += bufs[i]->pkt_len;
> + vhost_update_single_packet_xstats(r, bufs[i]);
> + }
>
> nb_missed = nb_bufs - nb_tx;
>
> r->stats.pkts += nb_tx;
> r->stats.bytes += nb_bytes;
> - r->stats.missed_pkts += nb_bufs - nb_tx;
> + r->stats.missed_pkts += nb_missed;
>
> - vhost_update_packet_xstats(r, bufs, nb_tx, nb_bytes, nb_missed);
> + r->stats.xstats[VHOST_BYTE] += nb_bytes;
> + r->stats.xstats[VHOST_MISSED_PKT] += nb_missed;
> + r->stats.xstats[VHOST_UNICAST_PKT] += nb_missed;
>
> /* According to RFC2863, ifHCOutUcastPkts, ifHCOutMulticastPkts and
> * ifHCOutBroadcastPkts counters are increased when packets are not
>
^ permalink raw reply [flat|nested] 11+ messages in thread* Re: [dpdk-dev] [PATCH v2] net/vhost: merge vhost stats loop in vhost Tx/Rx
2021-10-15 12:16 ` Maxime Coquelin
@ 2021-10-16 8:59 ` Gaoxiang Liu
0 siblings, 0 replies; 11+ messages in thread
From: Gaoxiang Liu @ 2021-10-16 8:59 UTC (permalink / raw)
To: maxime.coquelin; +Cc: Xia, Chenbo, dev, liugaoxiang
Hi, Maxime
I agree with you.The inline should be added to vhost_update_single_packet_xstats function.
I will fix it in [PATCH v3].
Thanks,
Gaoxiang
发自 网易邮箱大师
---- 回复的原邮件 ----
| 发件人 | Maxime Coquelin<maxime.coquelin@redhat.com> |
| 日期 | 2021年10月15日 20:16 |
| 收件人 | Gaoxiang Liu<gaoxiangliu0@163.com>、chenbo.xia@intel.com<chenbo.xia@intel.com> |
| 抄送至 | dev@dpdk.org<dev@dpdk.org>、liugaoxiang@huawei.com<liugaoxiang@huawei.com> |
| 主题 | Re: [PATCH v2] net/vhost: merge vhost stats loop in vhost Tx/Rx |
Hi,
On 9/28/21 03:43, Gaoxiang Liu wrote:
> To improve performance in vhost Tx/Rx, merge vhost stats loop.
> eth_vhost_tx has 2 loop of send num iteraion.
> It can be merge into one.
> eth_vhost_rx has the same issue as Tx.
>
> Fixes: 4d6cf2ac93dc ("net/vhost: add extended statistics")
Please remove the Fixes tag, this is an optimization, not a fix.
>
> Signed-off-by: Gaoxiang Liu <gaoxiangliu0@163.com>
> ---
>
> v2:
> * Fix coding style issues.
> ---
> drivers/net/vhost/rte_eth_vhost.c | 62 ++++++++++++++-----------------
> 1 file changed, 28 insertions(+), 34 deletions(-)
>
> diff --git a/drivers/net/vhost/rte_eth_vhost.c b/drivers/net/vhost/rte_eth_vhost.c
> index a202931e9a..a4129980f2 100644
> --- a/drivers/net/vhost/rte_eth_vhost.c
> +++ b/drivers/net/vhost/rte_eth_vhost.c
> @@ -336,38 +336,29 @@ vhost_count_xcast_packets(struct vhost_queue *vq,
> }
>
> static void
> -vhost_update_packet_xstats(struct vhost_queue *vq, struct rte_mbuf **bufs,
> - uint16_t count, uint64_t nb_bytes,
> - uint64_t nb_missed)
> +vhost_update_single_packet_xstats(struct vhost_queue *vq, struct rte_mbuf *buf)
I tried to build without and with your patch, and I think that what can
explain most of the performance difference is that without your patch
the function is not inlined, whereas it is implicitely inlined with your
patch applied.
I agree with your patch, but I think we might add __rte_always_inline to
this function to make it explicit. What do you think?
Other than that:
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Thanks,
Maxime
> {
> uint32_t pkt_len = 0;
> - uint64_t i = 0;
> uint64_t index;
> struct vhost_stats *pstats = &vq->stats;
>
> - pstats->xstats[VHOST_BYTE] += nb_bytes;
> - pstats->xstats[VHOST_MISSED_PKT] += nb_missed;
> - pstats->xstats[VHOST_UNICAST_PKT] += nb_missed;
> -
> - for (i = 0; i < count ; i++) {
> - pstats->xstats[VHOST_PKT]++;
> - pkt_len = bufs[i]->pkt_len;
> - if (pkt_len == 64) {
> - pstats->xstats[VHOST_64_PKT]++;
> - } else if (pkt_len > 64 && pkt_len < 1024) {
> - index = (sizeof(pkt_len) * 8)
> - - __builtin_clz(pkt_len) - 5;
> - pstats->xstats[index]++;
> - } else {
> - if (pkt_len < 64)
> - pstats->xstats[VHOST_UNDERSIZE_PKT]++;
> - else if (pkt_len <= 1522)
> - pstats->xstats[VHOST_1024_TO_1522_PKT]++;
> - else if (pkt_len > 1522)
> - pstats->xstats[VHOST_1523_TO_MAX_PKT]++;
> - }
> - vhost_count_xcast_packets(vq, bufs[i]);
> + pstats->xstats[VHOST_PKT]++;
> + pkt_len = buf->pkt_len;
> + if (pkt_len == 64) {
> + pstats->xstats[VHOST_64_PKT]++;
> + } else if (pkt_len > 64 && pkt_len < 1024) {
> + index = (sizeof(pkt_len) * 8)
> + - __builtin_clz(pkt_len) - 5;
> + pstats->xstats[index]++;
> + } else {
> + if (pkt_len < 64)
> + pstats->xstats[VHOST_UNDERSIZE_PKT]++;
> + else if (pkt_len <= 1522)
> + pstats->xstats[VHOST_1024_TO_1522_PKT]++;
> + else if (pkt_len > 1522)
> + pstats->xstats[VHOST_1523_TO_MAX_PKT]++;
> }
> + vhost_count_xcast_packets(vq, buf);
> }
>
> static uint16_t
> @@ -376,7 +367,6 @@ eth_vhost_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
> struct vhost_queue *r = q;
> uint16_t i, nb_rx = 0;
> uint16_t nb_receive = nb_bufs;
> - uint64_t nb_bytes = 0;
>
> if (unlikely(rte_atomic32_read(&r->allow_queuing) == 0))
> return 0;
> @@ -411,11 +401,11 @@ eth_vhost_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
> if (r->internal->vlan_strip)
> rte_vlan_strip(bufs[i]);
>
> - nb_bytes += bufs[i]->pkt_len;
> - }
> + r->stats.bytes += bufs[i]->pkt_len;
> + r->stats.xstats[VHOST_BYTE] += bufs[i]->pkt_len;
>
> - r->stats.bytes += nb_bytes;
> - vhost_update_packet_xstats(r, bufs, nb_rx, nb_bytes, 0);
> + vhost_update_single_packet_xstats(r, bufs[i]);
> + }
>
> out:
> rte_atomic32_set(&r->while_queuing, 0);
> @@ -471,16 +461,20 @@ eth_vhost_tx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
> break;
> }
>
> - for (i = 0; likely(i < nb_tx); i++)
> + for (i = 0; likely(i < nb_tx); i++) {
> nb_bytes += bufs[i]->pkt_len;
> + vhost_update_single_packet_xstats(r, bufs[i]);
> + }
>
> nb_missed = nb_bufs - nb_tx;
>
> r->stats.pkts += nb_tx;
> r->stats.bytes += nb_bytes;
> - r->stats.missed_pkts += nb_bufs - nb_tx;
> + r->stats.missed_pkts += nb_missed;
>
> - vhost_update_packet_xstats(r, bufs, nb_tx, nb_bytes, nb_missed);
> + r->stats.xstats[VHOST_BYTE] += nb_bytes;
> + r->stats.xstats[VHOST_MISSED_PKT] += nb_missed;
> + r->stats.xstats[VHOST_UNICAST_PKT] += nb_missed;
>
> /* According to RFC2863, ifHCOutUcastPkts, ifHCOutMulticastPkts and
> * ifHCOutBroadcastPkts counters are increased when packets are not
>
^ permalink raw reply [flat|nested] 11+ messages in thread
* [dpdk-dev] [PATCH v3] net/vhost: merge vhost stats loop in vhost Tx/Rx
2021-09-28 1:43 ` Gaoxiang Liu
2021-10-06 10:42 ` Gaoxiang Liu
2021-10-15 12:16 ` Maxime Coquelin
@ 2021-10-17 23:19 ` Gaoxiang Liu
2021-10-21 10:04 ` Maxime Coquelin
` (2 more replies)
2 siblings, 3 replies; 11+ messages in thread
From: Gaoxiang Liu @ 2021-10-17 23:19 UTC (permalink / raw)
To: maxime.coquelin, chenbo.xia; +Cc: dev, liugaoxiang, Gaoxiang Liu
To improve performance in vhost Tx/Rx, merge vhost stats loop.
eth_vhost_tx has 2 loop of send num iteraion.
It can be merge into one.
eth_vhost_rx has the same issue as Tx.
Signed-off-by: Gaoxiang Liu <gaoxiangliu0@163.com>
---
v2:
* Fix coding style issues.
v3:
* add __rte_always_inline to vhost_update_single_packet_xstats.
---
drivers/net/vhost/rte_eth_vhost.c | 64 ++++++++++++++-----------------
1 file changed, 29 insertions(+), 35 deletions(-)
diff --git a/drivers/net/vhost/rte_eth_vhost.c b/drivers/net/vhost/rte_eth_vhost.c
index a202931e9a..021195ae57 100644
--- a/drivers/net/vhost/rte_eth_vhost.c
+++ b/drivers/net/vhost/rte_eth_vhost.c
@@ -335,39 +335,30 @@ vhost_count_xcast_packets(struct vhost_queue *vq,
}
}
-static void
-vhost_update_packet_xstats(struct vhost_queue *vq, struct rte_mbuf **bufs,
- uint16_t count, uint64_t nb_bytes,
- uint64_t nb_missed)
+static __rte_always_inline void
+vhost_update_single_packet_xstats(struct vhost_queue *vq, struct rte_mbuf *buf)
{
uint32_t pkt_len = 0;
- uint64_t i = 0;
uint64_t index;
struct vhost_stats *pstats = &vq->stats;
- pstats->xstats[VHOST_BYTE] += nb_bytes;
- pstats->xstats[VHOST_MISSED_PKT] += nb_missed;
- pstats->xstats[VHOST_UNICAST_PKT] += nb_missed;
-
- for (i = 0; i < count ; i++) {
- pstats->xstats[VHOST_PKT]++;
- pkt_len = bufs[i]->pkt_len;
- if (pkt_len == 64) {
- pstats->xstats[VHOST_64_PKT]++;
- } else if (pkt_len > 64 && pkt_len < 1024) {
- index = (sizeof(pkt_len) * 8)
- - __builtin_clz(pkt_len) - 5;
- pstats->xstats[index]++;
- } else {
- if (pkt_len < 64)
- pstats->xstats[VHOST_UNDERSIZE_PKT]++;
- else if (pkt_len <= 1522)
- pstats->xstats[VHOST_1024_TO_1522_PKT]++;
- else if (pkt_len > 1522)
- pstats->xstats[VHOST_1523_TO_MAX_PKT]++;
- }
- vhost_count_xcast_packets(vq, bufs[i]);
+ pstats->xstats[VHOST_PKT]++;
+ pkt_len = buf->pkt_len;
+ if (pkt_len == 64) {
+ pstats->xstats[VHOST_64_PKT]++;
+ } else if (pkt_len > 64 && pkt_len < 1024) {
+ index = (sizeof(pkt_len) * 8)
+ - __builtin_clz(pkt_len) - 5;
+ pstats->xstats[index]++;
+ } else {
+ if (pkt_len < 64)
+ pstats->xstats[VHOST_UNDERSIZE_PKT]++;
+ else if (pkt_len <= 1522)
+ pstats->xstats[VHOST_1024_TO_1522_PKT]++;
+ else if (pkt_len > 1522)
+ pstats->xstats[VHOST_1523_TO_MAX_PKT]++;
}
+ vhost_count_xcast_packets(vq, buf);
}
static uint16_t
@@ -376,7 +367,6 @@ eth_vhost_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
struct vhost_queue *r = q;
uint16_t i, nb_rx = 0;
uint16_t nb_receive = nb_bufs;
- uint64_t nb_bytes = 0;
if (unlikely(rte_atomic32_read(&r->allow_queuing) == 0))
return 0;
@@ -411,11 +401,11 @@ eth_vhost_rx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
if (r->internal->vlan_strip)
rte_vlan_strip(bufs[i]);
- nb_bytes += bufs[i]->pkt_len;
- }
+ r->stats.bytes += bufs[i]->pkt_len;
+ r->stats.xstats[VHOST_BYTE] += bufs[i]->pkt_len;
- r->stats.bytes += nb_bytes;
- vhost_update_packet_xstats(r, bufs, nb_rx, nb_bytes, 0);
+ vhost_update_single_packet_xstats(r, bufs[i]);
+ }
out:
rte_atomic32_set(&r->while_queuing, 0);
@@ -471,16 +461,20 @@ eth_vhost_tx(void *q, struct rte_mbuf **bufs, uint16_t nb_bufs)
break;
}
- for (i = 0; likely(i < nb_tx); i++)
+ for (i = 0; likely(i < nb_tx); i++) {
nb_bytes += bufs[i]->pkt_len;
+ vhost_update_single_packet_xstats(r, bufs[i]);
+ }
nb_missed = nb_bufs - nb_tx;
r->stats.pkts += nb_tx;
r->stats.bytes += nb_bytes;
- r->stats.missed_pkts += nb_bufs - nb_tx;
+ r->stats.missed_pkts += nb_missed;
- vhost_update_packet_xstats(r, bufs, nb_tx, nb_bytes, nb_missed);
+ r->stats.xstats[VHOST_BYTE] += nb_bytes;
+ r->stats.xstats[VHOST_MISSED_PKT] += nb_missed;
+ r->stats.xstats[VHOST_UNICAST_PKT] += nb_missed;
/* According to RFC2863, ifHCOutUcastPkts, ifHCOutMulticastPkts and
* ifHCOutBroadcastPkts counters are increased when packets are not
--
2.32.0
^ permalink raw reply [flat|nested] 11+ messages in thread* Re: [dpdk-dev] [PATCH v3] net/vhost: merge vhost stats loop in vhost Tx/Rx
2021-10-17 23:19 ` [dpdk-dev] [PATCH v3] " Gaoxiang Liu
@ 2021-10-21 10:04 ` Maxime Coquelin
2021-10-21 12:34 ` Maxime Coquelin
2021-10-21 18:56 ` Ferruh Yigit
2 siblings, 0 replies; 11+ messages in thread
From: Maxime Coquelin @ 2021-10-21 10:04 UTC (permalink / raw)
To: Gaoxiang Liu, chenbo.xia; +Cc: dev, liugaoxiang
On 10/18/21 01:19, Gaoxiang Liu wrote:
> To improve performance in vhost Tx/Rx, merge vhost stats loop.
> eth_vhost_tx has 2 loop of send num iteraion.
> It can be merge into one.
> eth_vhost_rx has the same issue as Tx.
>
> Signed-off-by: Gaoxiang Liu <gaoxiangliu0@163.com>
> ---
>
> v2:
> * Fix coding style issues.
>
> v3:
> * add __rte_always_inline to vhost_update_single_packet_xstats.
> ---
> drivers/net/vhost/rte_eth_vhost.c | 64 ++++++++++++++-----------------
> 1 file changed, 29 insertions(+), 35 deletions(-)
>
Reviewed-by: Maxime Coquelin <maxime.coquelin@redhat.com>
Thanks!
Maxime
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [dpdk-dev] [PATCH v3] net/vhost: merge vhost stats loop in vhost Tx/Rx
2021-10-17 23:19 ` [dpdk-dev] [PATCH v3] " Gaoxiang Liu
2021-10-21 10:04 ` Maxime Coquelin
@ 2021-10-21 12:34 ` Maxime Coquelin
2021-10-21 18:56 ` Ferruh Yigit
2 siblings, 0 replies; 11+ messages in thread
From: Maxime Coquelin @ 2021-10-21 12:34 UTC (permalink / raw)
To: Gaoxiang Liu, chenbo.xia; +Cc: dev, liugaoxiang
On 10/18/21 01:19, Gaoxiang Liu wrote:
> To improve performance in vhost Tx/Rx, merge vhost stats loop.
> eth_vhost_tx has 2 loop of send num iteraion.
> It can be merge into one.
> eth_vhost_rx has the same issue as Tx.
>
> Signed-off-by: Gaoxiang Liu <gaoxiangliu0@163.com>
> ---
>
> v2:
> * Fix coding style issues.
>
> v3:
> * add __rte_always_inline to vhost_update_single_packet_xstats.
> ---
> drivers/net/vhost/rte_eth_vhost.c | 64 ++++++++++++++-----------------
> 1 file changed, 29 insertions(+), 35 deletions(-)
>
Applied to dpdk-next-virtio/main.
Thanks,
Maxime
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [dpdk-dev] [PATCH v3] net/vhost: merge vhost stats loop in vhost Tx/Rx
2021-10-17 23:19 ` [dpdk-dev] [PATCH v3] " Gaoxiang Liu
2021-10-21 10:04 ` Maxime Coquelin
2021-10-21 12:34 ` Maxime Coquelin
@ 2021-10-21 18:56 ` Ferruh Yigit
2021-10-22 1:35 ` Gaoxiang Liu
2 siblings, 1 reply; 11+ messages in thread
From: Ferruh Yigit @ 2021-10-21 18:56 UTC (permalink / raw)
To: Gaoxiang Liu, maxime.coquelin, chenbo.xia; +Cc: dev, liugaoxiang
On 10/18/2021 12:19 AM, Gaoxiang Liu wrote:
> To improve performance in vhost Tx/Rx, merge vhost stats loop.
> eth_vhost_tx has 2 loop of send num iteraion.
> It can be merge into one.
> eth_vhost_rx has the same issue as Tx.
>
> Signed-off-by: Gaoxiang Liu<gaoxiangliu0@163.com>
Hi Gaoxiang,
We are trying to keep unique identity in the git repo, your previous in git
records are with your work email, so I will update this commit to use same:
Signed-off-by: Gaoxiang Liu <liugaoxiang@huawei.com>
Please let us know if you have an objection.
^ permalink raw reply [flat|nested] 11+ messages in thread
* Re: [dpdk-dev] [PATCH v3] net/vhost: merge vhost stats loop in vhost Tx/Rx
2021-10-21 18:56 ` Ferruh Yigit
@ 2021-10-22 1:35 ` Gaoxiang Liu
0 siblings, 0 replies; 11+ messages in thread
From: Gaoxiang Liu @ 2021-10-22 1:35 UTC (permalink / raw)
To: Yigit, Ferruh; +Cc: maxime.coquelin, Xia, Chenbo, dev, liugaoxiang
Hi Ferruh,
liugaoxiang@huawei.com is also my email address. You can update the commit to use it.
Thanks,
Gaoxiang
发自 网易邮箱大师
---- 回复的原邮件 ----
| 发件人 | Ferruh Yigit<ferruh.yigit@intel.com> |
| 日期 | 2021年10月22日 02:56 |
| 收件人 | Gaoxiang Liu<gaoxiangliu0@163.com>、maxime.coquelin@redhat.com<maxime.coquelin@redhat.com>、chenbo.xia@intel.com<chenbo.xia@intel.com> |
| 抄送至 | dev@dpdk.org<dev@dpdk.org>、liugaoxiang@huawei.com<liugaoxiang@huawei.com> |
| 主题 | Re: [dpdk-dev] [PATCH v3] net/vhost: merge vhost stats loop in vhost Tx/Rx |
On 10/18/2021 12:19 AM, Gaoxiang Liu wrote:
> To improve performance in vhost Tx/Rx, merge vhost stats loop.
> eth_vhost_tx has 2 loop of send num iteraion.
> It can be merge into one.
> eth_vhost_rx has the same issue as Tx.
>
> Signed-off-by: Gaoxiang Liu<gaoxiangliu0@163.com>
Hi Gaoxiang,
We are trying to keep unique identity in the git repo, your previous in git
records are with your work email, so I will update this commit to use same:
Signed-off-by: Gaoxiang Liu <liugaoxiang@huawei.com>
Please let us know if you have an objection.
^ permalink raw reply [flat|nested] 11+ messages in thread