Currently, we store the statistics in the independent fields of virtnet_stats,
this is not scalable when we want to add more counters. As suggested by Michael,
this patch convert it to an array and use the enum as the index to access them.
Signed-off-by: Jason Wang <[email protected]>
---
drivers/net/virtio_net.c | 30 +++++++++++++++++-------------
1 files changed, 17 insertions(+), 13 deletions(-)
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 5214b1e..6e4aa6f 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -41,13 +41,17 @@ module_param(gso, bool, 0444);
#define VIRTNET_SEND_COMMAND_SG_MAX 2
#define VIRTNET_DRIVER_VERSION "1.0.0"
+enum virtnet_stats_type {
+ VIRTNET_TX_BYTES,
+ VIRTNET_TX_PACKETS,
+ VIRTNET_RX_BYTES,
+ VIRTNET_RX_PACKETS,
+ VIRTNET_NUM_STATS,
+};
+
struct virtnet_stats {
struct u64_stats_sync syncp;
- u64 tx_bytes;
- u64 tx_packets;
-
- u64 rx_bytes;
- u64 rx_packets;
+ u64 data[VIRTNET_NUM_STATS];
};
struct virtnet_info {
@@ -301,8 +305,8 @@ static void receive_buf(struct net_device *dev, void *buf, unsigned int len)
hdr = skb_vnet_hdr(skb);
u64_stats_update_begin(&stats->syncp);
- stats->rx_bytes += skb->len;
- stats->rx_packets++;
+ stats->data[VIRTNET_RX_BYTES] += skb->len;
+ stats->data[VIRTNET_RX_PACKETS]++;
u64_stats_update_end(&stats->syncp);
if (hdr->hdr.flags & VIRTIO_NET_HDR_F_NEEDS_CSUM) {
@@ -566,8 +570,8 @@ static unsigned int free_old_xmit_skbs(struct virtnet_info *vi)
pr_debug("Sent skb %p\n", skb);
u64_stats_update_begin(&stats->syncp);
- stats->tx_bytes += skb->len;
- stats->tx_packets++;
+ stats->data[VIRTNET_TX_BYTES] += skb->len;
+ stats->data[VIRTNET_TX_PACKETS]++;
u64_stats_update_end(&stats->syncp);
tot_sgs += skb_vnet_hdr(skb)->num_sg;
@@ -704,10 +708,10 @@ static struct rtnl_link_stats64 *virtnet_stats(struct net_device *dev,
do {
start = u64_stats_fetch_begin(&stats->syncp);
- tpackets = stats->tx_packets;
- tbytes = stats->tx_bytes;
- rpackets = stats->rx_packets;
- rbytes = stats->rx_bytes;
+ tpackets = stats->data[VIRTNET_TX_PACKETS];
+ tbytes = stats->data[VIRTNET_TX_BYTES];
+ rpackets = stats->data[VIRTNET_RX_PACKETS];
+ rbytes = stats->data[VIRTNET_RX_BYTES];
} while (u64_stats_fetch_retry(&stats->syncp, start));
tot->rx_packets += rpackets;
Satistics counters is useful for debugging and performance optimization, so this
patch lets virtio_net driver collect following and export them to userspace
through "ethtool -S":
- number of packets sent/received
- number of bytes sent/received
- number of callbacks for tx/rx
- number of kick for tx/rx
- number of bytes/packets queued for tx
As virtnet_stats were per-cpu, so both per-cpu and gloabl satistics were
collected like:
NIC statistics:
tx_bytes[0]: 1731209929
tx_packets[0]: 60685
tx_kicks[0]: 63
tx_callbacks[0]: 73
tx_queued_bytes[0]: 1935749360
tx_queued_packets[0]: 80652
rx_bytes[0]: 2695648
rx_packets[0]: 40767
rx_kicks[0]: 1
rx_callbacks[0]: 2077
tx_bytes[1]: 9105588697
tx_packets[1]: 344150
tx_kicks[1]: 162
tx_callbacks[1]: 905
tx_queued_bytes[1]: 8901049412
tx_queued_packets[1]: 324184
rx_bytes[1]: 23679828
rx_packets[1]: 358770
rx_kicks[1]: 6
rx_callbacks[1]: 17717
tx_bytes: 10836798626
tx_packets: 404835
tx_kicks: 225
tx_callbacks: 978
tx_queued_bytes: 10836798772
tx_queued_packets: 404836
rx_bytes: 26375476
rx_packets: 399537
rx_kicks: 7
rx_callbacks: 19794
TODO:
- more statistics
- calculate the pending bytes/pkts
Signed-off-by: Jason Wang <[email protected]>
---
Changes from v1:
- style & typo fixs
- convert the statistics fields to array
- use unlikely()
---
drivers/net/virtio_net.c | 115 +++++++++++++++++++++++++++++++++++++++++++++-
1 files changed, 113 insertions(+), 2 deletions(-)
diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
index 6e4aa6f..909a0a7 100644
--- a/drivers/net/virtio_net.c
+++ b/drivers/net/virtio_net.c
@@ -44,8 +44,14 @@ module_param(gso, bool, 0444);
enum virtnet_stats_type {
VIRTNET_TX_BYTES,
VIRTNET_TX_PACKETS,
+ VIRTNET_TX_KICKS,
+ VIRTNET_TX_CBS,
+ VIRTNET_TX_Q_BYTES,
+ VIRTNET_TX_Q_PACKETS,
VIRTNET_RX_BYTES,
VIRTNET_RX_PACKETS,
+ VIRTNET_RX_KICKS,
+ VIRTNET_RX_CBS,
VIRTNET_NUM_STATS,
};
@@ -54,6 +60,21 @@ struct virtnet_stats {
u64 data[VIRTNET_NUM_STATS];
};
+static struct {
+ char string[ETH_GSTRING_LEN];
+} virtnet_stats_str_attr[] = {
+ { "tx_bytes" },
+ { "tx_packets" },
+ { "tx_kicks" },
+ { "tx_callbacks" },
+ { "tx_queued_bytes" },
+ { "tx_queued_packets" },
+ { "rx_bytes" },
+ { "rx_packets" },
+ { "rx_kicks" },
+ { "rx_callbacks" },
+};
+
struct virtnet_info {
struct virtio_device *vdev;
struct virtqueue *rvq, *svq, *cvq;
@@ -146,6 +167,11 @@ static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
static void skb_xmit_done(struct virtqueue *svq)
{
struct virtnet_info *vi = svq->vdev->priv;
+ struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
+
+ u64_stats_update_begin(&stats->syncp);
+ stats->data[VIRTNET_TX_CBS]++;
+ u64_stats_update_end(&stats->syncp);
/* Suppress further interrupts. */
virtqueue_disable_cb(svq);
@@ -465,6 +491,7 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
{
int err;
bool oom;
+ struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
do {
if (vi->mergeable_rx_bufs)
@@ -481,13 +508,24 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
} while (err > 0);
if (unlikely(vi->num > vi->max))
vi->max = vi->num;
- virtqueue_kick(vi->rvq);
+ if (virtqueue_kick_prepare(vi->rvq)) {
+ virtqueue_notify(vi->rvq);
+ u64_stats_update_begin(&stats->syncp);
+ stats->data[VIRTNET_RX_KICKS]++;
+ u64_stats_update_end(&stats->syncp);
+ }
return !oom;
}
static void skb_recv_done(struct virtqueue *rvq)
{
struct virtnet_info *vi = rvq->vdev->priv;
+ struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
+
+ u64_stats_update_begin(&stats->syncp);
+ stats->data[VIRTNET_RX_CBS]++;
+ u64_stats_update_end(&stats->syncp);
+
/* Schedule NAPI, Suppress further interrupts if successful. */
if (napi_schedule_prep(&vi->napi)) {
virtqueue_disable_cb(rvq);
@@ -630,7 +668,9 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
{
struct virtnet_info *vi = netdev_priv(dev);
+ struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
int capacity;
+ bool kick;
/* Free up any pending old buffers before queueing new ones. */
free_old_xmit_skbs(vi);
@@ -655,7 +695,17 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
kfree_skb(skb);
return NETDEV_TX_OK;
}
- virtqueue_kick(vi->svq);
+
+ kick = virtqueue_kick_prepare(vi->svq);
+ if (unlikely(kick))
+ virtqueue_notify(vi->svq);
+
+ u64_stats_update_begin(&stats->syncp);
+ if (unlikely(kick))
+ stats->data[VIRTNET_TX_KICKS]++;
+ stats->data[VIRTNET_TX_Q_BYTES] += skb->len;
+ stats->data[VIRTNET_TX_Q_PACKETS]++;
+ u64_stats_update_end(&stats->syncp);
/* Don't wait up for transmitted skbs to be freed. */
skb_orphan(skb);
@@ -943,10 +993,71 @@ static void virtnet_get_drvinfo(struct net_device *dev,
}
+static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
+{
+ int i, cpu;
+ switch (stringset) {
+ case ETH_SS_STATS:
+ for_each_possible_cpu(cpu)
+ for (i = 0; i < VIRTNET_NUM_STATS; i++) {
+ sprintf(buf, "%s[%u]",
+ virtnet_stats_str_attr[i].string, cpu);
+ buf += ETH_GSTRING_LEN;
+ }
+ for (i = 0; i < VIRTNET_NUM_STATS; i++) {
+ memcpy(buf, virtnet_stats_str_attr[i].string,
+ ETH_GSTRING_LEN);
+ buf += ETH_GSTRING_LEN;
+ }
+ break;
+ }
+}
+
+static int virtnet_get_sset_count(struct net_device *dev, int sset)
+{
+ switch (sset) {
+ case ETH_SS_STATS:
+ return VIRTNET_NUM_STATS * (num_possible_cpus() + 1);
+ default:
+ return -EOPNOTSUPP;
+ }
+}
+
+static void virtnet_get_ethtool_stats(struct net_device *dev,
+ struct ethtool_stats *stats, u64 *buf)
+{
+ struct virtnet_info *vi = netdev_priv(dev);
+ int cpu, i;
+ unsigned int start;
+ struct virtnet_stats sample, total;
+
+ memset(&total, 0, sizeof(total));
+
+ for_each_possible_cpu(cpu) {
+ struct virtnet_stats *s = per_cpu_ptr(vi->stats, cpu);
+ do {
+ start = u64_stats_fetch_begin(&s->syncp);
+ memcpy(&sample.data, &s->data,
+ sizeof(u64) * VIRTNET_NUM_STATS);
+ } while (u64_stats_fetch_retry(&s->syncp, start));
+
+ for (i = 0; i < VIRTNET_NUM_STATS; i++) {
+ *buf = sample.data[i];
+ total.data[i] += sample.data[i];
+ buf++;
+ }
+ }
+
+ memcpy(buf, &total.data, sizeof(u64) * VIRTNET_NUM_STATS);
+}
+
static const struct ethtool_ops virtnet_ethtool_ops = {
.get_drvinfo = virtnet_get_drvinfo,
.get_link = ethtool_op_get_link,
.get_ringparam = virtnet_get_ringparam,
+ .get_ethtool_stats = virtnet_get_ethtool_stats,
+ .get_strings = virtnet_get_strings,
+ .get_sset_count = virtnet_get_sset_count,
};
#define MIN_MTU 68
On Wed, 2012-06-06 at 15:52 +0800, Jason Wang wrote:
> Currently, we store the statistics in the independent fields of virtnet_stats,
> this is not scalable when we want to add more counters. As suggested by Michael,
> this patch convert it to an array and use the enum as the index to access them.
>
> Signed-off-by: Jason Wang <[email protected]>
> ---
> drivers/net/virtio_net.c | 30 +++++++++++++++++-------------
> 1 files changed, 17 insertions(+), 13 deletions(-)
>
> struct virtnet_stats {
> struct u64_stats_sync syncp;
> - u64 tx_bytes;
> - u64 tx_packets;
> -
> - u64 rx_bytes;
> - u64 rx_packets;
> + u64 data[VIRTNET_NUM_STATS];
> };
>
Interesting, but I fear you'll have a lot of problems.
Current code is buggy, and you are adding more possible races.
We could have one cpu doing the :
u64_stats_update_begin(&stats->syncp);
stats->rx_bytes += skb->len;
stats->rx_packets++;
u64_stats_update_end(&stats->syncp);
And another one doing :
u64_stats_update_begin(&stats->syncp);
stats->tx_bytes += skb->len;
stats->tx_packets++;
u64_stats_update_end(&stats->syncp);
And one syncp sequence increment can be lost, since both cpus are
basically doing this at the same time :
write_seqcount_begin(&syncp->seq);
I'll send a fix in a separate thread.
On Wed, Jun 06, 2012 at 03:52:17PM +0800, Jason Wang wrote:
> Satistics counters is useful for debugging and performance optimization, so this
> patch lets virtio_net driver collect following and export them to userspace
> through "ethtool -S":
>
> - number of packets sent/received
> - number of bytes sent/received
> - number of callbacks for tx/rx
> - number of kick for tx/rx
> - number of bytes/packets queued for tx
>
> As virtnet_stats were per-cpu, so both per-cpu and gloabl satistics were
> collected like:
>
> NIC statistics:
> tx_bytes[0]: 1731209929
> tx_packets[0]: 60685
> tx_kicks[0]: 63
> tx_callbacks[0]: 73
> tx_queued_bytes[0]: 1935749360
> tx_queued_packets[0]: 80652
> rx_bytes[0]: 2695648
> rx_packets[0]: 40767
> rx_kicks[0]: 1
> rx_callbacks[0]: 2077
> tx_bytes[1]: 9105588697
> tx_packets[1]: 344150
> tx_kicks[1]: 162
> tx_callbacks[1]: 905
> tx_queued_bytes[1]: 8901049412
> tx_queued_packets[1]: 324184
> rx_bytes[1]: 23679828
> rx_packets[1]: 358770
> rx_kicks[1]: 6
> rx_callbacks[1]: 17717
> tx_bytes: 10836798626
> tx_packets: 404835
> tx_kicks: 225
> tx_callbacks: 978
> tx_queued_bytes: 10836798772
> tx_queued_packets: 404836
> rx_bytes: 26375476
> rx_packets: 399537
> rx_kicks: 7
> rx_callbacks: 19794
>
> TODO:
>
> - more statistics
> - calculate the pending bytes/pkts
>
Do we need that? pending is (queued - packets), no?
> Signed-off-by: Jason Wang <[email protected]>
>
> ---
> Changes from v1:
>
> - style & typo fixs
> - convert the statistics fields to array
> - use unlikely()
> ---
> drivers/net/virtio_net.c | 115 +++++++++++++++++++++++++++++++++++++++++++++-
> 1 files changed, 113 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 6e4aa6f..909a0a7 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -44,8 +44,14 @@ module_param(gso, bool, 0444);
> enum virtnet_stats_type {
> VIRTNET_TX_BYTES,
> VIRTNET_TX_PACKETS,
> + VIRTNET_TX_KICKS,
> + VIRTNET_TX_CBS,
> + VIRTNET_TX_Q_BYTES,
> + VIRTNET_TX_Q_PACKETS,
> VIRTNET_RX_BYTES,
> VIRTNET_RX_PACKETS,
> + VIRTNET_RX_KICKS,
> + VIRTNET_RX_CBS,
> VIRTNET_NUM_STATS,
> };
>
> @@ -54,6 +60,21 @@ struct virtnet_stats {
> u64 data[VIRTNET_NUM_STATS];
> };
>
> +static struct {
static const?
> + char string[ETH_GSTRING_LEN];
> +} virtnet_stats_str_attr[] = {
> + { "tx_bytes" },
> + { "tx_packets" },
> + { "tx_kicks" },
> + { "tx_callbacks" },
> + { "tx_queued_bytes" },
> + { "tx_queued_packets" },
> + { "rx_bytes" },
> + { "rx_packets" },
> + { "rx_kicks" },
> + { "rx_callbacks" },
> +};
> +
> struct virtnet_info {
> struct virtio_device *vdev;
> struct virtqueue *rvq, *svq, *cvq;
> @@ -146,6 +167,11 @@ static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
> static void skb_xmit_done(struct virtqueue *svq)
> {
> struct virtnet_info *vi = svq->vdev->priv;
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
> +
> + u64_stats_update_begin(&stats->syncp);
> + stats->data[VIRTNET_TX_CBS]++;
> + u64_stats_update_end(&stats->syncp);
>
> /* Suppress further interrupts. */
> virtqueue_disable_cb(svq);
> @@ -465,6 +491,7 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
> {
> int err;
> bool oom;
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>
> do {
> if (vi->mergeable_rx_bufs)
> @@ -481,13 +508,24 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
> } while (err > 0);
> if (unlikely(vi->num > vi->max))
> vi->max = vi->num;
> - virtqueue_kick(vi->rvq);
> + if (virtqueue_kick_prepare(vi->rvq)) {
if (unlikely())
also move stats here where they are actually used?
> + virtqueue_notify(vi->rvq);
> + u64_stats_update_begin(&stats->syncp);
> + stats->data[VIRTNET_RX_KICKS]++;
> + u64_stats_update_end(&stats->syncp);
> + }
> return !oom;
> }
>
> static void skb_recv_done(struct virtqueue *rvq)
> {
> struct virtnet_info *vi = rvq->vdev->priv;
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
> +
> + u64_stats_update_begin(&stats->syncp);
> + stats->data[VIRTNET_RX_CBS]++;
> + u64_stats_update_end(&stats->syncp);
> +
> /* Schedule NAPI, Suppress further interrupts if successful. */
> if (napi_schedule_prep(&vi->napi)) {
> virtqueue_disable_cb(rvq);
> @@ -630,7 +668,9 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
> static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> {
> struct virtnet_info *vi = netdev_priv(dev);
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
> int capacity;
> + bool kick;
>
> /* Free up any pending old buffers before queueing new ones. */
> free_old_xmit_skbs(vi);
> @@ -655,7 +695,17 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> kfree_skb(skb);
> return NETDEV_TX_OK;
> }
> - virtqueue_kick(vi->svq);
> +
> + kick = virtqueue_kick_prepare(vi->svq);
> + if (unlikely(kick))
> + virtqueue_notify(vi->svq);
> +
> + u64_stats_update_begin(&stats->syncp);
> + if (unlikely(kick))
> + stats->data[VIRTNET_TX_KICKS]++;
> + stats->data[VIRTNET_TX_Q_BYTES] += skb->len;
> + stats->data[VIRTNET_TX_Q_PACKETS]++;
> + u64_stats_update_end(&stats->syncp);
>
> /* Don't wait up for transmitted skbs to be freed. */
> skb_orphan(skb);
> @@ -943,10 +993,71 @@ static void virtnet_get_drvinfo(struct net_device *dev,
>
> }
>
> +static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
> +{
> + int i, cpu;
> + switch (stringset) {
> + case ETH_SS_STATS:
> + for_each_possible_cpu(cpu)
> + for (i = 0; i < VIRTNET_NUM_STATS; i++) {
> + sprintf(buf, "%s[%u]",
> + virtnet_stats_str_attr[i].string, cpu);
> + buf += ETH_GSTRING_LEN;
I would do
ret = snprintf(buf, ETH_GSTRING_LEN, ...)
BUG_ON(ret >= ETH_GSTRING_LEN);
here to make it more robust.
> + }
> + for (i = 0; i < VIRTNET_NUM_STATS; i++) {
> + memcpy(buf, virtnet_stats_str_attr[i].string,
> + ETH_GSTRING_LEN);
> + buf += ETH_GSTRING_LEN;
> + }
So why not just memcpy the whole array there?
memcpy(buf, virtnet_stats_str_attr,
sizeof virtnet_stats_str_attr);
> + break;
> + }
> +}
> +
> +static int virtnet_get_sset_count(struct net_device *dev, int sset)
> +{
> + switch (sset) {
> + case ETH_SS_STATS:
also add
BUILD_BUG_ON(VIRTNET_NUM_STATS != (sizeof virtnet_stats_str_attr) / ETH_GSTRING_LEN);
> + return VIRTNET_NUM_STATS * (num_possible_cpus() + 1);
> + default:
> + return -EOPNOTSUPP;
> + }
> +}
> +
> +static void virtnet_get_ethtool_stats(struct net_device *dev,
> + struct ethtool_stats *stats, u64 *buf)
> +{
> + struct virtnet_info *vi = netdev_priv(dev);
> + int cpu, i;
> + unsigned int start;
> + struct virtnet_stats sample, total;
> +
> + memset(&total, 0, sizeof(total));
sizeof total
when operand is a variable,
to distinguish from when it is a type.
> +
> + for_each_possible_cpu(cpu) {
> + struct virtnet_stats *s = per_cpu_ptr(vi->stats, cpu);
> + do {
> + start = u64_stats_fetch_begin(&s->syncp);
> + memcpy(&sample.data, &s->data,
> + sizeof(u64) * VIRTNET_NUM_STATS);
> + } while (u64_stats_fetch_retry(&s->syncp, start));
> +
> + for (i = 0; i < VIRTNET_NUM_STATS; i++) {
> + *buf = sample.data[i];
> + total.data[i] += sample.data[i];
> + buf++;
> + }
> + }
> +
> + memcpy(buf, &total.data, sizeof(u64) * VIRTNET_NUM_STATS);
> +}
> +
> static const struct ethtool_ops virtnet_ethtool_ops = {
> .get_drvinfo = virtnet_get_drvinfo,
> .get_link = ethtool_op_get_link,
> .get_ringparam = virtnet_get_ringparam,
> + .get_ethtool_stats = virtnet_get_ethtool_stats,
> + .get_strings = virtnet_get_strings,
> + .get_sset_count = virtnet_get_sset_count,
> };
>
> #define MIN_MTU 68
On Wed, Jun 06, 2012 at 03:52:17PM +0800, Jason Wang wrote:
> Satistics counters is useful for debugging and performance optimization, so this
> patch lets virtio_net driver collect following and export them to userspace
> through "ethtool -S":
>
> - number of packets sent/received
> - number of bytes sent/received
> - number of callbacks for tx/rx
> - number of kick for tx/rx
> - number of bytes/packets queued for tx
>
> As virtnet_stats were per-cpu, so both per-cpu and gloabl satistics were
> collected like:
>
> NIC statistics:
> tx_bytes[0]: 1731209929
> tx_packets[0]: 60685
> tx_kicks[0]: 63
> tx_callbacks[0]: 73
> tx_queued_bytes[0]: 1935749360
> tx_queued_packets[0]: 80652
> rx_bytes[0]: 2695648
> rx_packets[0]: 40767
> rx_kicks[0]: 1
> rx_callbacks[0]: 2077
> tx_bytes[1]: 9105588697
> tx_packets[1]: 344150
> tx_kicks[1]: 162
> tx_callbacks[1]: 905
> tx_queued_bytes[1]: 8901049412
> tx_queued_packets[1]: 324184
> rx_bytes[1]: 23679828
> rx_packets[1]: 358770
> rx_kicks[1]: 6
> rx_callbacks[1]: 17717
> tx_bytes: 10836798626
> tx_packets: 404835
> tx_kicks: 225
> tx_callbacks: 978
> tx_queued_bytes: 10836798772
> tx_queued_packets: 404836
> rx_bytes: 26375476
> rx_packets: 399537
> rx_kicks: 7
> rx_callbacks: 19794
>
> TODO:
>
> - more statistics
> - calculate the pending bytes/pkts
>
> Signed-off-by: Jason Wang <[email protected]>
>
> ---
> Changes from v1:
>
> - style & typo fixs
> - convert the statistics fields to array
> - use unlikely()
> ---
> drivers/net/virtio_net.c | 115 +++++++++++++++++++++++++++++++++++++++++++++-
> 1 files changed, 113 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 6e4aa6f..909a0a7 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -44,8 +44,14 @@ module_param(gso, bool, 0444);
> enum virtnet_stats_type {
> VIRTNET_TX_BYTES,
> VIRTNET_TX_PACKETS,
> + VIRTNET_TX_KICKS,
> + VIRTNET_TX_CBS,
> + VIRTNET_TX_Q_BYTES,
> + VIRTNET_TX_Q_PACKETS,
What about counting the time we spend with queue
stopped and # of times we stop the queue?
> VIRTNET_RX_BYTES,
> VIRTNET_RX_PACKETS,
> + VIRTNET_RX_KICKS,
> + VIRTNET_RX_CBS,
What about a counter for oom on rx?
> VIRTNET_NUM_STATS,
> };
>
> @@ -54,6 +60,21 @@ struct virtnet_stats {
> u64 data[VIRTNET_NUM_STATS];
> };
>
> +static struct {
> + char string[ETH_GSTRING_LEN];
> +} virtnet_stats_str_attr[] = {
> + { "tx_bytes" },
> + { "tx_packets" },
> + { "tx_kicks" },
> + { "tx_callbacks" },
> + { "tx_queued_bytes" },
> + { "tx_queued_packets" },
> + { "rx_bytes" },
> + { "rx_packets" },
> + { "rx_kicks" },
> + { "rx_callbacks" },
> +};
> +
> struct virtnet_info {
> struct virtio_device *vdev;
> struct virtqueue *rvq, *svq, *cvq;
> @@ -146,6 +167,11 @@ static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
> static void skb_xmit_done(struct virtqueue *svq)
> {
> struct virtnet_info *vi = svq->vdev->priv;
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
> +
> + u64_stats_update_begin(&stats->syncp);
> + stats->data[VIRTNET_TX_CBS]++;
> + u64_stats_update_end(&stats->syncp);
>
> /* Suppress further interrupts. */
> virtqueue_disable_cb(svq);
> @@ -465,6 +491,7 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
> {
> int err;
> bool oom;
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>
> do {
> if (vi->mergeable_rx_bufs)
> @@ -481,13 +508,24 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
> } while (err > 0);
> if (unlikely(vi->num > vi->max))
> vi->max = vi->num;
> - virtqueue_kick(vi->rvq);
> + if (virtqueue_kick_prepare(vi->rvq)) {
> + virtqueue_notify(vi->rvq);
> + u64_stats_update_begin(&stats->syncp);
> + stats->data[VIRTNET_RX_KICKS]++;
> + u64_stats_update_end(&stats->syncp);
> + }
> return !oom;
> }
>
> static void skb_recv_done(struct virtqueue *rvq)
> {
> struct virtnet_info *vi = rvq->vdev->priv;
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
> +
> + u64_stats_update_begin(&stats->syncp);
> + stats->data[VIRTNET_RX_CBS]++;
> + u64_stats_update_end(&stats->syncp);
> +
This data path so not entirely free.
I am guessing the overhead is not measureable but
did you check?
An alternative is to count when napi callbacks
are envoked. If we also count when weight was exceeded
we get almost the same result.
> /* Schedule NAPI, Suppress further interrupts if successful. */
> if (napi_schedule_prep(&vi->napi)) {
> virtqueue_disable_cb(rvq);
> @@ -630,7 +668,9 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
> static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> {
> struct virtnet_info *vi = netdev_priv(dev);
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
> int capacity;
> + bool kick;
>
> /* Free up any pending old buffers before queueing new ones. */
> free_old_xmit_skbs(vi);
> @@ -655,7 +695,17 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> kfree_skb(skb);
> return NETDEV_TX_OK;
> }
> - virtqueue_kick(vi->svq);
> +
> + kick = virtqueue_kick_prepare(vi->svq);
> + if (unlikely(kick))
> + virtqueue_notify(vi->svq);
> +
> + u64_stats_update_begin(&stats->syncp);
> + if (unlikely(kick))
> + stats->data[VIRTNET_TX_KICKS]++;
> + stats->data[VIRTNET_TX_Q_BYTES] += skb->len;
> + stats->data[VIRTNET_TX_Q_PACKETS]++;
> + u64_stats_update_end(&stats->syncp);
>
> /* Don't wait up for transmitted skbs to be freed. */
> skb_orphan(skb);
> @@ -943,10 +993,71 @@ static void virtnet_get_drvinfo(struct net_device *dev,
>
> }
>
> +static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
> +{
> + int i, cpu;
> + switch (stringset) {
> + case ETH_SS_STATS:
> + for_each_possible_cpu(cpu)
> + for (i = 0; i < VIRTNET_NUM_STATS; i++) {
> + sprintf(buf, "%s[%u]",
> + virtnet_stats_str_attr[i].string, cpu);
> + buf += ETH_GSTRING_LEN;
> + }
> + for (i = 0; i < VIRTNET_NUM_STATS; i++) {
> + memcpy(buf, virtnet_stats_str_attr[i].string,
> + ETH_GSTRING_LEN);
> + buf += ETH_GSTRING_LEN;
> + }
> + break;
> + }
> +}
> +
> +static int virtnet_get_sset_count(struct net_device *dev, int sset)
> +{
> + switch (sset) {
> + case ETH_SS_STATS:
> + return VIRTNET_NUM_STATS * (num_possible_cpus() + 1);
> + default:
> + return -EOPNOTSUPP;
> + }
> +}
> +
> +static void virtnet_get_ethtool_stats(struct net_device *dev,
> + struct ethtool_stats *stats, u64 *buf)
> +{
> + struct virtnet_info *vi = netdev_priv(dev);
> + int cpu, i;
> + unsigned int start;
> + struct virtnet_stats sample, total;
> +
> + memset(&total, 0, sizeof(total));
> +
> + for_each_possible_cpu(cpu) {
> + struct virtnet_stats *s = per_cpu_ptr(vi->stats, cpu);
> + do {
> + start = u64_stats_fetch_begin(&s->syncp);
> + memcpy(&sample.data, &s->data,
> + sizeof(u64) * VIRTNET_NUM_STATS);
> + } while (u64_stats_fetch_retry(&s->syncp, start));
> +
> + for (i = 0; i < VIRTNET_NUM_STATS; i++) {
> + *buf = sample.data[i];
> + total.data[i] += sample.data[i];
> + buf++;
> + }
> + }
> +
> + memcpy(buf, &total.data, sizeof(u64) * VIRTNET_NUM_STATS);
> +}
> +
> static const struct ethtool_ops virtnet_ethtool_ops = {
> .get_drvinfo = virtnet_get_drvinfo,
> .get_link = ethtool_op_get_link,
> .get_ringparam = virtnet_get_ringparam,
> + .get_ethtool_stats = virtnet_get_ethtool_stats,
> + .get_strings = virtnet_get_strings,
> + .get_sset_count = virtnet_get_sset_count,
> };
>
> #define MIN_MTU 68
On 06/06/2012 04:27 PM, Michael S. Tsirkin wrote:
> On Wed, Jun 06, 2012 at 03:52:17PM +0800, Jason Wang wrote:
>> Satistics counters is useful for debugging and performance optimization, so this
>> patch lets virtio_net driver collect following and export them to userspace
>> through "ethtool -S":
>>
>> - number of packets sent/received
>> - number of bytes sent/received
>> - number of callbacks for tx/rx
>> - number of kick for tx/rx
>> - number of bytes/packets queued for tx
>>
>> As virtnet_stats were per-cpu, so both per-cpu and gloabl satistics were
>> collected like:
>>
>> NIC statistics:
>> tx_bytes[0]: 1731209929
>> tx_packets[0]: 60685
>> tx_kicks[0]: 63
>> tx_callbacks[0]: 73
>> tx_queued_bytes[0]: 1935749360
>> tx_queued_packets[0]: 80652
>> rx_bytes[0]: 2695648
>> rx_packets[0]: 40767
>> rx_kicks[0]: 1
>> rx_callbacks[0]: 2077
>> tx_bytes[1]: 9105588697
>> tx_packets[1]: 344150
>> tx_kicks[1]: 162
>> tx_callbacks[1]: 905
>> tx_queued_bytes[1]: 8901049412
>> tx_queued_packets[1]: 324184
>> rx_bytes[1]: 23679828
>> rx_packets[1]: 358770
>> rx_kicks[1]: 6
>> rx_callbacks[1]: 17717
>> tx_bytes: 10836798626
>> tx_packets: 404835
>> tx_kicks: 225
>> tx_callbacks: 978
>> tx_queued_bytes: 10836798772
>> tx_queued_packets: 404836
>> rx_bytes: 26375476
>> rx_packets: 399537
>> rx_kicks: 7
>> rx_callbacks: 19794
>>
>> TODO:
>>
>> - more statistics
>> - calculate the pending bytes/pkts
>>
> Do we need that? pending is (queued - packets), no?
>
No, if we choose to calculate by tools.
>> Signed-off-by: Jason Wang<[email protected]>
>>
>> ---
>> Changes from v1:
>>
>> - style& typo fixs
>> - convert the statistics fields to array
>> - use unlikely()
>> ---
>> drivers/net/virtio_net.c | 115 +++++++++++++++++++++++++++++++++++++++++++++-
>> 1 files changed, 113 insertions(+), 2 deletions(-)
>>
>> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
>> index 6e4aa6f..909a0a7 100644
>> --- a/drivers/net/virtio_net.c
>> +++ b/drivers/net/virtio_net.c
>> @@ -44,8 +44,14 @@ module_param(gso, bool, 0444);
>> enum virtnet_stats_type {
>> VIRTNET_TX_BYTES,
>> VIRTNET_TX_PACKETS,
>> + VIRTNET_TX_KICKS,
>> + VIRTNET_TX_CBS,
>> + VIRTNET_TX_Q_BYTES,
>> + VIRTNET_TX_Q_PACKETS,
>> VIRTNET_RX_BYTES,
>> VIRTNET_RX_PACKETS,
>> + VIRTNET_RX_KICKS,
>> + VIRTNET_RX_CBS,
>> VIRTNET_NUM_STATS,
>> };
>>
>> @@ -54,6 +60,21 @@ struct virtnet_stats {
>> u64 data[VIRTNET_NUM_STATS];
>> };
>>
>> +static struct {
> static const?
>
Sorry, forget this.
>> + char string[ETH_GSTRING_LEN];
>> +} virtnet_stats_str_attr[] = {
>> + { "tx_bytes" },
>> + { "tx_packets" },
>> + { "tx_kicks" },
>> + { "tx_callbacks" },
>> + { "tx_queued_bytes" },
>> + { "tx_queued_packets" },
>> + { "rx_bytes" },
>> + { "rx_packets" },
>> + { "rx_kicks" },
>> + { "rx_callbacks" },
>> +};
>> +
>> struct virtnet_info {
>> struct virtio_device *vdev;
>> struct virtqueue *rvq, *svq, *cvq;
>> @@ -146,6 +167,11 @@ static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
>> static void skb_xmit_done(struct virtqueue *svq)
>> {
>> struct virtnet_info *vi = svq->vdev->priv;
>> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>> +
>> + u64_stats_update_begin(&stats->syncp);
>> + stats->data[VIRTNET_TX_CBS]++;
>> + u64_stats_update_end(&stats->syncp);
>>
>> /* Suppress further interrupts. */
>> virtqueue_disable_cb(svq);
>> @@ -465,6 +491,7 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
>> {
>> int err;
>> bool oom;
>> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>>
>> do {
>> if (vi->mergeable_rx_bufs)
>> @@ -481,13 +508,24 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
>> } while (err> 0);
>> if (unlikely(vi->num> vi->max))
>> vi->max = vi->num;
>> - virtqueue_kick(vi->rvq);
>> + if (virtqueue_kick_prepare(vi->rvq)) {
> if (unlikely())
> also move stats here where they are actually used?
Sure.
>> + virtqueue_notify(vi->rvq);
>> + u64_stats_update_begin(&stats->syncp);
>> + stats->data[VIRTNET_RX_KICKS]++;
>> + u64_stats_update_end(&stats->syncp);
>> + }
>> return !oom;
>> }
>>
>> static void skb_recv_done(struct virtqueue *rvq)
>> {
>> struct virtnet_info *vi = rvq->vdev->priv;
>> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>> +
>> + u64_stats_update_begin(&stats->syncp);
>> + stats->data[VIRTNET_RX_CBS]++;
>> + u64_stats_update_end(&stats->syncp);
>> +
>> /* Schedule NAPI, Suppress further interrupts if successful. */
>> if (napi_schedule_prep(&vi->napi)) {
>> virtqueue_disable_cb(rvq);
>> @@ -630,7 +668,9 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
>> static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
>> {
>> struct virtnet_info *vi = netdev_priv(dev);
>> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>> int capacity;
>> + bool kick;
>>
>> /* Free up any pending old buffers before queueing new ones. */
>> free_old_xmit_skbs(vi);
>> @@ -655,7 +695,17 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
>> kfree_skb(skb);
>> return NETDEV_TX_OK;
>> }
>> - virtqueue_kick(vi->svq);
>> +
>> + kick = virtqueue_kick_prepare(vi->svq);
>> + if (unlikely(kick))
>> + virtqueue_notify(vi->svq);
>> +
>> + u64_stats_update_begin(&stats->syncp);
>> + if (unlikely(kick))
>> + stats->data[VIRTNET_TX_KICKS]++;
>> + stats->data[VIRTNET_TX_Q_BYTES] += skb->len;
>> + stats->data[VIRTNET_TX_Q_PACKETS]++;
>> + u64_stats_update_end(&stats->syncp);
>>
>> /* Don't wait up for transmitted skbs to be freed. */
>> skb_orphan(skb);
>> @@ -943,10 +993,71 @@ static void virtnet_get_drvinfo(struct net_device *dev,
>>
>> }
>>
>> +static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
>> +{
>> + int i, cpu;
>> + switch (stringset) {
>> + case ETH_SS_STATS:
>> + for_each_possible_cpu(cpu)
>> + for (i = 0; i< VIRTNET_NUM_STATS; i++) {
>> + sprintf(buf, "%s[%u]",
>> + virtnet_stats_str_attr[i].string, cpu);
>> + buf += ETH_GSTRING_LEN;
> I would do
> ret = snprintf(buf, ETH_GSTRING_LEN, ...)
> BUG_ON(ret>= ETH_GSTRING_LEN);
> here to make it more robust.
Ok.
>> + }
>> + for (i = 0; i< VIRTNET_NUM_STATS; i++) {
>> + memcpy(buf, virtnet_stats_str_attr[i].string,
>> + ETH_GSTRING_LEN);
>> + buf += ETH_GSTRING_LEN;
>> + }
> So why not just memcpy the whole array there?
> memcpy(buf, virtnet_stats_str_attr,
> sizeof virtnet_stats_str_attr);
>
>> + break;
>> + }
>> +}
>> +
>> +static int virtnet_get_sset_count(struct net_device *dev, int sset)
>> +{
>> + switch (sset) {
>> + case ETH_SS_STATS:
> also add
> BUILD_BUG_ON(VIRTNET_NUM_STATS != (sizeof virtnet_stats_str_attr) / ETH_GSTRING_LEN);
>
Ok.
>> + return VIRTNET_NUM_STATS * (num_possible_cpus() + 1);
>> + default:
>> + return -EOPNOTSUPP;
>> + }
>> +}
>> +
>> +static void virtnet_get_ethtool_stats(struct net_device *dev,
>> + struct ethtool_stats *stats, u64 *buf)
>> +{
>> + struct virtnet_info *vi = netdev_priv(dev);
>> + int cpu, i;
>> + unsigned int start;
>> + struct virtnet_stats sample, total;
>> +
>> + memset(&total, 0, sizeof(total));
> sizeof total
> when operand is a variable,
> to distinguish from when it is a type.
Sure.
>> +
>> + for_each_possible_cpu(cpu) {
>> + struct virtnet_stats *s = per_cpu_ptr(vi->stats, cpu);
>> + do {
>> + start = u64_stats_fetch_begin(&s->syncp);
>> + memcpy(&sample.data,&s->data,
>> + sizeof(u64) * VIRTNET_NUM_STATS);
>> + } while (u64_stats_fetch_retry(&s->syncp, start));
>> +
>> + for (i = 0; i< VIRTNET_NUM_STATS; i++) {
>> + *buf = sample.data[i];
>> + total.data[i] += sample.data[i];
>> + buf++;
>> + }
>> + }
>> +
>> + memcpy(buf,&total.data, sizeof(u64) * VIRTNET_NUM_STATS);
>> +}
>> +
>> static const struct ethtool_ops virtnet_ethtool_ops = {
>> .get_drvinfo = virtnet_get_drvinfo,
>> .get_link = ethtool_op_get_link,
>> .get_ringparam = virtnet_get_ringparam,
>> + .get_ethtool_stats = virtnet_get_ethtool_stats,
>> + .get_strings = virtnet_get_strings,
>> + .get_sset_count = virtnet_get_sset_count,
>> };
>>
>> #define MIN_MTU 68
On Wed, 2012-06-06 at 15:52 +0800, Jason Wang wrote:
> Satistics counters is useful for debugging and performance optimization, so this
> patch lets virtio_net driver collect following and export them to userspace
> through "ethtool -S":
>
> - number of packets sent/received
> - number of bytes sent/received
> - number of callbacks for tx/rx
> - number of kick for tx/rx
> - number of bytes/packets queued for tx
>
> As virtnet_stats were per-cpu, so both per-cpu and gloabl satistics were
> collected like:
[...]
I would really like to see some sort of convention for presenting
per-queue statistics through ethtool. At the moment we have a complete
mess of different formats:
bnx2x: "[${index}]: ${name}"
be2net: "${qtype}q${index}: ${name}"
ehea: "PR${index} ${name}"
mlx4_en: "${qtype}${index}_${name}"
myri10ge: dummy stat names as headings
niu: dummy stat names as headings
s2io: "ring_${index}_${name}"
vmxnet3: dummy stat names as headings
vxge: "${name}_${index}"; also dummy stat names as headings
And you're introducing yet another format!
(Additionally some of the drivers are playing games with spaces and tabs
to make ethtool indent the stats the way they like. Ethtool statistics
are inconsistent enough already without drivers pulling that sort of
crap.
I'm inclined to make ethtool start stripping whitespace from stat names,
and *if* people can agree on a common format for per-queue statistic
names then I'll indent them *consistently*. Also, I would make such
stats optional, so you don't get hundreds of lines of crap by default.)
Ben.
--
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
From: Ben Hutchings <[email protected]>
Date: Thu, 7 Jun 2012 18:15:06 +0100
> I would really like to see some sort of convention for presenting
> per-queue statistics through ethtool. At the moment we have a complete
> mess of different formats:
Indeed. Probably ${QUEUE_TYPE}-${INDEX}-${STATISTIC} is best.
With an agreed upon list of queue types such as "rx", "tx", "rxtx"
etc.
On Thu, 2012-06-07 at 13:05 -0700, David Miller wrote:
> From: Ben Hutchings <[email protected]>
> Date: Thu, 7 Jun 2012 18:15:06 +0100
>
> > I would really like to see some sort of convention for presenting
> > per-queue statistics through ethtool. At the moment we have a complete
> > mess of different formats:
>
> Indeed. Probably ${QUEUE_TYPE}-${INDEX}-${STATISTIC} is best.
> With an agreed upon list of queue types such as "rx", "tx", "rxtx"
> etc.
I think we should leave the type names open-ended, as there are other
useful groupings like per-virtual-port. In that case the separator
should be chosen to allow arbitrary type names without ambiguity.
Ben.
--
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
On 06/07/2012 01:24 PM, Ben Hutchings wrote:
> On Thu, 2012-06-07 at 13:05 -0700, David Miller wrote:
>> From: Ben Hutchings<[email protected]>
>> Date: Thu, 7 Jun 2012 18:15:06 +0100
>>
>>> I would really like to see some sort of convention for presenting
>>> per-queue statistics through ethtool. At the moment we have a complete
>>> mess of different formats:
>>
>> Indeed. Probably ${QUEUE_TYPE}-${INDEX}-${STATISTIC} is best.
>> With an agreed upon list of queue types such as "rx", "tx", "rxtx"
>> etc.
>
> I think we should leave the type names open-ended, as there are other
> useful groupings like per-virtual-port. In that case the separator
> should be chosen to allow arbitrary type names without ambiguity.
So you mean like something along the lines of the presence of say '.'
indicating indent a level:
rx_bytes: 1234
myqueue1.rx_bytes: 234
myqueue2.rx_bytes: 345
...
rick jones
On Thu, 2012-06-07 at 13:39 -0700, Rick Jones wrote:
> On 06/07/2012 01:24 PM, Ben Hutchings wrote:
> > On Thu, 2012-06-07 at 13:05 -0700, David Miller wrote:
> >> From: Ben Hutchings<[email protected]>
> >> Date: Thu, 7 Jun 2012 18:15:06 +0100
> >>
> >>> I would really like to see some sort of convention for presenting
> >>> per-queue statistics through ethtool. At the moment we have a complete
> >>> mess of different formats:
> >>
> >> Indeed. Probably ${QUEUE_TYPE}-${INDEX}-${STATISTIC} is best.
> >> With an agreed upon list of queue types such as "rx", "tx", "rxtx"
> >> etc.
> >
> > I think we should leave the type names open-ended, as there are other
> > useful groupings like per-virtual-port. In that case the separator
> > should be chosen to allow arbitrary type names without ambiguity.
>
> So you mean like something along the lines of the presence of say '.'
> indicating indent a level:
>
> rx_bytes: 1234
> myqueue1.rx_bytes: 234
> myqueue2.rx_bytes: 345
> ...
Most drivers seem to want this sort of ordering/grouping:
group0.foo
group0.bar
...
group1.foo
group1.bar
...
but if we have a standard way of indicating groups of statistics then
the user can choose whether they want to reorder by type name.
Ben.
--
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
On Thu, 2012-06-07 at 21:56 +0100, Ben Hutchings wrote:
> On Thu, 2012-06-07 at 13:39 -0700, Rick Jones wrote:
> > On 06/07/2012 01:24 PM, Ben Hutchings wrote:
> > > On Thu, 2012-06-07 at 13:05 -0700, David Miller wrote:
> > >> From: Ben Hutchings<[email protected]>
> > >> Date: Thu, 7 Jun 2012 18:15:06 +0100
> > >>
> > >>> I would really like to see some sort of convention for presenting
> > >>> per-queue statistics through ethtool. At the moment we have a complete
> > >>> mess of different formats:
> > >>
> > >> Indeed. Probably ${QUEUE_TYPE}-${INDEX}-${STATISTIC} is best.
> > >> With an agreed upon list of queue types such as "rx", "tx", "rxtx"
> > >> etc.
> > >
> > > I think we should leave the type names open-ended, as there are other
> > > useful groupings like per-virtual-port. In that case the separator
> > > should be chosen to allow arbitrary type names without ambiguity.
> >
> > So you mean like something along the lines of the presence of say '.'
> > indicating indent a level:
> >
> > rx_bytes: 1234
> > myqueue1.rx_bytes: 234
> > myqueue2.rx_bytes: 345
> > ...
>
> Most drivers seem to want this sort of ordering/grouping:
>
> group0.foo
> group0.bar
> ...
> group1.foo
> group1.bar
> ...
>
> but if we have a standard way of indicating groups of statistics then
> the user can choose whether they want to reorder by type name.
I mean, whether they want to reorder/regroup by the final part of the
statistic name.
Ben.
--
Ben Hutchings, Staff Engineer, Solarflare
Not speaking for my employer; that's the marketing department's job.
They asked us to note that Solarflare product names are trademarked.
On Wed, Jun 06, 2012 at 03:52:17PM +0800, Jason Wang wrote:
> Satistics counters is useful for debugging and performance optimization, so this
> patch lets virtio_net driver collect following and export them to userspace
> through "ethtool -S":
>
> - number of packets sent/received
> - number of bytes sent/received
> - number of callbacks for tx/rx
> - number of kick for tx/rx
> - number of bytes/packets queued for tx
>
> As virtnet_stats were per-cpu, so both per-cpu and gloabl satistics were
> collected like:
>
> NIC statistics:
> tx_bytes[0]: 1731209929
> tx_packets[0]: 60685
> tx_kicks[0]: 63
> tx_callbacks[0]: 73
> tx_queued_bytes[0]: 1935749360
> tx_queued_packets[0]: 80652
> rx_bytes[0]: 2695648
> rx_packets[0]: 40767
> rx_kicks[0]: 1
> rx_callbacks[0]: 2077
> tx_bytes[1]: 9105588697
> tx_packets[1]: 344150
> tx_kicks[1]: 162
> tx_callbacks[1]: 905
> tx_queued_bytes[1]: 8901049412
> tx_queued_packets[1]: 324184
> rx_bytes[1]: 23679828
> rx_packets[1]: 358770
> rx_kicks[1]: 6
> rx_callbacks[1]: 17717
> tx_bytes: 10836798626
> tx_packets: 404835
> tx_kicks: 225
> tx_callbacks: 978
> tx_queued_bytes: 10836798772
> tx_queued_packets: 404836
> rx_bytes: 26375476
> rx_packets: 399537
> rx_kicks: 7
> rx_callbacks: 19794
>
> TODO:
>
> - more statistics
> - calculate the pending bytes/pkts
>
> Signed-off-by: Jason Wang <[email protected]>
>
> ---
> Changes from v1:
>
> - style & typo fixs
> - convert the statistics fields to array
> - use unlikely()
> ---
> drivers/net/virtio_net.c | 115 +++++++++++++++++++++++++++++++++++++++++++++-
> 1 files changed, 113 insertions(+), 2 deletions(-)
>
> diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
> index 6e4aa6f..909a0a7 100644
> --- a/drivers/net/virtio_net.c
> +++ b/drivers/net/virtio_net.c
> @@ -44,8 +44,14 @@ module_param(gso, bool, 0444);
> enum virtnet_stats_type {
> VIRTNET_TX_BYTES,
> VIRTNET_TX_PACKETS,
> + VIRTNET_TX_KICKS,
> + VIRTNET_TX_CBS,
> + VIRTNET_TX_Q_BYTES,
> + VIRTNET_TX_Q_PACKETS,
> VIRTNET_RX_BYTES,
> VIRTNET_RX_PACKETS,
> + VIRTNET_RX_KICKS,
> + VIRTNET_RX_CBS,
> VIRTNET_NUM_STATS,
> };
>
> @@ -54,6 +60,21 @@ struct virtnet_stats {
> u64 data[VIRTNET_NUM_STATS];
> };
>
> +static struct {
> + char string[ETH_GSTRING_LEN];
> +} virtnet_stats_str_attr[] = {
> + { "tx_bytes" },
> + { "tx_packets" },
> + { "tx_kicks" },
> + { "tx_callbacks" },
> + { "tx_queued_bytes" },
> + { "tx_queued_packets" },
> + { "rx_bytes" },
> + { "rx_packets" },
> + { "rx_kicks" },
> + { "rx_callbacks" },
> +};
> +
> struct virtnet_info {
> struct virtio_device *vdev;
> struct virtqueue *rvq, *svq, *cvq;
> @@ -146,6 +167,11 @@ static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
> static void skb_xmit_done(struct virtqueue *svq)
> {
> struct virtnet_info *vi = svq->vdev->priv;
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
> +
> + u64_stats_update_begin(&stats->syncp);
> + stats->data[VIRTNET_TX_CBS]++;
> + u64_stats_update_end(&stats->syncp);
>
> /* Suppress further interrupts. */
> virtqueue_disable_cb(svq);
> @@ -465,6 +491,7 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
> {
> int err;
> bool oom;
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>
> do {
> if (vi->mergeable_rx_bufs)
> @@ -481,13 +508,24 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
> } while (err > 0);
> if (unlikely(vi->num > vi->max))
> vi->max = vi->num;
> - virtqueue_kick(vi->rvq);
> + if (virtqueue_kick_prepare(vi->rvq)) {
> + virtqueue_notify(vi->rvq);
> + u64_stats_update_begin(&stats->syncp);
> + stats->data[VIRTNET_RX_KICKS]++;
> + u64_stats_update_end(&stats->syncp);
> + }
> return !oom;
> }
>
> static void skb_recv_done(struct virtqueue *rvq)
> {
> struct virtnet_info *vi = rvq->vdev->priv;
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
> +
> + u64_stats_update_begin(&stats->syncp);
> + stats->data[VIRTNET_RX_CBS]++;
> + u64_stats_update_end(&stats->syncp);
> +
> /* Schedule NAPI, Suppress further interrupts if successful. */
> if (napi_schedule_prep(&vi->napi)) {
> virtqueue_disable_cb(rvq);
> @@ -630,7 +668,9 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
> static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> {
> struct virtnet_info *vi = netdev_priv(dev);
> + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
> int capacity;
> + bool kick;
>
> /* Free up any pending old buffers before queueing new ones. */
> free_old_xmit_skbs(vi);
> @@ -655,7 +695,17 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> kfree_skb(skb);
> return NETDEV_TX_OK;
> }
> - virtqueue_kick(vi->svq);
> +
> + kick = virtqueue_kick_prepare(vi->svq);
> + if (unlikely(kick))
> + virtqueue_notify(vi->svq);
> +
> + u64_stats_update_begin(&stats->syncp);
> + if (unlikely(kick))
> + stats->data[VIRTNET_TX_KICKS]++;
> + stats->data[VIRTNET_TX_Q_BYTES] += skb->len;
> + stats->data[VIRTNET_TX_Q_PACKETS]++;
is this statistic interesting?
how about decrementing when we free?
this way we see how many are pending..
> + u64_stats_update_end(&stats->syncp);
>
> /* Don't wait up for transmitted skbs to be freed. */
> skb_orphan(skb);
> @@ -943,10 +993,71 @@ static void virtnet_get_drvinfo(struct net_device *dev,
>
> }
>
> +static void virtnet_get_strings(struct net_device *dev, u32 stringset, u8 *buf)
> +{
> + int i, cpu;
> + switch (stringset) {
> + case ETH_SS_STATS:
> + for_each_possible_cpu(cpu)
> + for (i = 0; i < VIRTNET_NUM_STATS; i++) {
> + sprintf(buf, "%s[%u]",
> + virtnet_stats_str_attr[i].string, cpu);
> + buf += ETH_GSTRING_LEN;
> + }
> + for (i = 0; i < VIRTNET_NUM_STATS; i++) {
> + memcpy(buf, virtnet_stats_str_attr[i].string,
> + ETH_GSTRING_LEN);
> + buf += ETH_GSTRING_LEN;
> + }
> + break;
> + }
> +}
> +
> +static int virtnet_get_sset_count(struct net_device *dev, int sset)
> +{
> + switch (sset) {
> + case ETH_SS_STATS:
> + return VIRTNET_NUM_STATS * (num_possible_cpus() + 1);
> + default:
> + return -EOPNOTSUPP;
> + }
> +}
> +
> +static void virtnet_get_ethtool_stats(struct net_device *dev,
> + struct ethtool_stats *stats, u64 *buf)
> +{
> + struct virtnet_info *vi = netdev_priv(dev);
> + int cpu, i;
> + unsigned int start;
> + struct virtnet_stats sample, total;
> +
> + memset(&total, 0, sizeof(total));
> +
> + for_each_possible_cpu(cpu) {
> + struct virtnet_stats *s = per_cpu_ptr(vi->stats, cpu);
> + do {
> + start = u64_stats_fetch_begin(&s->syncp);
> + memcpy(&sample.data, &s->data,
> + sizeof(u64) * VIRTNET_NUM_STATS);
> + } while (u64_stats_fetch_retry(&s->syncp, start));
> +
> + for (i = 0; i < VIRTNET_NUM_STATS; i++) {
> + *buf = sample.data[i];
> + total.data[i] += sample.data[i];
> + buf++;
> + }
> + }
> +
> + memcpy(buf, &total.data, sizeof(u64) * VIRTNET_NUM_STATS);
> +}
> +
> static const struct ethtool_ops virtnet_ethtool_ops = {
> .get_drvinfo = virtnet_get_drvinfo,
> .get_link = ethtool_op_get_link,
> .get_ringparam = virtnet_get_ringparam,
> + .get_ethtool_stats = virtnet_get_ethtool_stats,
> + .get_strings = virtnet_get_strings,
> + .get_sset_count = virtnet_get_sset_count,
> };
>
> #define MIN_MTU 68
On 06/08/2012 04:56 AM, Ben Hutchings wrote:
> On Thu, 2012-06-07 at 13:39 -0700, Rick Jones wrote:
>> On 06/07/2012 01:24 PM, Ben Hutchings wrote:
>>> On Thu, 2012-06-07 at 13:05 -0700, David Miller wrote:
>>>> From: Ben Hutchings<[email protected]>
>>>> Date: Thu, 7 Jun 2012 18:15:06 +0100
>>>>
>>>>> I would really like to see some sort of convention for presenting
>>>>> per-queue statistics through ethtool. At the moment we have a complete
>>>>> mess of different formats:
>>>> Indeed. Probably ${QUEUE_TYPE}-${INDEX}-${STATISTIC} is best.
>>>> With an agreed upon list of queue types such as "rx", "tx", "rxtx"
>>>> etc.
>>> I think we should leave the type names open-ended, as there are other
>>> useful groupings like per-virtual-port. In that case the separator
>>> should be chosen to allow arbitrary type names without ambiguity.
>> So you mean like something along the lines of the presence of say '.'
>> indicating indent a level:
>>
>> rx_bytes: 1234
>> myqueue1.rx_bytes: 234
>> myqueue2.rx_bytes: 345
>> ...
> Most drivers seem to want this sort of ordering/grouping:
>
> group0.foo
> group0.bar
> ...
> group1.foo
> group1.bar
> ...
>
> but if we have a standard way of indicating groups of statistics then
> the user can choose whether they want to reorder by type name.
>
> Ben.
>
Yes, it looks to me that the per-queue satistics were better:
- Simple and less synchronization.
- Good for future virtio-net multiqueue merging.
On 06/08/2012 06:19 AM, Michael S. Tsirkin wrote:
> On Wed, Jun 06, 2012 at 03:52:17PM +0800, Jason Wang wrote:
>> > Satistics counters is useful for debugging and performance optimization, so this
>> > patch lets virtio_net driver collect following and export them to userspace
>> > through "ethtool -S":
>> >
>> > - number of packets sent/received
>> > - number of bytes sent/received
>> > - number of callbacks for tx/rx
>> > - number of kick for tx/rx
>> > - number of bytes/packets queued for tx
>> >
>> > As virtnet_stats were per-cpu, so both per-cpu and gloabl satistics were
>> > collected like:
>> >
>> > NIC statistics:
>> > tx_bytes[0]: 1731209929
>> > tx_packets[0]: 60685
>> > tx_kicks[0]: 63
>> > tx_callbacks[0]: 73
>> > tx_queued_bytes[0]: 1935749360
>> > tx_queued_packets[0]: 80652
>> > rx_bytes[0]: 2695648
>> > rx_packets[0]: 40767
>> > rx_kicks[0]: 1
>> > rx_callbacks[0]: 2077
>> > tx_bytes[1]: 9105588697
>> > tx_packets[1]: 344150
>> > tx_kicks[1]: 162
>> > tx_callbacks[1]: 905
>> > tx_queued_bytes[1]: 8901049412
>> > tx_queued_packets[1]: 324184
>> > rx_bytes[1]: 23679828
>> > rx_packets[1]: 358770
>> > rx_kicks[1]: 6
>> > rx_callbacks[1]: 17717
>> > tx_bytes: 10836798626
>> > tx_packets: 404835
>> > tx_kicks: 225
>> > tx_callbacks: 978
>> > tx_queued_bytes: 10836798772
>> > tx_queued_packets: 404836
>> > rx_bytes: 26375476
>> > rx_packets: 399537
>> > rx_kicks: 7
>> > rx_callbacks: 19794
>> >
>> > TODO:
>> >
>> > - more statistics
>> > - calculate the pending bytes/pkts
>> >
>> > Signed-off-by: Jason Wang<[email protected]>
>> >
>> > ---
>> > Changes from v1:
>> >
>> > - style& typo fixs
>> > - convert the statistics fields to array
>> > - use unlikely()
>> > ---
>> > drivers/net/virtio_net.c | 115 +++++++++++++++++++++++++++++++++++++++++++++-
>> > 1 files changed, 113 insertions(+), 2 deletions(-)
>> >
>> > diff --git a/drivers/net/virtio_net.c b/drivers/net/virtio_net.c
>> > index 6e4aa6f..909a0a7 100644
>> > --- a/drivers/net/virtio_net.c
>> > +++ b/drivers/net/virtio_net.c
>> > @@ -44,8 +44,14 @@ module_param(gso, bool, 0444);
>> > enum virtnet_stats_type {
>> > VIRTNET_TX_BYTES,
>> > VIRTNET_TX_PACKETS,
>> > + VIRTNET_TX_KICKS,
>> > + VIRTNET_TX_CBS,
>> > + VIRTNET_TX_Q_BYTES,
>> > + VIRTNET_TX_Q_PACKETS,
>> > VIRTNET_RX_BYTES,
>> > VIRTNET_RX_PACKETS,
>> > + VIRTNET_RX_KICKS,
>> > + VIRTNET_RX_CBS,
>> > VIRTNET_NUM_STATS,
>> > };
>> >
>> > @@ -54,6 +60,21 @@ struct virtnet_stats {
>> > u64 data[VIRTNET_NUM_STATS];
>> > };
>> >
>> > +static struct {
>> > + char string[ETH_GSTRING_LEN];
>> > +} virtnet_stats_str_attr[] = {
>> > + { "tx_bytes" },
>> > + { "tx_packets" },
>> > + { "tx_kicks" },
>> > + { "tx_callbacks" },
>> > + { "tx_queued_bytes" },
>> > + { "tx_queued_packets" },
>> > + { "rx_bytes" },
>> > + { "rx_packets" },
>> > + { "rx_kicks" },
>> > + { "rx_callbacks" },
>> > +};
>> > +
>> > struct virtnet_info {
>> > struct virtio_device *vdev;
>> > struct virtqueue *rvq, *svq, *cvq;
>> > @@ -146,6 +167,11 @@ static struct page *get_a_page(struct virtnet_info *vi, gfp_t gfp_mask)
>> > static void skb_xmit_done(struct virtqueue *svq)
>> > {
>> > struct virtnet_info *vi = svq->vdev->priv;
>> > + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>> > +
>> > + u64_stats_update_begin(&stats->syncp);
>> > + stats->data[VIRTNET_TX_CBS]++;
>> > + u64_stats_update_end(&stats->syncp);
>> >
>> > /* Suppress further interrupts. */
>> > virtqueue_disable_cb(svq);
>> > @@ -465,6 +491,7 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
>> > {
>> > int err;
>> > bool oom;
>> > + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>> >
>> > do {
>> > if (vi->mergeable_rx_bufs)
>> > @@ -481,13 +508,24 @@ static bool try_fill_recv(struct virtnet_info *vi, gfp_t gfp)
>> > } while (err> 0);
>> > if (unlikely(vi->num> vi->max))
>> > vi->max = vi->num;
>> > - virtqueue_kick(vi->rvq);
>> > + if (virtqueue_kick_prepare(vi->rvq)) {
>> > + virtqueue_notify(vi->rvq);
>> > + u64_stats_update_begin(&stats->syncp);
>> > + stats->data[VIRTNET_RX_KICKS]++;
>> > + u64_stats_update_end(&stats->syncp);
>> > + }
>> > return !oom;
>> > }
>> >
>> > static void skb_recv_done(struct virtqueue *rvq)
>> > {
>> > struct virtnet_info *vi = rvq->vdev->priv;
>> > + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>> > +
>> > + u64_stats_update_begin(&stats->syncp);
>> > + stats->data[VIRTNET_RX_CBS]++;
>> > + u64_stats_update_end(&stats->syncp);
>> > +
>> > /* Schedule NAPI, Suppress further interrupts if successful. */
>> > if (napi_schedule_prep(&vi->napi)) {
>> > virtqueue_disable_cb(rvq);
>> > @@ -630,7 +668,9 @@ static int xmit_skb(struct virtnet_info *vi, struct sk_buff *skb)
>> > static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
>> > {
>> > struct virtnet_info *vi = netdev_priv(dev);
>> > + struct virtnet_stats *stats = this_cpu_ptr(vi->stats);
>> > int capacity;
>> > + bool kick;
>> >
>> > /* Free up any pending old buffers before queueing new ones. */
>> > free_old_xmit_skbs(vi);
>> > @@ -655,7 +695,17 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
>> > kfree_skb(skb);
>> > return NETDEV_TX_OK;
>> > }
>> > - virtqueue_kick(vi->svq);
>> > +
>> > + kick = virtqueue_kick_prepare(vi->svq);
>> > + if (unlikely(kick))
>> > + virtqueue_notify(vi->svq);
>> > +
>> > + u64_stats_update_begin(&stats->syncp);
>> > + if (unlikely(kick))
>> > + stats->data[VIRTNET_TX_KICKS]++;
>> > + stats->data[VIRTNET_TX_Q_BYTES] += skb->len;
>> > + stats->data[VIRTNET_TX_Q_PACKETS]++;
> is this statistic interesting?
> how about decrementing when we free?
> this way we see how many are pending..
>
Currently we didn't have per-vq statistics but per-cpu, so the skb could
be sent by one vcpu and freed by another.
Pehaps another reason to use per-queue satistics.
On Fri, Jun 08, 2012 at 11:35:25AM +0800, Jason Wang wrote:
> >>> @@ -655,7 +695,17 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> >>> kfree_skb(skb);
> >>> return NETDEV_TX_OK;
> >>> }
> >>> - virtqueue_kick(vi->svq);
> >>> +
> >>> + kick = virtqueue_kick_prepare(vi->svq);
> >>> + if (unlikely(kick))
> >>> + virtqueue_notify(vi->svq);
> >>> +
> >>> + u64_stats_update_begin(&stats->syncp);
> >>> + if (unlikely(kick))
> >>> + stats->data[VIRTNET_TX_KICKS]++;
> >>> + stats->data[VIRTNET_TX_Q_BYTES] += skb->len;
> >>> + stats->data[VIRTNET_TX_Q_PACKETS]++;
> >is this statistic interesting?
> >how about decrementing when we free?
> >this way we see how many are pending..
> >
>
> Currently we didn't have per-vq statistics but per-cpu, so the skb
> could be sent by one vcpu and freed by another.
> Pehaps another reason to use per-queue satistics.
For transmit, it could be done easily as we both send and free skbs
under a lock. I'm not sure how acceptable it is to
take a lock in get_stats but send a separate patch like this
and we'll see what others say.
--
MST
On Fri, Jun 08, 2012 at 11:35:25AM +0800, Jason Wang wrote:
> >>> @@ -655,7 +695,17 @@ static netdev_tx_t start_xmit(struct sk_buff *skb, struct net_device *dev)
> >>> kfree_skb(skb);
> >>> return NETDEV_TX_OK;
> >>> }
> >>> - virtqueue_kick(vi->svq);
> >>> +
> >>> + kick = virtqueue_kick_prepare(vi->svq);
> >>> + if (unlikely(kick))
> >>> + virtqueue_notify(vi->svq);
> >>> +
> >>> + u64_stats_update_begin(&stats->syncp);
> >>> + if (unlikely(kick))
> >>> + stats->data[VIRTNET_TX_KICKS]++;
> >>> + stats->data[VIRTNET_TX_Q_BYTES] += skb->len;
> >>> + stats->data[VIRTNET_TX_Q_PACKETS]++;
> >is this statistic interesting?
> >how about decrementing when we free?
> >this way we see how many are pending..
> >
>
> Currently we didn't have per-vq statistics but per-cpu, so the skb
> could be sent by one vcpu and freed by another.
> Pehaps another reason to use per-queue satistics.
Just to stress these things do not need to contradict:
you can have per cpu stats for each queue.
--
MST