feat(metrics/net): add tap tx latency metric

This metric measures how long Firecracker VMM thread
is blocked on the write syscalls when accessing a tap device.
By looking at it, we will be able to see what portion
of tx net latency is attributed to factors external
to Firecracker.

(cherry picked from commit f7d8a33)

Signed-off-by: Nikita Kalyazin <[email protected]>

Author: kalyazin

src/vmm/src/devices/virtio/net/device.rs

@@ -490,6 +490,7 @@ impl Net {
             });
         }
 
+        let _metric = net_metrics.tap_write_agg.record_latency_metrics();
         match Self::write_tap(tap, frame_iovec) {
             Ok(_) => {
                 let len = frame_iovec.len() as u64;

src/vmm/src/devices/virtio/net/metrics.rs

@@ -85,7 +85,7 @@ use std::sync::{Arc, RwLock};
 use serde::ser::SerializeMap;
 use serde::{Serialize, Serializer};
 
-use crate::logger::{IncMetric, SharedIncMetric};
+use crate::logger::{IncMetric, LatencyAggregateMetrics, SharedIncMetric};
 
 /// map of network interface id and metrics
 /// this should be protected by a lock before accessing.
@@ -107,7 +107,7 @@ impl NetMetricsPerDevice {
                 .write()
                 .unwrap()
                 .metrics
-                .insert(iface_id.clone(), Arc::new(NetDeviceMetrics::default()));
+                .insert(iface_id.clone(), Arc::new(NetDeviceMetrics::new()));
         }
         METRICS
             .read()
@@ -184,6 +184,8 @@ pub struct NetDeviceMetrics {
     pub tap_read_fails: SharedIncMetric,
     /// Number of times writing to TAP failed.
     pub tap_write_fails: SharedIncMetric,
+    /// Duration of all tap write operations.
+    pub tap_write_agg: LatencyAggregateMetrics,
     /// Number of transmitted bytes.
     pub tx_bytes_count: SharedIncMetric,
     /// Number of malformed TX frames.
@@ -207,6 +209,14 @@ pub struct NetDeviceMetrics {
 }
 
 impl NetDeviceMetrics {
+    /// Const default construction.
+    pub fn new() -> Self {
+        Self {
+            tap_write_agg: LatencyAggregateMetrics::new(),
+            ..Default::default()
+        }
+    }
+
     /// Net metrics are SharedIncMetric where the diff of current vs
     /// old is serialized i.e. serialize_u64(current-old).
     /// So to have the aggregate serialized in same way we need to
@@ -239,6 +249,9 @@ impl NetDeviceMetrics {
         self.rx_count.add(other.rx_count.fetch_diff());
         self.tap_read_fails.add(other.tap_read_fails.fetch_diff());
         self.tap_write_fails.add(other.tap_write_fails.fetch_diff());
+        self.tap_write_agg
+            .sum_us
+            .add(other.tap_write_agg.sum_us.fetch_diff());
         self.tx_bytes_count.add(other.tx_bytes_count.fetch_diff());
         self.tx_malformed_frames
             .add(other.tx_malformed_frames.fetch_diff());

tests/host_tools/fcmetrics.py

@@ -240,6 +240,9 @@ def validate_fc_metrics(metrics):
             "read_agg",
             "write_agg",
         ],
+        "net": [
+            "tap_write_agg",
+        ],
     }
 
     # validate timestamp before jsonschema validation which some more time