fix(metrics): probe AMD L3 uncore and disable when unusable (fixes #660)

cmd/metrics/loader_legacy.go

@@ -184,6 +184,10 @@ func isCollectableEvent(event EventDefinition, metadata Metadata) bool {
 	// - their corresponding device is not found
 	// - not in system-wide collection scope
 	if event.Device != "cpu" && event.Device != "" {
+		if !metadata.SupportsUncore {
+			slog.Debug("Uncore events not supported on target", slog.String("event", event.Name))
+			return false
+		}
 		if flagScope == scopeProcess || flagScope == scopeCgroup {
 			slog.Debug("Uncore events not supported in process or cgroup scope", slog.String("event", event.Name))
 			return false

cmd/metrics/metadata.go

@@ -48,6 +48,7 @@ const (
 	scriptKernelVersion          = "kernel version"
 	scriptARMSlots               = "arm slots"
 	scriptARMCPUID               = "arm cpuid"
+	scriptPerfStatAMDUncoreProbe = "perf stat amd uncore probe"
 )
 
 // CommonMetadata -- common to all architectures
@@ -213,6 +214,13 @@ BEGIN {
 		Architectures:  []string{cpus.X86Architecture},
 		Depends:        []string{"perf"},
 	},
+	{
+		Name:           scriptPerfStatAMDUncoreProbe,
+		ScriptTemplate: `perf stat -a -e "l3/event=0x4,umask=0xff,enallcores=0x1,enallslices=0x1,threadmask=0x3,name='l3_lookup_state.all_coherent_accesses_to_l3'/" sleep 1`,
+		Architectures:  []string{cpus.X86Architecture},
+		Vendors:        []string{cpus.AMDVendor},
+		Depends:        []string{"perf"},
+	},
 	{
 		Name:           scriptPerfStatFixedInstr,
 		ScriptTemplate: "perf stat -a -e '{{{.InstructionsList}}}' sleep 1",

cmd/metrics/metadata_x86.go

@@ -149,6 +149,18 @@ func (c *X86MetadataCollector) CollectMetadata(t target.Target, noRoot bool, noS
 	}
 	metadata.SupportsUncore = c.checkUncoreSupport(metadata.UncoreDeviceIDs, isAMDArchitecture)
 
+	// On AMD, probe that L3 uncore actually works (e.g. GCP VMs expose l3 in sysfs but perf cannot use it).
+	// If the probe fails, disable uncore so collection uses core-only events and still produces metrics.
+	if isAMDArchitecture && metadata.SupportsUncore {
+		if output, ok := scriptOutputs[scriptPerfStatAMDUncoreProbe]; ok {
+			if !getSupportsAMDUncore(output) {
+				slog.Warn("AMD L3 uncore probe failed, disabling uncore metrics", slog.String("stderr", output.Stderr))
+				metadata.SupportsUncore = false
+				removeUncoreDevices(metadata.UncoreDeviceIDs, "l3", "df")
+			}
+		}
+	}
+
 	return metadata, nil
 }
 
@@ -229,6 +241,32 @@ func (c *X86MetadataCollector) checkUncoreSupport(uncoreDeviceIDs map[string][]i
 	return false
 }
 
+// getSupportsAMDUncore returns true if the AMD uncore probe script succeeded (L3 PMU is usable).
+// On some VMs (e.g. GCP AMD Turin), sysfs lists amd_l3 but perf cannot use it.
+func getSupportsAMDUncore(output script.ScriptOutput) bool {
+	if output.Exitcode != 0 {
+		return false
+	}
+	stderr := output.Stderr
+	if strings.Contains(stderr, "Unable to find PMU or event on a PMU of 'l3'") {
+		return false
+	}
+	if strings.Contains(stderr, "event syntax error") && strings.Contains(stderr, "l3") {
+		return false
+	}
+	if strings.Contains(stderr, "<not supported>") {
+		return false
+	}
+	return true
+}
+
+// removeUncoreDevices removes the given device names from the map (used when uncore probe fails).
+func removeUncoreDevices(uncoreDeviceIDs map[string][]int, deviceNames ...string) {
+	for _, name := range deviceNames {
+		delete(uncoreDeviceIDs, name)
+	}
+}
+
 // --- x86-specific helper functions ---
 
 // getUncoreDeviceIDs returns a map of device type to list of device indices.