numalib: provide a fallback for topology scanning on linux

.changelog/19457.txt

@@ -0,0 +1,3 @@
+```release-note:bug
+client: Fixed a bug where clients are unable to detect CPU topology in certain conditions
+```

client/lib/numalib/detect_default.go

@@ -5,15 +5,6 @@
 
 package numalib
 
-import (
-	"context"
-	"time"
-
-	"github.com/hashicorp/nomad/client/lib/idset"
-	"github.com/hashicorp/nomad/client/lib/numalib/hw"
-	"github.com/shirou/gopsutil/v3/cpu"
-)
-
 // PlatformScanners returns the set of SystemScanner for systems without a
 // specific implementation.
 func PlatformScanners() []SystemScanner {
@@ -22,40 +13,10 @@ func PlatformScanners() []SystemScanner {
 	}
 }
 
-const (
-	nodeID   = hw.NodeID(0)
-	socketID = hw.SocketID(0)
-	maxSpeed = hw.KHz(0)
-)
-
 // Generic implements SystemScanner as a fallback for operating systems without
 // a specific implementation.
 type Generic struct{}
 
 func (g *Generic) ScanSystem(top *Topology) {
-	// hardware may or may not be NUMA, but for now we only
-	// detect such topology on linux systems
-	top.NodeIDs = idset.Empty[hw.NodeID]()
-	top.NodeIDs.Insert(nodeID)
-
-	// cores
-	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
-	defer cancel()
-
-	count, err := cpu.CountsWithContext(ctx, true)
-	if err != nil {
-		return
-	}
-	top.Cores = make([]Core, count)
-
-	infos, err := cpu.InfoWithContext(ctx)
-	if err != nil || len(infos) == 0 {
-		return
-	}
-
-	for i := 0; i < count; i++ {
-		info := infos[0]
-		speed := hw.KHz(hw.MHz(info.Mhz) * 1000)
-		top.insert(nodeID, socketID, hw.CoreID(i), Performance, maxSpeed, speed)
-	}
+	scanGeneric(top)
 }

client/lib/numalib/detect_generic.go

@@ -0,0 +1,47 @@
+// Copyright (c) HashiCorp, Inc.
+// SPDX-License-Identifier: BUSL-1.1
+
+package numalib
+
+import (
+	"context"
+	"time"
+
+	"github.com/hashicorp/nomad/client/lib/idset"
+	"github.com/hashicorp/nomad/client/lib/numalib/hw"
+	"github.com/shirou/gopsutil/v3/cpu"
+)
+
+const (
+	genericNodeID   = hw.NodeID(0)
+	genericSocketID = hw.SocketID(0)
+	genericMaxSpeed = hw.KHz(0)
+)
+
+func scanGeneric(top *Topology) {
+	// hardware may or may not be NUMA, but for now we only
+	// detect such topology on linux systems
+	top.NodeIDs = idset.Empty[hw.NodeID]()
+	top.NodeIDs.Insert(genericNodeID)
+
+	// cores
+	ctx, cancel := context.WithTimeout(context.Background(), 10*time.Second)
+	defer cancel()
+
+	count, err := cpu.CountsWithContext(ctx, true)
+	if err != nil {
+		return
+	}
+	top.Cores = make([]Core, count)
+
+	infos, err := cpu.InfoWithContext(ctx)
+	if err != nil || len(infos) == 0 {
+		return
+	}
+
+	for i := 0; i < count; i++ {
+		info := infos[0]
+		speed := hw.KHz(hw.MHz(info.Mhz) * 1000)
+		top.insert(genericNodeID, genericSocketID, hw.CoreID(i), Performance, genericMaxSpeed, speed)
+	}
+}

client/lib/numalib/detect_linux.go

@@ -23,6 +23,7 @@ func PlatformScanners() []SystemScanner {
 		new(Smbios),
 		new(Cgroups1),
 		new(Cgroups2),
+		new(Fallback),
 	}
 }
 
@@ -209,3 +210,40 @@ func scanIDs(top *Topology, content string) {
 		}
 	}
 }
+
+// Fallback detects if the NUMA aware topology scanning was unable to construct
+// a valid model of the system. This will be common on Nomad clients running in
+// containers, erroneous hypervisors, or without root.
+type Fallback struct{}
+
+func (s *Fallback) ScanSystem(top *Topology) {
+	broken := false
+
+	switch {
+	case top.NodeIDs.Empty():
+		broken = true
+	case len(top.Distances) == 0:
+		broken = true
+	case top.NumCores() <= 0:
+		broken = true
+	case top.TotalCompute() <= 0:
+		broken = true
+	case top.UsableCompute() <= 0:
+		broken = true
+	case top.UsableCores().Empty():
+		broken = true
+	}
+
+	if !broken {
+		return
+	}
+
+	// we have a broken topology; reset it and fallback to the generic scanner
+	// basically treating this client like a windows / unsupported OS
+	top.NodeIDs = nil
+	top.Distances = nil
+	top.Cores = nil
+
+	// invoke the generic scanner
+	scanGeneric(top)
+}