Amend -off-cpu-threshold value

cli_flags.go

@@ -25,7 +25,7 @@ const (
 	defaultProbabilisticThreshold = tracer.ProbabilisticThresholdMax
 	defaultProbabilisticInterval  = 1 * time.Minute
 	defaultArgSendErrorFrames     = false
-	defaultOffCPUThreshold        = support.OffCPUThresholdMax
+	defaultOffCPUThreshold        = 0
 
 	// This is the X in 2^(n + x) where n is the default hardcoded map size value
 	defaultArgMapScaleFactor = 0
@@ -63,11 +63,10 @@ var (
 		"If zero, monotonic-realtime clock sync will be performed once, " +
 		"on agent startup, but not periodically."
 	sendErrorFramesHelp = "Send error frames (devfiler only, breaks Kibana)"
-	offCPUThresholdHelp = fmt.Sprintf("If set to a value between 1 and %d will enable "+
-		"off-cpu profiling: Every time an off-cpu entry point is hit, a random number between "+
-		"0 and %d is chosen. If the given threshold is greater than this random number, the "+
-		"off-cpu trace is collected and reported.",
-		support.OffCPUThresholdMax-1, support.OffCPUThresholdMax-1)
+	offCPUThresholdHelp = fmt.Sprintf("The per-mille chance for an off-cpu event being recorded. "+
+		"Valid values are in the range [1..%d], and 0 to disable off-cpu profiling."+
+		"Default is %d.",
+		support.OffCPUThresholdMax, defaultOffCPUThreshold)
 )
 
 // Package-scope variable, so that conditionally compiled other components can refer

internal/controller/config.go

@@ -8,7 +8,9 @@ import (
 	"time"
 
 	log "github.com/sirupsen/logrus"
+
 	"go.opentelemetry.io/ebpf-profiler/reporter"
+	"go.opentelemetry.io/ebpf-profiler/support"
 	"go.opentelemetry.io/ebpf-profiler/tracer"
 )
 
@@ -89,6 +91,14 @@ func (cfg *Config) Validate() error {
 		)
 	}
 
+	if cfg.OffCPUThreshold > support.OffCPUThresholdMax {
+		return fmt.Errorf(
+			"invalid argument for off-cpu-threshold. Value "+
+				"should be between 1 and %d, or 0 to disable off-cpu profiling",
+			support.OffCPUThresholdMax,
+		)
+	}
+
 	if !cfg.NoKernelVersionCheck {
 		major, minor, patch, err := tracer.GetCurrentKernelVersion()
 		if err != nil {

internal/controller/controller.go

@@ -11,7 +11,6 @@ import (
 	"go.opentelemetry.io/ebpf-profiler/host"
 	"go.opentelemetry.io/ebpf-profiler/metrics"
 	"go.opentelemetry.io/ebpf-profiler/reporter"
-	"go.opentelemetry.io/ebpf-profiler/support"
 	"go.opentelemetry.io/ebpf-profiler/times"
 	"go.opentelemetry.io/ebpf-profiler/tracehandler"
 	"go.opentelemetry.io/ebpf-profiler/tracer"
@@ -108,7 +107,7 @@ func (c *Controller) Start(ctx context.Context) error {
 	}
 	log.Info("Attached tracer program")
 
-	if c.config.OffCPUThreshold < support.OffCPUThresholdMax {
+	if c.config.OffCPUThreshold > 0 {
 		if err := trc.StartOffCPUProfiling(); err != nil {
 			return fmt.Errorf("failed to start off-cpu profiling: %v", err)
 		}

support/ebpf/off_cpu.ebpf.c

@@ -37,7 +37,7 @@ int tracepoint__sched_switch(void *ctx)
     return ERR_UNREACHABLE;
   }
 
-  if (bpf_get_prandom_u32() % OFF_CPU_THRESHOLD_MAX > syscfg->off_cpu_threshold) {
+  if (bpf_get_prandom_u32() % OFF_CPU_THRESHOLD_MAX >= syscfg->off_cpu_threshold) {
     return 0;
   }
 

tracer/tracer.go

@@ -490,7 +490,7 @@ func initializeMapsAndPrograms(kernelSymbols *libpf.SymbolMap, cfg *Config) (
 		return nil, nil, fmt.Errorf("failed to load perf eBPF programs: %v", err)
 	}
 
-	if cfg.OffCPUThreshold < support.OffCPUThresholdMax {
+	if cfg.OffCPUThreshold > 0 {
 		if err = loadKProbeUnwinders(coll, ebpfProgs, ebpfMaps["kprobe_progs"], tailCallProgs,
 			cfg.BPFVerifierLogLevel, ebpfMaps["perf_progs"].FD()); err != nil {
 			return nil, nil, fmt.Errorf("failed to load kprobe eBPF programs: %v", err)
@@ -554,23 +554,22 @@ func loadAllMaps(coll *cebpf.CollectionSpec, cfg *Config,
 	// On modern systems /proc/sys/kernel/pid_max defaults to 4194304.
 	// Try to fit this PID space scaled down with cfg.OffCPUThreshold into
 	// this map.
-	adaption["sched_times"] = (4194304 / support.OffCPUThresholdMax) * cfg.OffCPUThreshold
+	adaption["sched_times"] = (4194304 * cfg.OffCPUThreshold) / support.OffCPUThresholdMax
 
 	for i := support.StackDeltaBucketSmallest; i <= support.StackDeltaBucketLargest; i++ {
 		mapName := fmt.Sprintf("exe_id_to_%d_stack_deltas", i)
 		adaption[mapName] = 1 << uint32(exeIDToStackDeltasSize+cfg.MapScaleFactor)
 	}
 
 	for mapName, mapSpec := range coll.Maps {
+		if mapName == "sched_times" && cfg.OffCPUThreshold == 0 {
+			// Off CPU Profiling is disabled. So do not load this map.
+			continue
+		}
 		if newSize, ok := adaption[mapName]; ok {
 			log.Debugf("Size of eBPF map %s: %v", mapName, newSize)
 			mapSpec.MaxEntries = newSize
 		}
-		if mapName == "sched_times" &&
-			cfg.OffCPUThreshold >= support.OffCPUThresholdMax {
-			// Off CPU Profiling is not enabled. So do not load this map.
-			continue
-		}
 		ebpfMap, err := cebpf.NewMap(mapSpec)
 		if err != nil {
 			return fmt.Errorf("failed to load %s: %v", mapName, err)