Increase maximum frames and make maximum tail calls configurable

Author: umanwizard

internal/controller/controller.go

@@ -20,6 +20,8 @@ import (
 
 const MiB = 1 << 20
 
+const MaxTailCalls = 29
+
 // Controller is an instance that runs, manages and stops the agent.
 type Controller struct {
 	config   *Config
@@ -85,6 +87,7 @@ func (c *Controller) Start(ctx context.Context) error {
 		ProbabilisticInterval:  c.config.ProbabilisticInterval,
 		ProbabilisticThreshold: c.config.ProbabilisticThreshold,
 		OffCPUThreshold:        uint32(c.config.OffCPUThreshold),
+		MaxTailCalls:           MaxTailCalls,
 	})
 	if err != nil {
 		return fmt.Errorf("failed to load eBPF tracer: %w", err)

support/ebpf/tracemgmt.h

@@ -204,6 +204,10 @@ static inline PerCPURecord *get_per_cpu_record(void)
 static inline PerCPURecord *get_pristine_per_cpu_record()
 {
   PerCPURecord *record = get_per_cpu_record();
+  int key0 = 0;
+  SystemConfig *syscfg = bpf_map_lookup_elem(&system_config, &key0);
+  if (!syscfg)
+    return NULL;
   if (!record)
     return record;
 
@@ -235,7 +239,7 @@ static inline PerCPURecord *get_pristine_per_cpu_record()
   record->luajitUnwindState.cframe         = 0;
   record->luajitUnwindState.is_jit         = false;
   record->unwindersDone                    = 0;
-  record->tailCalls                        = 0;
+  record->tailCallsRemaining               = syscfg->max_tail_calls;
   record->ratelimitAction                  = RATELIMIT_ACTION_DEFAULT;
   record->customLabelsState.go_m_ptr       = NULL;
 
@@ -516,8 +520,8 @@ static inline __attribute__((__always_inline__)) void tail_call(void *ctx, int n
     return;
   }
 
-  if (record->tailCalls >= 29) {
-    // The maximum tail call count we need to support on older kernels is 32. At this point
+  if (!record->tailCallsRemaining) {
+    // At this point
     // there is a chance that continuing unwinding the stack would further increase the number of
     // tail calls. As a result we might lose the unwound stack as no further tail calls are left
     // to report it to user space. To make sure we do not run into this issue we stop unwinding
@@ -526,7 +530,7 @@ static inline __attribute__((__always_inline__)) void tail_call(void *ctx, int n
     record->state.unwind_error = ERR_MAX_TAIL_CALLS;
     increment_metric(metricID_MaxTailCalls);
   }
-  record->tailCalls += 1;
+  --record->tailCallsRemaining;
 
   bpf_tail_call(ctx, &perf_progs, next);
 }

support/ebpf/tracer.ebpf.debug.amd64

@@ -386,7 +386,7 @@ typedef enum TraceOrigin {
 // MAX_FRAME_UNWINDS defines the maximum number of frames per
 // Trace we can unwind and respect the limit of eBPF instructions,
 // limit of tail calls and limit of stack size per eBPF program.
-#define MAX_FRAME_UNWINDS 128
+#define MAX_FRAME_UNWINDS 256
 
 // MAX_NON_ERROR_FRAME_UNWINDS defines the maximum number of frames
 // to be pushed by unwinders while still leaving space for an error frame.
@@ -886,7 +886,7 @@ typedef struct PerCPURecord {
   u32 unwindersDone;
 
   // tailCalls tracks the number of calls to bpf_tail_call().
-  u8 tailCalls;
+  u8 tailCallsRemaining;
 
   // ratelimitAction determines the PID event rate limiting mode
   u8 ratelimitAction;
@@ -1051,6 +1051,9 @@ typedef struct SystemConfig {
 
   // Enables the temporary hack that drops pure errors frames in unwind_stop.
   bool drop_error_only_traces;
+
+  // Maximum number of eBPF tail calls before giving up.
+  u8 max_tail_calls;
 } SystemConfig;
 
 // Avoid including all of arch/arm64/include/uapi/asm/ptrace.h by copying the

support/ebpf/tracer.ebpf.debug.arm64

@@ -50,6 +50,7 @@ func StartTracer(ctx context.Context, t *testing.T, et tracertypes.IncludedTrace
 		SamplesPerSecond:       20,
 		ProbabilisticInterval:  100,
 		ProbabilisticThreshold: 100,
+		MaxTailCalls: 29,
 	})
 	require.NoError(t, err)
 

support/ebpf/tracer.ebpf.release.amd64

@@ -227,7 +227,8 @@ func determineStackLayout(coll *cebpf.CollectionSpec, maps map[string]*cebpf.Map
 
 func loadSystemConfig(coll *cebpf.CollectionSpec, maps map[string]*cebpf.Map,
 	kernelSymbols *libpf.SymbolMap, includeTracers types.IncludedTracers,
-	offCPUThreshold uint32, filterErrorFrames bool) error {
+	offCPUThreshold uint32, filterErrorFrames bool,
+	maxTailCalls uint8) error {
 	pacMask := pacmask.GetPACMask()
 	if pacMask != 0 {
 		log.Infof("Determined PAC mask to be 0x%016X", pacMask)
@@ -238,6 +239,7 @@ func loadSystemConfig(coll *cebpf.CollectionSpec, maps map[string]*cebpf.Map,
 		inverse_pac_mask:       ^C.u64(pacMask),
 		drop_error_only_traces: C.bool(filterErrorFrames),
 		off_cpu_threshold:      C.u32(offCPUThreshold),
+		max_tail_calls:         C.u8(maxTailCalls),
 	}
 
 	if err := parseBTF(&syscfg); err != nil {