Add prctl monitor

Author: nsavoire

internal/controller/controller.go

@@ -162,6 +162,11 @@ func (c *Controller) Start(ctx context.Context) error {
 	// So if you change this log line update also the system test.
 	log.Info("Attached sched monitor")
 
+	if err := trc.AttachPrctlMonitor(); err != nil {
+		return fmt.Errorf("failed to attach prctl monitor: %w", err)
+	}
+	log.Info("Attached prctl monitor")
+
 	if err := c.startTraceHandling(ctx, trc); err != nil {
 		return fmt.Errorf("failed to start trace handling: %w", err)
 	}

metrics/ids.go

@@ -2096,5 +2096,12 @@
     "name": "UnwindErrBadDTVRead",
     "field": "bpf.errors.bad_dtv_read",
     "id": 286
+  },
+  {
+    "description": "Number of prctl(PR_SET_VMA) calls naming an anonymous mapping OTEL_CTX",
+    "type": "counter",
+    "name": "NumPrctlSetVmaOtelCtx",
+    "field": "bpf.num_prctl_set_vma_otel_ctx",
+    "id": 287
   }
 ]

metrics/metrics.json

@@ -64,8 +64,16 @@ func Test_ProcessContext(t *testing.T) {
 	tests := map[string]struct {
 		exeName string
 		args    []string
+		env     []string
 	}{
 		"glibc_exe": {exeName: "processctx_exe_glibc"},
+		// Publishes the process context after a delay, so the profiler discovers
+		// the PID before the publication and the prctl monitor must trigger a
+		// resync to pick up the OTEL_CTX mapping.
+		"glibc_exe_delayed_publish": {
+			exeName: "processctx_exe_glibc",
+			env:     []string{"OTEL_PROCESS_CTX_PUBLISH_DELAY_MS=200"},
+		},
 		// "musl_exe":     {exeName: "processctx_exe_musl"},
 		// "glibc_lib":    {exeName: "processctx_lib_glibc"},
 		// "musl_lib":     {exeName: "processctx_lib_musl"},
@@ -107,6 +115,9 @@ func Test_ProcessContext(t *testing.T) {
 
 			cmd := exec.CommandContext(ctx, filepath.Join(exeDir, tc.exeName), tc.args...)
 			cmd.Stderr = os.Stderr
+			if len(tc.env) > 0 {
+				cmd.Env = append(os.Environ(), tc.env...)
+			}
 			require.NoError(t, cmd.Start())
 
 			wg := sync.WaitGroup{}

processcontext/integrationtests/processcontext_integration_test.go

@@ -6,6 +6,7 @@
 #include <stdbool.h>
 #include <stdint.h>
 #include <stdio.h>
+#include <stdlib.h>
 #include <unistd.h>
 
 #include "processctx_lib.h"
@@ -82,6 +83,19 @@ int main(int argc, char *argv[]) {
 
   signal(SIGTERM, handle_sigterm);
 
+  // OTEL_PROCESS_CTX_PUBLISH_DELAY_MS lets the test exercise the prctl monitor
+  // path: the profiler discovers the PID first, then the publish prctl fires
+  // and triggers a resync. We burn CPU rather than sleep so the process stays
+  // on-CPU and is sampled by the profiler's perf event, which drives process
+  // synchronization before the context is published.
+  const char *delay_str = getenv("OTEL_PROCESS_CTX_PUBLISH_DELAY_MS");
+  if (delay_str != NULL) {
+    int delay_ms = atoi(delay_str);
+    if (delay_ms > 0) {
+      burn(delay_ms);
+    }
+  }
+
   if (init_process_context()) {
     fprintf(stderr, "Failed to initialize process context\n");
     return 1;

processcontext/integrationtests/testdata/processctx.c

@@ -545,10 +545,8 @@ func (pm *ProcessManager) SynchronizeProcess(pr process.Process) {
 	numParseErrors, err := pr.IterateMappings(func(m process.RawMapping) bool {
 		if processcontext.IsContextMapping(m.IsExecutable(), m.Path) {
 			contextMappingAddr = m.Vaddr
-			// Even if process context is not found, it might be published in the future.
-			// For now, we rely on a new call to synchronizeMappings to pick it up.
-			// TODO: Add some kind of polling mechanism or a hook on prctl to be notified
-			// when the process context is published.
+			// The eBPF hook on prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME) will trigger a
+			// PID resynchronization when the process names its context mapping "OTEL_CTX".
 		}
 
 		// Executable mappings and VDSO, converted directly to libpf.FrameMapping

processmanager/processinfo.go

@@ -0,0 +1,65 @@
+// This file contains the code for the tracepoint on prctl(PR_SET_VMA, PR_SET_VMA_ANON_NAME, ...)
+// to detect when a process names an anonymous memory mapping "OTEL_CTX".
+
+#include "bpfdefs.h"
+#include "tracemgmt.h"
+
+#include "types.h"
+
+#ifndef TESTING_COREDUMP
+
+  // prctl constants from include/uapi/linux/prctl.h
+  #define PR_SET_VMA           0x53564d41
+  #define PR_SET_VMA_ANON_NAME 0
+
+// See /sys/kernel/tracing/events/syscalls/sys_enter_prctl/format
+struct sys_enter_prctl_ctx {
+  unsigned char skip[16]; // common fields (8) + __syscall_nr (4) + pad (4)
+  unsigned long option;   // prctl option
+  unsigned long arg2;     // sub-option (PR_SET_VMA_ANON_NAME for PR_SET_VMA)
+  unsigned long arg3;     // addr
+  unsigned long arg4;     // len
+  unsigned long arg5;     // name (user-space pointer)
+};
+
+// tracepoint__sys_enter_prctl hooks prctl() calls to detect when a process
+// names an anonymous VMA "OTEL_CTX". This triggers a PID resynchronization
+// so the profiler can discover the newly published process context mapping.
+SEC("tracepoint/syscalls/sys_enter_prctl")
+int tracepoint__sys_enter_prctl(struct sys_enter_prctl_ctx *ctx)
+{
+  if (ctx->option != PR_SET_VMA || ctx->arg2 != PR_SET_VMA_ANON_NAME) {
+    goto exit;
+  }
+
+  u64 pid_tgid = bpf_get_current_pid_tgid();
+  u32 pid      = pid_tgid >> 32;
+
+  if (!bpf_map_lookup_elem(&reported_pids, &pid) && !pid_information_exists(pid)) {
+    // Only report PIDs that we explicitly track. This avoids sending kernel worker PIDs
+    // to userspace.
+    goto exit;
+  }
+
+  // Read the VMA name from user-space. We only need 9 bytes ("OTEL_CTX" + NUL).
+  __attribute__((aligned(8))) char name[9] = {};
+  if (bpf_probe_read_user(name, sizeof(name), (void *)ctx->arg5)) {
+    goto exit;
+  }
+
+  // Check for an exact "OTEL_CTX" match. We avoid bpf_strncmp (kernel 5.17+).
+  // Instead, compare as a u64 for the 8 characters plus a byte check for the
+  // NUL terminator.
+  if (*(u64 *)name != *(u64 *)"OTEL_CTX" || name[8] != '\0') {
+    goto exit;
+  }
+
+  if (report_pid(ctx, pid_tgid, RATELIMIT_ACTION_DEFAULT)) {
+    increment_metric(metricID_NumPrctlSetVmaOtelCtx);
+  }
+
+exit:
+  return 0;
+}
+
+#endif

support/ebpf/prctl_monitor.ebpf.c

@@ -328,6 +328,9 @@ enum {
   // number of failures to read TLS variables via the DTV
   metricID_UnwindErrBadDTVRead,
 
+  // number of prctl(PR_SET_VMA) calls naming an anonymous mapping OTEL_CTX
+  metricID_NumPrctlSetVmaOtelCtx,
+
   //
   // Metric IDs above are for counters (cumulative values)
   //