diff --git a/src/PerfCounters.cc b/src/PerfCounters.cc
index 3884ce9e21d..fce76630e76 100644
--- a/src/PerfCounters.cc
+++ b/src/PerfCounters.cc
@@ -761,8 +761,6 @@ PTData PerfCounters::extract_pt_data() {
   PTData result;
   if (pt_state) {
     result = std::move(pt_state->pt_data);
-    pt_state->stop();
-    start_pt(tid, *pt_state);
   }
   return result;
 }
diff --git a/src/ProcessorTraceDecoder.cc b/src/ProcessorTraceDecoder.cc
index e6290139711..4692d4ef0b3 100644
--- a/src/ProcessorTraceDecoder.cc
+++ b/src/ProcessorTraceDecoder.cc
@@ -36,16 +36,41 @@ int read_mem_callback(uint8_t *buffer, size_t size,
       read_mem(ip, buffer, size);
 }
 
-void ProcessorTraceDecoder::init() {
+static constexpr uint8_t injected_header_packets[] =
+    { /*PSB*/ 0x02, 0x82, 0x02, 0x82, 0x02, 0x82, 0x02, 0x82,
+              0x02, 0x82, 0x02, 0x82, 0x02, 0x82, 0x02, 0x82,
+      /*CBR 31*/ 0x02, 0x03, 0x1f, 0x00,
+      /*PSBEND*/ 0x02, 0x23 };
+static constexpr size_t PSB_LEN = 16;
+
+void ProcessorTraceDecoder::init(const PTData& trace_data) {
   if (trace_data.data.empty()) {
     return;
   }
 
+  if (trace_data.data.size() < PSB_LEN ||
+      memcmp(injected_header_packets, trace_data.data.data(), PSB_LEN)) {
+    full_trace_data.reserve(sizeof(injected_header_packets) + trace_data.data.size());
+    full_trace_data.insert(full_trace_data.end(), injected_header_packets,
+                           injected_header_packets + sizeof(injected_header_packets));
+  }
+  full_trace_data.insert(full_trace_data.end(), trace_data.data.begin(),
+                         trace_data.data.end());
+
+  init_decoder();
+}
+
+void ProcessorTraceDecoder::dump_full_trace_data_to_file() {
+  ScopedFd out("/tmp/ptdata", O_WRONLY | O_CREAT | O_TRUNC, 0700);
+  write_all(out, full_trace_data.data(), full_trace_data.size());
+}
+
+void ProcessorTraceDecoder::init_decoder() {
   pt_config config;
   memset(&config, 0, sizeof(config));
   config.size = sizeof(config);
-  config.begin = const_cast<uint8_t*>(trace_data.data.data());
-  config.end = const_cast<uint8_t*>(trace_data.data.data() + trace_data.data.size());
+  config.begin = const_cast<uint8_t*>(full_trace_data.data());
+  config.end = const_cast<uint8_t*>(full_trace_data.data() + full_trace_data.size());
 
   decoder = pt_insn_alloc_decoder(&config);
   if (!decoder) {
@@ -61,6 +86,12 @@ void ProcessorTraceDecoder::init() {
   need_sync = true;
 }
 
+ProcessorTraceDecoder::ProcessorTraceDecoder(Task* task, std::vector<uint8_t> full_trace_data,
+                                             Mode mode)
+  : task(task), full_trace_data(full_trace_data), mode(mode), need_sync(false) {
+  init_decoder();
+}
+
 static vector<uint8_t>* cached_rr_page_for_recording[2];
 
 static const vector<uint8_t>& rr_page_for_recording(SupportedArch arch) {
@@ -120,7 +151,7 @@ void ProcessorTraceDecoder::maybe_process_events(int status) {
 }
 
 string ProcessorTraceDecoder::internal_error_context_string() {
-  ProcessorTraceDecoder helper(task, trace_data, mode);
+  ProcessorTraceDecoder helper(task, full_trace_data, mode);
   vector<Instruction> instructions;
   while (true) {
     int pt_status = 0;
diff --git a/src/ProcessorTraceDecoder.h b/src/ProcessorTraceDecoder.h
index f93b0cd428a..784c7437501 100644
--- a/src/ProcessorTraceDecoder.h
+++ b/src/ProcessorTraceDecoder.h
@@ -22,9 +22,8 @@ class ProcessorTraceDecoder {
   };
 
   ProcessorTraceDecoder(Task* t, const PTData& trace_data, Mode mode)
-    : task(t), trace_data(trace_data), decoder(nullptr), mode(mode),
-      need_sync(false) {
-    init();
+    : task(t), decoder(nullptr), mode(mode), need_sync(false) {
+    init(trace_data);
   }
   ~ProcessorTraceDecoder();
 
@@ -32,33 +31,39 @@ class ProcessorTraceDecoder {
     remote_code_ptr address;
   };
 
-  void init()
+  bool next_instruction(Instruction* out, int* pt_status = nullptr)
 #ifdef INTEL_PT_DECODING
     ;
 #else
   {
     FATAL() << "Intel PT support not built";
+    return false;
   }
 #endif
 
-  bool next_instruction(Instruction* out, int* pt_status = nullptr)
+  int read_mem(uint64_t ip, uint8_t *buffer, size_t size);
+
+private:
 #ifdef INTEL_PT_DECODING
-    ;
+  void init(const PTData& trace_data);
 #else
+  void init(const PTData&)
   {
     FATAL() << "Intel PT support not built";
-    return false;
   }
 #endif
+  void init_decoder();
 
-  int read_mem(uint64_t ip, uint8_t *buffer, size_t size);
+  void dump_full_trace_data_to_file();
+
+  ProcessorTraceDecoder(Task* task, std::vector<uint8_t> full_trace_data,
+                        Mode mode);
 
-private:
   void maybe_process_events(int status);
   std::string internal_error_context_string();
 
   Task* const task;
-  const PTData& trace_data;
+  std::vector<uint8_t> full_trace_data;
   pt_insn_decoder* decoder;
   Mode mode;
   bool need_sync;