feat: add support for CPython 3.13

We add support for CPython 3.13

.github/workflows/tests.yml

@@ -132,7 +132,7 @@ jobs:
       - name: Run functional Austin tests (with sudo)
         run: |
           ulimit -c unlimited
-          echo "core.%p" | sudo tee /proc/sys/kernel/core_pattern
+          sudo echo "core.%p" | sudo tee /proc/sys/kernel/core_pattern
           sudo -E env PATH="$PATH" .venv/bin/pytest --pastebin=failed -svr a test/functional -k "not austinp"
         if: always()
 

src/py_proc.c

@@ -235,6 +235,30 @@ _py_proc__infer_python_version(py_proc_t * self) {
 
   int major = 0, minor = 0, patch = 0;
 
+  // Starting with Python 3.13 we can use the PyRuntime structure
+  if (isvalid(self->symbols[DYNSYM_RUNTIME])) {
+    _Py_DebugOffsets py_d;
+    if (fail(py_proc__get_type(self, self->symbols[DYNSYM_RUNTIME], py_d))) {
+      log_e("Cannot copy PyRuntimeState structure from remote address");
+      FAIL;
+    }
+
+    if (0 == memcmp(py_d.v3_13.cookie, "xdebugpy", sizeof(py_d.v3_13.cookie))) {
+      uint64_t version = py_d.v3_13.version;
+      major = (version>>24) & 0xFF;
+      minor = (version>>16) & 0xFF;
+      patch = (version>>8)  & 0xFF;
+
+      log_d("Python version (from debug offsets): %d.%d.%d", major, minor, patch);
+
+      self->py_v = get_version_descriptor(major, minor, patch);
+
+      init_version_descriptor(self->py_v, &py_d);
+
+      SUCCESS;
+    }
+  }
+
   // Starting with Python 3.11 we can rely on the Py_Version symbol
   if (isvalid(self->symbols[DYNSYM_HEX_VERSION])) {
     unsigned long py_version = 0;
@@ -322,17 +346,14 @@ _py_proc__check_interp_state(py_proc_t * self, void * raddr) {
 
   V_DESC(self->py_v);
 
-  PyInterpreterState is;
-  PyThreadState      tstate_head;
-
-  if (py_proc__get_type(self, raddr, is)) {
+  if (py_proc__copy_v(self, is, raddr, self->is)) {
     log_ie("Cannot get remote interpreter state");
     FAIL;
   }
+  log_d("Interpreter state buffer %p", self->is);
+  void * tstate_head_addr = V_FIELD_PTR(void *, self->is, py_is, o_tstate_head);
 
-  void * tstate_head_addr = V_FIELD(void *, is, py_is, o_tstate_head);
-
-  if (fail(py_proc__get_type(self, tstate_head_addr, tstate_head))) {
+  if (fail(py_proc__copy_v(self, thread, tstate_head_addr, self->ts))) {
     log_e(
       "Cannot copy PyThreadState head at %p from PyInterpreterState instance",
       tstate_head_addr
@@ -342,7 +363,7 @@ _py_proc__check_interp_state(py_proc_t * self, void * raddr) {
 
   log_t("PyThreadState head loaded @ %p", V_FIELD(void *, is, py_is, o_tstate_head));
 
-  if (V_FIELD(void*, tstate_head, py_thread, o_interp) != raddr) {
+  if (V_FIELD_PTR(void*, self->ts, py_thread, o_interp) != raddr) {
     log_d("PyThreadState head does not point to interpreter state");
     set_error(EPROC);
     FAIL;
@@ -358,7 +379,7 @@ _py_proc__check_interp_state(py_proc_t * self, void * raddr) {
     raddr, V_FIELD(void *, is, py_is, o_tstate_head)
   );
 
-  raddr_t thread_raddr = {self->proc_ref, V_FIELD(void *, is, py_is, o_tstate_head)};
+  raddr_t thread_raddr = {self->proc_ref, V_FIELD_PTR(void *, self->is, py_is, o_tstate_head)};
   py_thread_t thread;
 
   if (fail(py_thread__fill_from_raddr(&thread, &thread_raddr, self))) {
@@ -560,6 +581,45 @@ _py_proc__get_current_thread_state_raddr(py_proc_t * self) {
   return (void *) -1;
 }
 
+// ----------------------------------------------------------------------------
+static void
+_py_proc__free_local_buffers(py_proc_t * self) {
+  sfree(self->is);
+  sfree(self->ts);
+}
+
+// ----------------------------------------------------------------------------
+static int
+_py_proc__init_local_buffers(py_proc_t * self) {
+  if (!isvalid(self)) {
+    set_error(EPROC);
+    FAIL;
+  }
+
+  self->is = calloc(1, self->py_v->py_is.size);
+  if (!isvalid(self->is)) {
+    log_e("Cannot allocate memory for PyInterpreterState");
+    goto error;
+  }
+
+  self->ts = calloc(1, self->py_v->py_thread.size);
+  if (!isvalid(self->ts)) {
+    log_e("Cannot allocate memory for PyThreadState");
+    goto error;
+  }
+
+  log_d("Local buffers initialised");
+
+  SUCCESS;
+
+error:
+  set_error(ENOMEM);
+
+  _py_proc__free_local_buffers(self);
+
+  FAIL;
+}
+
 // ----------------------------------------------------------------------------
 static int
 _py_proc__find_interpreter_state(py_proc_t * self) {
@@ -575,6 +635,9 @@ _py_proc__find_interpreter_state(py_proc_t * self) {
   if (fail(_py_proc__infer_python_version(self)))
     FAIL;
 
+  if (fail(_py_proc__init_local_buffers(self)))
+    FAIL;
+
   if (self->sym_loaded || isvalid(self->map.runtime.base)) {
     // Try to resolve the symbols or the runtime section, if we have them
 
@@ -703,6 +766,10 @@ py_proc_new(int child) {
 
   py_proc->child = child;
   py_proc->gc_state_raddr = NULL;
+  py_proc->py_v = NULL;
+
+  py_proc->is = NULL;
+  py_proc->ts = NULL;
 
   _prehash_symbols();
 
@@ -1200,15 +1267,13 @@ py_proc__sample(py_proc_t * self) {
 
   V_DESC(self->py_v);
 
-  PyInterpreterState is;
-
   do {
-    if (fail(py_proc__get_type(self, current_interp, is))) {
+    if (fail(py_proc__copy_v(self, is, current_interp, self->is))) {
       log_ie("Failed to get interpreter state while sampling");
       FAIL;
     }
 
-    void * tstate_head = V_FIELD(void *, is, py_is, o_tstate_head);
+    void * tstate_head = V_FIELD_PTR(void *, self->is, py_is, o_tstate_head);
     if (!isvalid(tstate_head))
       // Maybe the interpreter state is in an invalid state. We'll try again
       // unless there is a fatal error.
@@ -1223,7 +1288,7 @@ py_proc__sample(py_proc_t * self) {
     time_delta = gettime() - self->timestamp;
     #endif
 
-    int result = _py_proc__sample_interpreter(self, &is, time_delta);
+    int result = _py_proc__sample_interpreter(self, self->is, time_delta);
 
     #ifdef NATIVE
     if (fail(_py_proc__resume_threads(self, &raddr))) {
@@ -1234,7 +1299,7 @@ py_proc__sample(py_proc_t * self) {
 
     if (fail(result))
       FAIL;
-  } while (isvalid(current_interp = V_FIELD(void *, is, py_is, o_next)));
+  } while (isvalid(current_interp = V_FIELD_PTR(void *, self->is, py_is, o_next)));
 
   #ifdef NATIVE
   self->timestamp = gettime();
@@ -1321,6 +1386,8 @@ py_proc__destroy(py_proc_t * self) {
   hash_table__destroy(self->base_table);
   #endif
 
+  _py_proc__free_local_buffers(self);
+
   sfree(self->bin_path);
   sfree(self->lib_path);
   sfree(self->extra);

src/py_proc.h

@@ -99,6 +99,10 @@ typedef struct {
   hash_table_t    * base_table;
   #endif
 
+  // Local buffers
+  PyInterpreterState * is;
+  PyThreadState      * ts;
+
   // Platform-dependent fields
   proc_extra_info * extra;
 } py_proc_t;
@@ -208,6 +212,19 @@ py_proc__sample(py_proc_t *);
  */
 #define py_proc__get_type(self, raddr, dt) (py_proc__memcpy(self, raddr, sizeof(dt), &dt))
 
+/**
+ * Make a local copy of a remote structure.
+ *
+ * @param self  the process object.
+ * @param type  the type of the structure.
+ * @param raddr the remote address of the structure.
+ * @param dest  the destination address.
+ *
+ * @return 0 on success.
+ */
+#define py_proc__copy_v(self, type, raddr, dest) (py_proc__memcpy(self, raddr, py_v->py_##type.size, dest))
+
+
 /**
  * Log the Python interpreter version
  * @param self  the process object.

src/py_proc_list.c

@@ -176,7 +176,7 @@ py_proc_list__sample(py_proc_list_t * self) {
   for (py_proc_item_t * item = self->first; item != NULL; /* item = item->next */) {
     log_t("Sampling process with PID %d", item->py_proc->pid);
     stopwatch_start();
-    if (fail(py_proc__sample(item->py_proc))) {
+    if (!isvalid(item->py_proc->py_v) || fail(py_proc__sample(item->py_proc))) {
       py_proc__wait(item->py_proc);
       py_proc_item_t * next = item->next;
       _py_proc_list__remove(self, item);

src/py_thread.c

@@ -424,7 +424,7 @@ static inline int
 _py_thread__unwind_iframe_stack(py_thread_t * self, void * iframe_raddr) {
   int invalid = FALSE; 
   void * curr = iframe_raddr;
-  
+
   while (isvalid(curr)) {
     if (fail(_py_thread__push_iframe(self, &curr))) {
       log_d("Failed to retrieve iframe #%d", stack_pointer());
@@ -457,8 +457,6 @@ static inline int
 _py_thread__unwind_cframe_stack(py_thread_t * self) {
   PyCFrame cframe;
 
-  int invalid = FALSE;
-
   _py_thread__read_stack(self);
 
   stack_reset();
@@ -470,11 +468,7 @@ _py_thread__unwind_cframe_stack(py_thread_t * self) {
     FAIL;
   }
 
-  invalid = fail(_py_thread__unwind_iframe_stack(self, V_FIELD(void *, cframe, py_cframe, o_current_frame)));
-  if (invalid)
-    return invalid;
-
-  return invalid;
+  return fail(_py_thread__unwind_iframe_stack(self, V_FIELD(void *, cframe, py_cframe, o_current_frame)));
 }
 
 
@@ -954,7 +948,13 @@ py_thread__emit_collapsed_stack(py_thread_t * self, int64_t interp_id, ctime_t t
   V_DESC(self->proc->py_v);
 
   if (isvalid(self->top_frame)) {
-    if (V_MIN(3, 11)) {
+    if (V_MIN(3, 13)) {
+      if (fail(_py_thread__unwind_iframe_stack(self, self->top_frame))) {
+        emit_invalid_frame();
+        error = TRUE;
+      }
+    }  
+    else if (V_MIN(3, 11)) {
       if (fail(_py_thread__unwind_cframe_stack(self))) {
         emit_invalid_frame();
         error = TRUE;

src/python/runtime.h

@@ -156,4 +156,109 @@ typedef union {
   _PyRuntimeState3_12 v3_12;
 } _PyRuntimeState;
 
+
+// Starting with CPython 3.13, we can retrieve the offsets from a dedicated
+// data structure
+
+typedef struct _Py_DebugOffsets3_13 {
+    char cookie[8];  // xdebugpy
+    uint64_t version;
+    // Runtime state offset;
+    struct _runtime_state {
+        uint64_t size;
+        uint64_t finalizing;
+        uint64_t interpreters_head;
+    } runtime_state;
+
+    // Interpreter state offset;
+    struct _interpreter_state {
+        uint64_t size;
+        uint64_t id;
+        uint64_t next;
+        uint64_t threads_head;
+        uint64_t gc;
+        uint64_t imports_modules;
+        uint64_t sysdict;
+        uint64_t builtins;
+        uint64_t ceval_gil;
+        uint64_t gil_runtime_state_locked;
+        uint64_t gil_runtime_state_holder;
+    } interpreter_state;
+
+    // Thread state offset;
+    struct _thread_state{
+        uint64_t size;
+        uint64_t prev;
+        uint64_t next;
+        uint64_t interp;
+        uint64_t current_frame;
+        uint64_t thread_id;
+        uint64_t native_thread_id;
+        uint64_t datastack_chunk;
+        uint64_t status;
+    } thread_state;
+
+    // InterpreterFrame offset;
+    struct _interpreter_frame {
+        uint64_t size;
+        uint64_t previous;
+        uint64_t executable;
+        uint64_t instr_ptr;
+        uint64_t localsplus;
+        uint64_t owner;
+    } interpreter_frame;
+
+    // Code object offset;
+    struct _code_object {
+        uint64_t size;
+        uint64_t filename;
+        uint64_t name;
+        uint64_t qualname;
+        uint64_t linetable;
+        uint64_t firstlineno;
+        uint64_t argcount;
+        uint64_t localsplusnames;
+        uint64_t localspluskinds;
+        uint64_t co_code_adaptive;
+    } code_object;
+
+    // PyObject offset;
+    struct _pyobject {
+        uint64_t size;
+        uint64_t ob_type;
+    } pyobject;
+
+    // PyTypeObject object offset;
+    struct _type_object {
+        uint64_t size;
+        uint64_t tp_name;
+    } type_object;
+
+    // PyTuple object offset;
+    struct _tuple_object {
+        uint64_t size;
+        uint64_t ob_item;
+    } tuple_object;
+
+    // Unicode object offset;
+    struct _unicode_object {
+        uint64_t size;
+        uint64_t state;
+        uint64_t length;
+        size_t asciiobject_size;
+    } unicode_object;
+
+    // GC runtime state offset;
+    struct _gc {
+        uint64_t size;
+        uint64_t collecting;
+    } gc;
+} _Py_DebugOffsets3_13;
+
+
+typedef union {
+  _Py_DebugOffsets3_13 v3_13;
+} _Py_DebugOffsets;
+
+
 #endif