add windows to CI testing and fix race condition

.github/workflows/tests.yml

@@ -20,16 +20,18 @@ jobs:
         os: [ubuntu-latest]
         python: ["3.9", "3.10", "3.11", "3.12", "3.13", "3.14", "3.15.0-rc.2", "pypy-3.11"]
         experimental: [false]
-          # include:
-          #   - os: macos-latest
-          #     python: "3.10"
-          #     experimental: false
-          #   - os: windows-latest
-          #     python: "3.10"
-          #     experimental: false
+        include:
+          # Windows on the earliest and latest supported versions:
+          - os: windows-latest
+            python: "3.9"
+            experimental: false
+          - os: windows-latest
+            python: "3.15.0-rc.2"
+            experimental: false
     steps:
     - uses: actions/checkout@v7
     - name: Install libunwind
+      if: runner.os == 'Linux'
       run: |
         sudo apt install -y libunwind-dev
         pkg-config libunwind --cflags --libs-only-l

dev_requirements.txt

@@ -1,5 +1,4 @@
 six
 requests
 backports.shutil_which
-pytz
 colorama

pyproject.toml

@@ -16,7 +16,6 @@ requires-python = ">=3.9,<3.16"
 dependencies = [
     "requests",
     "six",
-    "pytz",
     "colorama",
 ]
 classifiers = [

src/vmprof_win.c

@@ -8,6 +8,8 @@
 
 volatile int thread_started = 0;
 volatile int enabled = 0;
+/* Set by the sampler thread while it takes and writes one sample. */
+static volatile LONG sampler_busy = 0;
 #ifndef RPYTHON_VMPROF
 static PY_WIN_THREAD_STATE *target_tstate = NULL;
 #endif
@@ -16,11 +18,25 @@ HANDLE write_mutex;
 
 int prepare_concurrent_bufs(void)
 {
-    if (!(write_mutex = CreateMutex(NULL, FALSE, NULL)))
-        return -1;
+    if (write_mutex == NULL) {
+        if (!(write_mutex = CreateMutex(NULL, FALSE, NULL)))
+            return -1;
+    }
     return 0;
 }
 
+/* Wait for the sampler thread to finish the sample it may be in the middle
+   of. Callers clear `enabled` first so that no new sample starts; without
+   this wait a stale sample could be written into the next profile, ahead of
+   its header. */
+static void wait_for_sampler(void)
+{
+    MemoryBarrier();
+    while (sampler_busy) {
+        SwitchToThread();
+    }
+}
+
 int vmprof_register_virtual_function(char *code_name, intptr_t code_uid,
                                      int auto_retry)
 {
@@ -126,8 +142,9 @@ int vmprof_snapshot_thread(DWORD thread_id, PY_WIN_THREAD_STATE *tstate, prof_st
 #else
         frame = PyThreadState_GetFrame(tstate);
 #endif
+        /* leave room for the thread id appended below */
         depth = vmp_walk_and_record_stack(frame, stack->stack,
-                                          MAX_STACK_DEPTH, 0, 0);
+                                          MAX_STACK_DEPTH - 1, 0, 0);
 #ifdef _MSC_VER
     } __except (EXCEPTION_EXECUTE_HANDLER) {
         depth = -1;
@@ -193,19 +210,25 @@ long __stdcall vmprof_mainloop(void *arg)
     // cpython version
     while (1) {
         Sleep(vmprof_get_profile_interval_usec() * 1000);
+        sampler_busy = 1;
+        MemoryBarrier();
         if (!enabled) {
+            sampler_busy = 0;
             continue;
         }
         tstate = get_current_thread_state();
         if (!tstate)
             tstate = target_tstate;
-        if (!tstate)
+        if (!tstate) {
+            sampler_busy = 0;
             continue;
+        }
         depth = vmprof_snapshot_thread(tstate->thread_id, tstate, stack);
         if (depth > 0) {
             vmp_write_all((char*)stack + offsetof(prof_stacktrace_s, marker),
                           SIZEOF_PROF_STACKTRACE + depth * sizeof(void*));
         }
+        sampler_busy = 0;
     }
 #else
     // pypy version
@@ -255,10 +278,9 @@ int vmprof_enable(int memory, int native, int real_time)
 RPY_EXTERN
 int vmprof_disable(void)
 {
-    char marker = MARKER_TRAILER;
-    (void)vmp_write_time_now(MARKER_TRAILER);
-
     enabled = 0;
+    wait_for_sampler();
+    (void)vmp_write_time_now(MARKER_TRAILER);
 #ifndef RPYTHON_VMPROF
     target_tstate = NULL;
 #endif
@@ -270,6 +292,9 @@ RPY_EXTERN
 void vmprof_ignore_signals(int ignored)
 {
     enabled = !ignored;
+    if (ignored) {
+        wait_for_sampler();
+    }
 }
 
 int vmp_native_enable(void)

test_requirements.txt

@@ -7,6 +7,5 @@ setuptools>=77
 pytest
 hypothesis
 colorama
-pytz
 attrs
 py

vmprof/reader.py

@@ -114,15 +114,16 @@ def detect_file_sizes(self):
             little = False
             self.setup_once(little_endian=little, word_size=4, addr_size=4)
         else:
-            firstbytes = self.read(8)
-            if firstbytes[0] == three:
+            secondbytes = self.read(8)
+            if secondbytes[0] == three:
                 little = True
                 self.setup_once(little_endian=little, word_size=8, addr_size=8)
-            elif firstbytes[7] == three:
+            elif secondbytes[7] == three:
                 little = False
                 self.setup_once(little_endian=little, word_size=8, addr_size=8)
             else:
-                raise NotImplementedError("could not determine word and addr size")
+                raise NotImplementedError("could not determine word and addr size, "
+                                          "file starts with %r" % (firstbytes + secondbytes,))
 
         # determine if it is windows 64 bit
         # even though it migt be a 64bit log, teh addr_size is now 4
@@ -243,7 +244,7 @@ def read_timeval(self):
 
     def read_timezone(self):
         timezone = self.read(8).strip(b'\x00')
-        # we should use pytz and parse iso8601 if we really support time zones
+        # we should use zoneinfo and parse iso8601 if we really support time zones
         return None
 
     def read_all(self):

vmprof/test/test_run.py

@@ -7,7 +7,6 @@
 import time
 import gzip
 import time
-import pytz
 import vmprof
 import six
 from cffi import FFI

vmshare/binary.py

@@ -1,7 +1,7 @@
 import sys
 import struct
 import array
-import pytz
+import zoneinfo
 
 WORD_SIZE = struct.calcsize('L')
 if sys.maxsize == 2**63-1:
@@ -82,7 +82,10 @@ def read_timezone(fileobj):
     timezone = fileobj.read(8).strip(b'\x00')
     timezone = timezone.decode('ascii')
     if timezone:
-        return pytz.timezone(timezone)
+        try:
+            return zoneinfo.ZoneInfo(timezone)
+        except (zoneinfo.ZoneInfoNotFoundError, ValueError):
+            return None
     return None
 
 def encode_le_u16(value):