Elevate Windows production worthiness

- SQLite file locking now works on Windows
- SQLite will now use fdatasync() on non-Apple platforms
- Fix Ctrl-C handler on Windows to not crash with TLS
- Signals now work in multithreaded apps on Windows
- fcntl() will now accurately report EINVAL errors
- fcntl() now has excellent --strace logging
- Token bucket replenish now go 100x faster
- *NSYNC cancellations now work on Windows
- Support closefrom() on NetBSD

libc/runtime/clone.c

@@ -43,6 +43,7 @@
 #include "libc/thread/openbsd.internal.h"
 #include "libc/thread/thread.h"
 #include "libc/thread/tls.h"
+#include "libc/thread/tls2.h"
 #include "libc/thread/xnu.internal.h"
 
 #define __NR_thr_new                      455
@@ -97,11 +98,7 @@ WinThreadEntry(int rdi,                                 // rcx
                int rdx,                                 // r8
                struct CloneArgs *wt) {                  // r9
   int rc;
-  if (wt->tls) {
-    asm("mov\t%1,%%gs:%0"
-        : "=m"(*((long *)0x1480 + __tls_index))
-        : "r"(wt->tls));
-  }
+  if (wt->tls) __set_tls_win32(wt->tls);
   *wt->ptid = wt->tid;
   *wt->ctid = wt->tid;
   rc = WinThreadLaunch(wt->arg, wt->tid, wt->func, (intptr_t)wt & -16);

libc/runtime/cocmd.c

@@ -48,7 +48,7 @@
 #define STATE_QUOTED_VAR 4
 #define STATE_WHITESPACE 5
 
-#define READ24LE(s) READ32LE(s "\0")
+#define READ24(s) READ32LE(s "\0")
 
 struct Env {
   char *s;
@@ -333,16 +333,16 @@ static int Test(void) {
   if (n == 4) {
     w = READ32LE(args[2]) & 0x00ffffff;
     if ((w & 65535) == READ16LE("=")) return !!strcmp(args[1], args[3]);
-    if (w == READ24LE("==")) return !!strcmp(args[1], args[3]);
-    if (w == READ24LE("!=")) return !strcmp(args[1], args[3]);
+    if (w == READ24("==")) return !!strcmp(args[1], args[3]);
+    if (w == READ24("!=")) return !strcmp(args[1], args[3]);
   } else if (n == 3) {
     w = READ32LE(args[1]) & 0x00ffffff;
-    if (w == READ24LE("-n")) return !(strlen(args[2]) > 0);
-    if (w == READ24LE("-z")) return !(strlen(args[2]) == 0);
-    if (w == READ24LE("-e")) return !!stat(args[2], &st);
-    if (w == READ24LE("-f")) return !stat(args[2], &st) && S_ISREG(st.st_mode);
-    if (w == READ24LE("-d")) return !stat(args[2], &st) && S_ISDIR(st.st_mode);
-    if (w == READ24LE("-h")) return !stat(args[2], &st) && S_ISLNK(st.st_mode);
+    if (w == READ24("-n")) return !(strlen(args[2]) > 0);
+    if (w == READ24("-z")) return !(strlen(args[2]) == 0);
+    if (w == READ24("-e")) return !!stat(args[2], &st);
+    if (w == READ24("-f")) return !(!stat(args[2], &st) && S_ISREG(st.st_mode));
+    if (w == READ24("-d")) return !(!stat(args[2], &st) && S_ISDIR(st.st_mode));
+    if (w == READ24("-h")) return !(!stat(args[2], &st) && S_ISLNK(st.st_mode));
   }
   return 1;
 }

libc/runtime/enable_threads.c

@@ -19,17 +19,15 @@
 #include "libc/assert.h"
 #include "libc/calls/calls.h"
 #include "libc/intrin/strace.internal.h"
-#include "libc/thread/tls.h"
 #include "libc/runtime/runtime.h"
+#include "libc/thread/tls.h"
 
 extern int __threadcalls_end[];
 extern int __threadcalls_start[];
 #pragma weak __threadcalls_start
 #pragma weak __threadcalls_end
 
-privileged void __enable_threads(void) {
-  if (__threaded) return;
-  STRACE("__enable_threads()");
+static privileged dontinline void FixupLocks(void) {
   __morph_begin();
   /*
    * _NOPL("__threadcalls", func)
@@ -55,5 +53,11 @@ privileged void __enable_threads(void) {
     _base[*p + 2] = 0xe8;
   }
   __morph_end();
+}
+
+void __enable_threads(void) {
+  if (__threaded) return;
+  STRACE("__enable_threads()");
+  FixupLocks();
   __threaded = gettid();
 }

libc/runtime/enable_tls.c

@@ -23,6 +23,7 @@
 #include "libc/errno.h"
 #include "libc/intrin/asan.internal.h"
 #include "libc/intrin/asancodes.h"
+#include "libc/intrin/atomic.h"
 #include "libc/intrin/bits.h"
 #include "libc/intrin/weaken.h"
 #include "libc/log/libfatal.internal.h"
@@ -99,6 +100,7 @@ _Alignas(TLS_ALIGNMENT) static char __static_tls[5008];
  * and your `errno` variable also won't be thread safe anymore.
  */
 privileged void __enable_tls(void) {
+  int tid;
   size_t siz;
   struct CosmoTib *tib;
   char *mem, *tls;
@@ -133,12 +135,13 @@ privileged void __enable_tls(void) {
   if (IsLinux()) {
     // gnu/systemd guarantees pid==tid for the main thread so we can
     // avoid issuing a superfluous system call at startup in program
-    tib->tib_tid = __pid;
+    tid = __pid;
   } else {
-    tib->tib_tid = sys_gettid();
+    tid = sys_gettid();
   }
+  atomic_store_explicit(&tib->tib_tid, tid, memory_order_relaxed);
   _pthread_main.tib = tib;
-  _pthread_main.tid = tib->tib_tid;
+  _pthread_main.tid = tid;
   _pthread_main.flags = PT_MAINTHREAD;
   __repmovsb(tls, _tdata_start, _TLDZ);
 

libc/runtime/fork.c

@@ -23,6 +23,7 @@
 #include "libc/calls/syscall-sysv.internal.h"
 #include "libc/calls/syscall_support-sysv.internal.h"
 #include "libc/dce.h"
+#include "libc/intrin/atomic.h"
 #include "libc/intrin/strace.internal.h"
 #include "libc/nt/process.h"
 #include "libc/runtime/internal.h"
@@ -56,7 +57,9 @@ int _fork(uint32_t dwCreationFlags) {
     parent = __pid;
     __pid = dx;
     if (__tls_enabled) {
-      __get_tls()->tib_tid = IsLinux() ? dx : sys_gettid();
+      atomic_store_explicit(&__get_tls()->tib_tid,
+                            IsLinux() ? dx : sys_gettid(),
+                            memory_order_relaxed);
     }
     if (!IsWindows()) sys_sigprocmask(SIG_SETMASK, &old, 0);
     STRACE("fork() → 0 (child of %d)", parent);