Make POSIX threads improvements

- Ensure SIGTHR isn't blocked in newly created threads
- Use TIB rather than thread_local for thread atexits
- Make POSIX thread keys atomic within thread
- Don't bother logging prctl() to --strace
- Log thread destructor names to --strace

Author: jart

libc/calls/prctl.c

@@ -53,15 +53,5 @@ int prctl(int operation, ...) {
     rc = enosys();
   }
 
-#if SYSDEBUG
-  if (operation == PR_CAPBSET_READ || operation == PR_CAPBSET_DROP) {
-    STRACE("prctl(%s, %s) → %d% m", DescribePrctlOperation(operation),
-           DescribeCapability(a), rc);
-  } else {
-    STRACE("prctl(%s, %p, %p, %p, %p) → %d% m",
-           DescribePrctlOperation(operation), a, b, c, d, rc);
-  }
-#endif
-
   return rc;
 }

libc/intrin/describeflags.internal.h

@@ -13,7 +13,6 @@ const char *DescribeFlags(char *, size_t, const struct DescribeFlags *, size_t,
 
 const char *DescribeArchPrctlCode(char[12], int) libcesque;
 const char *DescribeCancelState(char[12], int, int *) libcesque;
-const char *DescribeCapability(char[32], int) libcesque;
 const char *DescribeClockName(char[32], int) libcesque;
 const char *DescribeControlKeyState(char[64], uint32_t) libcesque;
 const char *DescribeDirfd(char[12], int) libcesque;
@@ -47,7 +46,6 @@ const char *DescribeOpenFlags(char[128], int) libcesque;
 const char *DescribeOpenMode(char[15], int, int) libcesque;
 const char *DescribePersonalityFlags(char[128], int) libcesque;
 const char *DescribePollFlags(char[64], int) libcesque;
-const char *DescribePrctlOperation(int) libcesque;
 const char *DescribeProtFlags(char[48], int) libcesque;
 const char *DescribePtrace(char[12], int) libcesque;
 const char *DescribePtraceEvent(char[32], int) libcesque;
@@ -69,9 +67,7 @@ const char *DescribeVirtualKeyCode(char[32], uint32_t) libcesque;
 const char *DescribeWhence(char[12], int) libcesque;
 const char *DescribeWhichPrio(char[12], int) libcesque;
 
-#define DescribeArchPrctlCode(x)     DescribeArchPrctlCode(alloca(12), x)
 #define DescribeCancelState(x, y)    DescribeCancelState(alloca(12), x, y)
-#define DescribeCapability(x)        DescribeCapability(alloca(32), x)
 #define DescribeClockName(x)         DescribeClockName(alloca(32), x)
 #define DescribeControlKeyState(x)   DescribeControlKeyState(alloca(64), x)
 #define DescribeDirfd(x)             DescribeDirfd(alloca(12), x)

libc/runtime/cxa_thread_atexit.c

@@ -17,6 +17,7 @@
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/intrin/cxaatexit.internal.h"
+#include "libc/intrin/strace.internal.h"
 #include "libc/intrin/weaken.h"
 #include "libc/mem/mem.h"
 #include "libc/nexgen32e/gc.internal.h"
@@ -30,8 +31,6 @@ struct Dtor {
   struct Dtor *next;
 };
 
-static _Thread_local struct Dtor *__cxa_thread_atexit_list;
-
 static void _pthread_unwind(struct CosmoTib *tib) {
   struct PosixThread *pt;
   struct _pthread_cleanup_buffer *cb;
@@ -42,34 +41,33 @@ static void _pthread_unwind(struct CosmoTib *tib) {
   }
 }
 
-static void __cxa_thread_unkey(struct CosmoTib *tib) {
-  void *val;
-  int i, j, gotsome;
-  pthread_key_dtor dtor;
-  for (j = 0; j < PTHREAD_DESTRUCTOR_ITERATIONS; ++j) {
-    for (gotsome = i = 0; i < PTHREAD_KEYS_MAX; ++i) {
-      if ((val = tib->tib_keys[i]) &&
-          (dtor = atomic_load_explicit(_pthread_key_dtor + i,
-                                       memory_order_relaxed)) &&
-          dtor != (pthread_key_dtor)-1) {
-        gotsome = 1;
-        tib->tib_keys[i] = 0;
+static void _pthread_unkey(struct CosmoTib *tib) {
+  for (int j = 0; j < PTHREAD_DESTRUCTOR_ITERATIONS; ++j) {
+    bool gotsome = false;
+    for (int i = 0; i < PTHREAD_KEYS_MAX; ++i) {
+      void *val;
+      pthread_key_dtor dtor;
+      if ((dtor = atomic_load_explicit(&_pthread_key_dtor[i],
+                                       memory_order_acquire)) &&
+          dtor != (pthread_key_dtor)-1 &&
+          (val = atomic_exchange_explicit(&tib->tib_keys[i], 0,
+                                          memory_order_relaxed))) {
+        STRACE("_pthread_unkey(%t, %p)", dtor, val);
+        gotsome = true;
         dtor(val);
       }
     }
-    if (!gotsome) {
+    if (!gotsome)
       break;
-    }
   }
 }
 
 static void _pthread_ungarbage(struct CosmoTib *tib) {
   struct Garbages *g;
   while ((g = tib->tib_garbages)) {
     tib->tib_garbages = 0;
-    while (g->i--) {
+    while (g->i--)
       ((void (*)(intptr_t))g->p[g->i].fn)(g->p[g->i].arg);
-    }
     _weaken(free)(g->p);
     _weaken(free)(g);
   }
@@ -82,10 +80,11 @@ void __cxa_thread_finalize(void) {
   _pthread_unwind(tib);
   if (tib->tib_nsync)
     _weaken(nsync_waiter_destroy)(tib->tib_nsync);
-  __cxa_thread_unkey(tib);
+  _pthread_unkey(tib);
   _pthread_ungarbage(tib);
-  while ((dtor = __cxa_thread_atexit_list)) {
-    __cxa_thread_atexit_list = dtor->next;
+  while ((dtor = tib->tib_atexit)) {
+    STRACE("__cxa_finalize(%t, %p)", dtor->fun, dtor->arg);
+    tib->tib_atexit = dtor->next;
     ((void (*)(void *))dtor->fun)(dtor->arg);
     _weaken(free)(dtor);
   }
@@ -97,9 +96,11 @@ int __cxa_thread_atexit_impl(void *fun, void *arg, void *dso_symbol) {
     return -1;
   if (!(dtor = _weaken(malloc)(sizeof(struct Dtor))))
     return -1;
+  struct CosmoTib *tib;
+  tib = __get_tls();
   dtor->fun = fun;
   dtor->arg = arg;
-  dtor->next = __cxa_thread_atexit_list;
-  __cxa_thread_atexit_list = dtor;
+  dtor->next = tib->tib_atexit;
+  tib->tib_atexit = dtor;
   return 0;
 }

libc/thread/pthread_cancel.c

@@ -54,9 +54,8 @@ long _pthread_cancel_ack(void) {
     pthread_exit(PTHREAD_CANCELED);
   }
   pt->pt_flags |= PT_NOCANCEL;
-  if (IsOpenbsd()) {
+  if (IsOpenbsd())
     pt->pt_flags |= PT_OPENBSD_KLUDGE;
-  }
   return ecanceled();
 }
 
@@ -351,6 +350,7 @@ static errno_t _pthread_cancel_everyone(void) {
  * @param thread may be 0 to cancel all threads except self
  * @return 0 on success, or errno on error
  * @raise ESRCH if system thread wasn't alive or we lost a race
+ * @cancelationpoint
  */
 errno_t pthread_cancel(pthread_t thread) {
   struct PosixThread *arg;
@@ -401,6 +401,7 @@ void pthread_testcancel(void) {
  *
  * @return 0 if not cancelled or cancelation is blocked or `ECANCELED`
  *     in masked mode when the calling thread has been cancelled
+ * @cancelationpoint
  */
 errno_t pthread_testcancel_np(void) {
   struct PosixThread *pt;

libc/thread/pthread_create.c

@@ -117,6 +117,7 @@ static int PosixThread(void *arg, int tid) {
   }
   // set long jump handler so pthread_exit can bring control back here
   if (!setjmp(pt->pt_exiter)) {
+    sigdelset(&pt->pt_attr.__sigmask, SIGTHR);
     if (IsWindows()) {
       atomic_store_explicit(&__get_tls()->tib_sigmask, pt->pt_attr.__sigmask,
                             memory_order_release);

libc/thread/pthread_key_create.c

@@ -48,9 +48,9 @@ int pthread_key_create(pthread_key_t *key, pthread_key_dtor dtor) {
   if (!dtor)
     dtor = (pthread_key_dtor)-1;
   for (i = 0; i < PTHREAD_KEYS_MAX; ++i) {
-    if (!(expect = atomic_load_explicit(_pthread_key_dtor + i,
-                                        memory_order_acquire)) &&
-        atomic_compare_exchange_strong_explicit(_pthread_key_dtor + i, &expect,
+    if (!(expect = atomic_load_explicit(&_pthread_key_dtor[i],
+                                        memory_order_relaxed)) &&
+        atomic_compare_exchange_strong_explicit(&_pthread_key_dtor[i], &expect,
                                                 dtor, memory_order_release,
                                                 memory_order_relaxed)) {
       *key = i;

libc/thread/pthread_key_delete.c

@@ -16,7 +16,6 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/assert.h"
 #include "libc/errno.h"
 #include "libc/intrin/atomic.h"
 #include "libc/thread/posixthread.internal.h"
@@ -32,10 +31,12 @@
  *
  * @param key was created by pthread_key_create()
  * @return 0 on success, or errno on error
+ * @raise EINVAL if `key` is invalid
  */
 int pthread_key_delete(pthread_key_t k) {
-  unassert(0 <= k && k < PTHREAD_KEYS_MAX);
-  unassert(atomic_load_explicit(_pthread_key_dtor + k, memory_order_acquire));
-  atomic_store_explicit(_pthread_key_dtor + k, 0, memory_order_release);
+  if (!(0 <= k && k < PTHREAD_KEYS_MAX))
+    return EINVAL;  // corrupt key identifier
+  if (!atomic_exchange_explicit(&_pthread_key_dtor[k], 0, memory_order_acq_rel))
+    return EINVAL;  // delete called twice
   return 0;
 }

libc/thread/thread.h

@@ -1,7 +1,7 @@
 #ifndef COSMOPOLITAN_LIBC_THREAD_THREAD_H_
 #define COSMOPOLITAN_LIBC_THREAD_THREAD_H_
 
-#define PTHREAD_KEYS_MAX              48
+#define PTHREAD_KEYS_MAX              46
 #define PTHREAD_STACK_MIN             65536
 #define PTHREAD_DESTRUCTOR_ITERATIONS 4
 

libc/thread/tls.h

@@ -39,7 +39,8 @@ struct CosmoTib {
   uint32_t tib_sigstack_flags;
   _Atomic(int) tib_relock_maps;
   void *tib_nsync;
-  void *tib_keys[47];
+  void *tib_atexit;
+  _Atomic(void *) tib_keys[46];
 } __attribute__((__aligned__(64)));
 
 extern int __threaded;

test/ctl/all_of_test.cc

@@ -20,9 +20,9 @@
 #include "ctl/array.h"
 #include "libc/mem/leaks.h"
 
-#include <algorithm>
-#include <array>
-#define ctl std
+// #include <algorithm>
+// #include <array>
+// #define ctl std
 
 int
 main()