Make spin locks go faster

libc/calls/getloadavg-nt.c

@@ -30,15 +30,27 @@
 
 static int cpus;
 static double load;
-static pthread_spinlock_t lock;
 static struct NtFileTime idle1, kern1, user1;
+static pthread_mutex_t getloadavg_lock;
+
+static void __getloadavg_lock(void) {
+  pthread_mutex_lock(&getloadavg_lock);
+}
+
+static void __getloadavg_unlock(void) {
+  pthread_mutex_unlock(&getloadavg_lock);
+}
+
+static void __getloadavg_wipe(void) {
+  pthread_mutex_init(&getloadavg_lock, 0);
+}
 
 textwindows int sys_getloadavg_nt(double *a, int n) {
   int i, rc;
   uint64_t elapsed, used;
   struct NtFileTime idle, kern, user;
   BLOCK_SIGNALS;
-  pthread_spin_lock(&lock);
+  __getloadavg_lock();
   if (GetSystemTimes(&idle, &kern, &user)) {
     elapsed = (FT(kern) - FT(kern1)) + (FT(user) - FT(user1));
     if (elapsed) {
@@ -54,7 +66,7 @@ textwindows int sys_getloadavg_nt(double *a, int n) {
   } else {
     rc = __winerr();
   }
-  pthread_spin_unlock(&lock);
+  __getloadavg_unlock();
   ALLOW_SIGNALS;
   return rc;
 }
@@ -65,5 +77,7 @@ __attribute__((__constructor__(40))) static textstartup void ntinitload(void) {
     cpus = __get_cpu_count() / 2;
     cpus = MAX(1, cpus);
     GetSystemTimes(&idle1, &kern1, &user1);
+    pthread_atfork(__getloadavg_lock, __getloadavg_unlock, __getloadavg_wipe);
+    __getloadavg_wipe();
   }
 }

libc/calls/sig.c

@@ -51,6 +51,7 @@
 #include "libc/sysv/consts/sicode.h"
 #include "libc/sysv/consts/ss.h"
 #include "libc/thread/posixthread.internal.h"
+#include "libc/thread/thread.h"
 #ifdef __x86_64__
 
 /**
@@ -64,6 +65,20 @@ struct SignalFrame {
   ucontext_t ctx;
 };
 
+static pthread_mutex_t __sig_lock_obj;
+
+static void __sig_wipe(void) {
+  pthread_mutex_init(&__sig_lock_obj, 0);
+}
+
+static void __sig_lock(void) {
+  pthread_mutex_lock(&__sig_lock_obj);
+}
+
+static void __sig_unlock(void) {
+  pthread_mutex_unlock(&__sig_lock_obj);
+}
+
 static textwindows bool __sig_ignored_by_default(int sig) {
   return sig == SIGURG ||   //
          sig == SIGCONT ||  //
@@ -318,11 +333,10 @@ static textwindows int __sig_killer(struct PosixThread *pt, int sig, int sic) {
   // take control of thread
   // suspending the thread happens asynchronously
   // however getting the context blocks until it's frozen
-  static pthread_spinlock_t killer_lock;
-  pthread_spin_lock(&killer_lock);
+  __sig_lock();
   if (SuspendThread(th) == -1u) {
     STRACE("SuspendThread failed w/ %d", GetLastError());
-    pthread_spin_unlock(&killer_lock);
+    __sig_unlock();
     return ESRCH;
   }
   struct NtContext nc;
@@ -330,10 +344,10 @@ static textwindows int __sig_killer(struct PosixThread *pt, int sig, int sic) {
   if (!GetThreadContext(th, &nc)) {
     STRACE("GetThreadContext failed w/ %d", GetLastError());
     ResumeThread(th);
-    pthread_spin_unlock(&killer_lock);
+    __sig_unlock();
     return ESRCH;
   }
-  pthread_spin_unlock(&killer_lock);
+  __sig_unlock();
 
   // we can't preempt threads that masked sig or are blocked
   // we can't preempt threads that are running in win32 code
@@ -612,6 +626,8 @@ __attribute__((__constructor__(10))) textstartup void __sig_init(void) {
     return;
   AddVectoredExceptionHandler(true, (void *)__sig_crash);
   SetConsoleCtrlHandler((void *)__sig_console, true);
+  pthread_atfork(__sig_lock, __sig_unlock, __sig_wipe);
+  __sig_wipe();
 }
 
 #endif /* __x86_64__ */

libc/intrin/cxalock.c

@@ -35,4 +35,5 @@ void __cxa_unlock(void) {
 
 __attribute__((__constructor__(60))) static textstartup void __cxa_init() {
   pthread_atfork(__cxa_lock, __cxa_unlock, __cxa_wipe);
+  __cxa_wipe();
 }

libc/intrin/maps.c

@@ -86,27 +86,15 @@ void __maps_init(void) {
 
 privileged bool __maps_lock(void) {
   struct CosmoTib *tib;
-  if (!__tls_enabled)
+  if (__tls_enabled)
     return false;
   tib = __get_tls_privileged();
   if (atomic_fetch_add_explicit(&tib->tib_relock_maps, 1, memory_order_relaxed))
     return true;
-  int backoff = 0;
-  while (atomic_exchange_explicit(&__maps.lock, 1, memory_order_acquire)) {
-    if (backoff < 7) {
-      volatile int i;
-      for (i = 0; i != 1 << backoff; i++) {
-      }
-      backoff++;
-    } else {
-      // STRACE("pthread_delay_np(__maps)");
-#if defined(__GNUC__) && defined(__aarch64__)
-      __asm__ volatile("yield");
-#elif defined(__GNUC__) && (defined(__x86_64__) || defined(__i386__))
-      __asm__ volatile("pause");
-#endif
-    }
-  }
+  while (atomic_exchange_explicit(&__maps.lock, 1, memory_order_acquire))
+    for (;;)
+      if (!atomic_load_explicit(&__maps.lock, memory_order_relaxed))
+        break;
   return false;
 }
 

libc/intrin/pthread_atfork_actual.c

@@ -16,21 +16,11 @@
 │ TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR             │
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
-#include "libc/atomic.h"
-#include "libc/calls/state.internal.h"
-#include "libc/cosmo.h"
-#include "libc/dce.h"
 #include "libc/errno.h"
-#include "libc/intrin/atomic.h"
-#include "libc/intrin/dll.h"
 #include "libc/intrin/strace.h"
 #include "libc/macros.internal.h"
-#include "libc/proc/proc.internal.h"
-#include "libc/runtime/runtime.h"
-#include "libc/str/str.h"
 #include "libc/thread/posixthread.internal.h"
 #include "libc/thread/thread.h"
-#include "libc/thread/tls.h"
 
 struct AtFork {
   struct AtFork *p[2];
@@ -38,16 +28,16 @@ struct AtFork {
 };
 
 static struct AtForks {
-  pthread_spinlock_t lock;
+  pthread_mutex_t lock;
   struct AtFork *list;
-  struct AtFork pool[64];
-  atomic_int allocated;
-} _atforks;
+  struct AtFork pool[256];
+  int allocated;
+} _atforks = {
+    PTHREAD_MUTEX_INITIALIZER,
+};
 
 static void _pthread_onfork(int i, const char *op) {
   struct AtFork *a;
-  if (!i)
-    pthread_spin_lock(&_atforks.lock);
   for (a = _atforks.list; a; a = a->p[!i]) {
     if (a->f[i]) {
       STRACE("pthread_atfork(%s, %t)", op, a->f[i]);
@@ -55,47 +45,41 @@ static void _pthread_onfork(int i, const char *op) {
     }
     _atforks.list = a;
   }
-  if (i)
-    pthread_spin_unlock(&_atforks.lock);
 }
 
 void _pthread_onfork_prepare(void) {
+  pthread_mutex_lock(&_atforks.lock);
   _pthread_onfork(0, "prepare");
 }
 
 void _pthread_onfork_parent(void) {
   _pthread_onfork(1, "parent");
+  pthread_mutex_unlock(&_atforks.lock);
 }
 
 void _pthread_onfork_child(void) {
+  pthread_mutex_init(&_atforks.lock, 0);
   _pthread_onfork(2, "child");
 }
 
-static struct AtFork *_pthread_atfork_alloc(void) {
-  int i, n = ARRAYLEN(_atforks.pool);
-  if (atomic_load_explicit(&_atforks.allocated, memory_order_relaxed) < n &&
-      (i = atomic_fetch_add(&_atforks.allocated, 1)) < n) {
-    return _atforks.pool + i;
-  } else {
-    return 0;
-  }
-}
-
 int _pthread_atfork(atfork_f prepare, atfork_f parent, atfork_f child) {
   int rc;
   struct AtFork *a;
-  if (!(a = _pthread_atfork_alloc()))
-    return ENOMEM;
+  pthread_mutex_lock(&_atforks.lock);
+  if (_atforks.allocated < ARRAYLEN(_atforks.pool)) {
+    a = &_atforks.pool[_atforks.allocated++];
     a->f[0] = prepare;
     a->f[1] = parent;
     a->f[2] = child;
-  pthread_spin_lock(&_atforks.lock);
     a->p[0] = 0;
     a->p[1] = _atforks.list;
     if (_atforks.list)
       _atforks.list->p[0] = a;
     _atforks.list = a;
-  pthread_spin_unlock(&_atforks.lock);
     rc = 0;
+  } else {
+    rc = ENOMEM;
+  }
+  pthread_mutex_unlock(&_atforks.lock);
   return rc;
 }

libc/intrin/pthread_mutex_lock.c

@@ -31,17 +31,16 @@
 #include "third_party/nsync/futex.internal.h"
 #include "third_party/nsync/mu.h"
 
-static void pthread_mutex_lock_naive(pthread_mutex_t *mutex, uint64_t word) {
+static void pthread_mutex_lock_spin(atomic_int *word) {
   int backoff = 0;
-  uint64_t lock;
   for (;;) {
-    word = MUTEX_UNLOCK(word);
-    lock = MUTEX_LOCK(word);
-    if (atomic_compare_exchange_weak_explicit(&mutex->_word, &word, lock,
-                                              memory_order_acquire,
-                                              memory_order_relaxed))
-      return;
-    backoff = pthread_delay_np(mutex, backoff);
+    if (!atomic_exchange_explicit(word, 1, memory_order_acquire))
+      break;
+    for (;;) {
+      if (!atomic_load_explicit(word, memory_order_relaxed))
+        break;
+      backoff = pthread_delay_np(word, backoff);
+    }
   }
 }
 
@@ -96,8 +95,13 @@ static errno_t pthread_mutex_lock_recursive(pthread_mutex_t *mutex,
       mutex->_pid = __pid;
       return 0;
     }
+    for (;;) {
+      word = atomic_load_explicit(&mutex->_word, memory_order_relaxed);
+      if (!MUTEX_LOCKED(word))
+        break;
       backoff = pthread_delay_np(mutex, backoff);
     }
+  }
 }
 
 static errno_t pthread_mutex_lock_impl(pthread_mutex_t *mutex) {
@@ -121,7 +125,7 @@ static errno_t pthread_mutex_lock_impl(pthread_mutex_t *mutex) {
     if (_weaken(nsync_futex_wait_)) {
       pthread_mutex_lock_drepper(&mutex->_futex, MUTEX_PSHARED(word));
     } else {
-      pthread_mutex_lock_naive(mutex, word);
+      pthread_mutex_lock_spin(&mutex->_futex);
     }
     return 0;
   }

libc/intrin/pthread_mutex_trylock.c

@@ -27,14 +27,8 @@
 #include "third_party/nsync/futex.internal.h"
 #include "third_party/nsync/mu.h"
 
-static errno_t pthread_mutex_trylock_naive(pthread_mutex_t *mutex,
-                                           uint64_t word) {
-  uint64_t lock;
-  word = MUTEX_UNLOCK(word);
-  lock = MUTEX_LOCK(word);
-  if (atomic_compare_exchange_weak_explicit(&mutex->_word, &word, lock,
-                                            memory_order_acquire,
-                                            memory_order_relaxed))
+static errno_t pthread_mutex_trylock_spin(atomic_int *word) {
+  if (!atomic_exchange_explicit(word, 1, memory_order_acquire))
     return 0;
   return EBUSY;
 }
@@ -116,7 +110,7 @@ errno_t pthread_mutex_trylock(pthread_mutex_t *mutex) {
     if (_weaken(nsync_futex_wait_)) {
       return pthread_mutex_trylock_drepper(&mutex->_futex);
     } else {
-      return pthread_mutex_trylock_naive(mutex, word);
+      return pthread_mutex_trylock_spin(&mutex->_futex);
     }
   }
 

libc/intrin/pthread_mutex_unlock.c

@@ -28,9 +28,8 @@
 #include "third_party/nsync/futex.internal.h"
 #include "third_party/nsync/mu.h"
 
-static void pthread_mutex_unlock_naive(pthread_mutex_t *mutex, uint64_t word) {
-  uint64_t lock = MUTEX_UNLOCK(word);
-  atomic_store_explicit(&mutex->_word, lock, memory_order_release);
+static void pthread_mutex_unlock_spin(atomic_int *word) {
+  atomic_store_explicit(word, 0, memory_order_release);
 }
 
 // see "take 3" algorithm in "futexes are tricky" by ulrich drepper
@@ -102,7 +101,7 @@ errno_t pthread_mutex_unlock(pthread_mutex_t *mutex) {
     if (_weaken(nsync_futex_wake_)) {
       pthread_mutex_unlock_drepper(&mutex->_futex, MUTEX_PSHARED(word));
     } else {
-      pthread_mutex_unlock_naive(mutex, word);
+      pthread_mutex_unlock_spin(&mutex->_futex);
     }
     return 0;
   }

libc/runtime/at_quick_exit.c

@@ -21,30 +21,46 @@
 #include "libc/runtime/runtime.h"
 #include "libc/thread/thread.h"
 
-static void (*funcs[32])(void);
 static int count;
-static pthread_spinlock_t lock;
-pthread_spinlock_t *const __at_quick_exit_lockptr = &lock;
+static void (*funcs[32])(void);
+static pthread_mutex_t __quick_exit_lock_obj;
+
+static void __quick_exit_wipe(void) {
+  pthread_mutex_init(&__quick_exit_lock_obj, 0);
+}
+
+static void __quick_exit_lock(void) {
+  pthread_mutex_lock(&__quick_exit_lock_obj);
+}
+
+static void __quick_exit_unlock(void) {
+  pthread_mutex_unlock(&__quick_exit_lock_obj);
+}
 
 void __funcs_on_quick_exit(void) {
   void (*func)(void);
-  pthread_spin_lock(&lock);
+  __quick_exit_lock();
   while (count) {
     func = funcs[--count];
-    pthread_spin_unlock(&lock);
+    __quick_exit_unlock();
     func();
-    pthread_spin_lock(&lock);
+    __quick_exit_lock();
   }
 }
 
 int at_quick_exit(void func(void)) {
   int res = 0;
-  pthread_spin_lock(&lock);
+  __quick_exit_lock();
   if (count == ARRAYLEN(funcs)) {
     res = -1;
   } else {
     funcs[count++] = func;
   }
-  pthread_spin_unlock(&lock);
+  __quick_exit_unlock();
   return res;
 }
+
+__attribute__((__constructor__(10))) textstartup void __quick_exit_init(void) {
+  pthread_atfork(__quick_exit_lock, __quick_exit_unlock, __quick_exit_wipe);
+  __quick_exit_wipe();
+}

libc/runtime/getsymboltable.c

@@ -17,6 +17,8 @@
 │ PERFORMANCE OF THIS SOFTWARE.                                                │
 ╚─────────────────────────────────────────────────────────────────────────────*/
 #include "libc/assert.h"
+#include "libc/atomic.h"
+#include "libc/cosmo.h"
 #include "libc/errno.h"
 #include "libc/intrin/promises.h"
 #include "libc/intrin/strace.h"
@@ -27,14 +29,12 @@
 #include "libc/runtime/symbols.internal.h"
 #include "libc/runtime/zipos.internal.h"
 #include "libc/str/str.h"
-#include "libc/thread/thread.h"
 #include "libc/x/x.h"
 #include "libc/zip.internal.h"
 #include "third_party/puff/puff.h"
 
 __static_yoink("__get_symbol");
 
-static pthread_spinlock_t g_lock;
 struct SymbolTable *__symtab;  // for kprintf
 
 static ssize_t GetZipFile(struct Zipos *zipos, const char *name) {
@@ -100,6 +100,25 @@ static struct SymbolTable *GetSymbolTableFromElf(void) {
   }
 }
 
+static void GetSymbolTableInit(void) {
+  struct Zipos *z;
+  int e = errno;
+  if (!__symtab && !__isworker) {
+    if (_weaken(__zipos_get) && (z = _weaken(__zipos_get)())) {
+      if ((__symtab = GetSymbolTableFromZip(z))) {
+        __symtab->names =
+            (uint32_t *)((char *)__symtab + __symtab->names_offset);
+        __symtab->name_base =
+            (char *)((char *)__symtab + __symtab->name_base_offset);
+      }
+    }
+    if (!__symtab) {
+      __symtab = GetSymbolTableFromElf();
+    }
+  }
+  errno = e;
+}
+
 /**
  * Returns symbol table singleton.
  *
@@ -121,24 +140,7 @@ static struct SymbolTable *GetSymbolTableFromElf(void) {
  * @return symbol table, or NULL if not found
  */
 struct SymbolTable *GetSymbolTable(void) {
-  struct Zipos *z;
-  if (pthread_spin_trylock(&g_lock))
-    return 0;
-  int e = errno;
-  if (!__symtab && !__isworker) {
-    if (_weaken(__zipos_get) && (z = _weaken(__zipos_get)())) {
-      if ((__symtab = GetSymbolTableFromZip(z))) {
-        __symtab->names =
-            (uint32_t *)((char *)__symtab + __symtab->names_offset);
-        __symtab->name_base =
-            (char *)((char *)__symtab + __symtab->name_base_offset);
-      }
-    }
-    if (!__symtab) {
-      __symtab = GetSymbolTableFromElf();
-    }
-  }
-  errno = e;
-  pthread_spin_unlock(&g_lock);
+  static atomic_uint once;
+  cosmo_once(&once, GetSymbolTableInit);
   return __symtab;
 }

libc/stdio/flockfile.c

@@ -55,9 +55,14 @@ static void __stdio_fork_parent(void) {
 
 static void __stdio_fork_child(void) {
   FILE *f;
-  for (int i = __fflush.handles.i; i--;)
+  for (int i = __fflush.handles.i; i--;) {
+    pthread_mutexattr_t attr;
+    pthread_mutexattr_init(&attr);
+    pthread_mutexattr_settype(&attr, PTHREAD_MUTEX_RECURSIVE);
     if ((f = __fflush.handles.p[i]))
-      f->lock = (pthread_mutex_t)PTHREAD_RECURSIVE_MUTEX_INITIALIZER_NP;
+      pthread_mutex_init(&f->lock, &attr);
+    pthread_mutexattr_destroy(&attr);
+  }
   pthread_mutex_init(&__fflush_lock_obj, 0);
 }
 

libc/thread/pthread_spin_lock.c

@@ -38,8 +38,12 @@
  * @see pthread_spin_init
  */
 errno_t pthread_spin_lock(pthread_spinlock_t *spin) {
-  while (atomic_exchange_explicit(&spin->_lock, 1, memory_order_acquire)) {
-    pthread_pause_np();
+  for (;;) {
+    if (!atomic_exchange_explicit(&spin->_lock, 1, memory_order_acquire))
+      break;
+    for (;;)
+      if (!atomic_load_explicit(&spin->_lock, memory_order_relaxed))
+        break;
   }
   return 0;
 }