htl: Add pshared semaphore support

The implementation is extremely similar to the nptl implementation, but
with slight differences in the futex interface. This fixes some of BZ
25521.

htl/Makefile

@@ -130,7 +130,7 @@ libpthread-routines := pt-attr pt-attr-destroy pt-attr-getdetachstate	    \
 									    \
 	sem-close sem-destroy sem-getvalue sem-init sem-open		    \
 	sem-post sem-timedwait sem-trywait sem-unlink			    \
-	sem-wait							    \
+	sem-wait sem-waitfast						    \
 									    \
 	shm-directory							    \
 									    \

htl/pt-internal.h

@@ -331,4 +331,37 @@ extern const struct __pthread_rwlockattr __pthread_default_rwlockattr;
 /* Default condition attributes.  */
 extern const struct __pthread_condattr __pthread_default_condattr;
 
+/* Semaphore encoding.
+   See nptl implementation for the details.  */
+struct new_sem
+{
+#if __HAVE_64B_ATOMICS
+  /* The data field holds both value (in the least-significant 32 bits) and
+     nwaiters.  */
+# if __BYTE_ORDER == __LITTLE_ENDIAN
+#  define SEM_VALUE_OFFSET 0
+# elif __BYTE_ORDER == __BIG_ENDIAN
+#  define SEM_VALUE_OFFSET 1
+# else
+#  error Unsupported byte order.
+# endif
+# define SEM_NWAITERS_SHIFT 32
+# define SEM_VALUE_MASK (~(unsigned int)0)
+  uint64_t data;
+  int pshared;
+#define __SEMAPHORE_INITIALIZER(value, pshared) \
+  { (value), (pshared) }
+#else
+# define SEM_VALUE_SHIFT 1
+# define SEM_NWAITERS_MASK ((unsigned int)1)
+  unsigned int value;
+  unsigned int nwaiters;
+  int pshared;
+#define __SEMAPHORE_INITIALIZER(value, pshared) \
+  { (value) << SEM_VALUE_SHIFT, 0, (pshared) }
+#endif
+};
+
+extern int __sem_waitfast (struct new_sem *isem, int definitive_result);
+
 #endif /* pt-internal.h */

sysdeps/htl/bits/semaphore.h

@@ -27,21 +27,15 @@
 #include <bits/pthread.h>
 
 /* User visible part of a semaphore.  */
-struct __semaphore
-{
-  __pthread_spinlock_t __lock;
-  struct __pthread *__queue;
-  int __pshared;
-  int __value;
-  void *__data;
-};
 
-typedef struct __semaphore sem_t;
+#define __SIZEOF_SEM_T	20
+
+typedef union
+{
+  char __size[__SIZEOF_SEM_T];
+  long int __align;
+} sem_t;
 
 #define SEM_FAILED ((void *) 0)
 
-/* Initializer for a semaphore.  */
-#define __SEMAPHORE_INITIALIZER(pshared, value) \
-  { __PTHREAD_SPIN_LOCK_INITIALIZER, NULL, (pshared), (value), NULL }
-
 #endif /* bits/semaphore.h */

sysdeps/htl/sem-destroy.c

@@ -24,7 +24,15 @@
 int
 __sem_destroy (sem_t *sem)
 {
-  if (sem->__queue)
+  struct new_sem *isem = (struct new_sem *) sem;
+  if (
+#if __HAVE_64B_ATOMICS
+      atomic_load_relaxed (&isem->data) >> SEM_NWAITERS_SHIFT
+#else
+      atomic_load_relaxed (&isem->value) & SEM_NWAITERS_MASK
+      || isem->nwaiters
+#endif
+      )
     /* There are threads waiting on *SEM.  */
     {
       errno = EBUSY;

sysdeps/htl/sem-getvalue.c

@@ -22,9 +22,13 @@
 int
 __sem_getvalue (sem_t *restrict sem, int *restrict value)
 {
-  __pthread_spin_wait (&sem->__lock);
+  struct new_sem *isem = (struct new_sem *) sem;
-  *value = sem->__value;
+
-  __pthread_spin_unlock (&sem->__lock);
+#if __HAVE_64B_ATOMICS
+  *value = atomic_load_relaxed (&isem->data) & SEM_VALUE_MASK;
+#else
+  *value = atomic_load_relaxed (&isem->value) >> SEM_VALUE_SHIFT;
+#endif
 
   return 0;
 }

sysdeps/htl/sem-init.c

@@ -24,12 +24,6 @@
 int
 __sem_init (sem_t *sem, int pshared, unsigned value)
 {
-  if (pshared != 0)
-    {
-      errno = EOPNOTSUPP;
-      return -1;
-    }
-
 #ifdef SEM_VALUE_MAX
   if (value > SEM_VALUE_MAX)
     {
@@ -38,7 +32,9 @@ __sem_init (sem_t *sem, int pshared, unsigned value)
     }
 #endif
 
-  *sem = (sem_t) __SEMAPHORE_INITIALIZER (pshared, value);
+  struct new_sem *isem = (struct new_sem *) sem;
+
+  *isem = (struct new_sem) __SEMAPHORE_INITIALIZER (value, pshared);
   return 0;
 }
 

sysdeps/htl/sem-post.c

@@ -19,48 +19,50 @@
 #include <semaphore.h>
 #include <assert.h>
 
+#include <hurdlock.h>
+
 #include <pt-internal.h>
 
 int
 __sem_post (sem_t *sem)
 {
-  struct __pthread *wakeup;
+  struct new_sem *isem = (struct new_sem *) sem;
+  int flags = isem->pshared ? GSYNC_SHARED : 0;
 
-  __pthread_spin_wait (&sem->__lock);
+#if __HAVE_64B_ATOMICS
-  if (sem->__value > 0)
+  uint64_t d = atomic_load_relaxed (&isem->data);
-    /* Do a quick up.  */
+
+  do
     {
-      if (sem->__value == SEM_VALUE_MAX)
+      if ((d & SEM_VALUE_MASK) == SEM_VALUE_MAX)
 	{
-	  __pthread_spin_unlock (&sem->__lock);
 	  errno = EOVERFLOW;
 	  return -1;
 	}
-
-      assert (sem->__queue == NULL);
-      sem->__value++;
-      __pthread_spin_unlock (&sem->__lock);
-      return 0;
     }
+  while (!atomic_compare_exchange_weak_release (&isem->data, &d, d + 1));
 
-  if (sem->__queue == NULL)
+  if ((d >> SEM_NWAITERS_SHIFT) != 0)
-    /* No one waiting.  */
+    /* Wake one waiter.  */
+    __lll_wake (((unsigned int *) &isem->data) + SEM_VALUE_OFFSET, flags);
+#else
+  unsigned int v = atomic_load_relaxed (&isem->value);
+
+  do
     {
-      sem->__value = 1;
+      if ((v >> SEM_VALUE_SHIFT) == SEM_VALUE_MAX)
-      __pthread_spin_unlock (&sem->__lock);
+	{
-      return 0;
+	  errno = EOVERFLOW;
+	  return -1;
+	}
     }
+  while (!atomic_compare_exchange_weak_release
+	  (&isem->value, &v, v + (1 << SEM_VALUE_SHIFT)));
 
-  /* Wake someone up.  */
+  if ((v & SEM_NWAITERS_MASK) != 0)
-
+    /* Wake one waiter.  */
-  /* First dequeue someone.  */
+    __lll_wake (&isem->value, flags);
-  wakeup = sem->__queue;
+#endif
-  __pthread_dequeue (wakeup);
-
-  /* Then drop the lock and transfer control.  */
-  __pthread_spin_unlock (&sem->__lock);
-
-  __pthread_wakeup (wakeup);
 
   return 0;
 }

sysdeps/htl/sem-timedwait.c

@@ -20,37 +20,27 @@
 #include <errno.h>
 #include <assert.h>
 #include <time.h>
+#include <hurdlock.h>
+#include <hurd/hurd.h>
+#include <sysdep-cancel.h>
 
 #include <pt-internal.h>
 
-struct cancel_ctx
+#if !__HAVE_64B_ATOMICS
-{
+static void
-  struct __pthread *wakeup;
+__sem_wait_32_finish (struct new_sem *isem);
-  sem_t *sem;
+#endif
-  int cancel_wake;
-};
 
 static void
-cancel_hook (void *arg)
+__sem_wait_cleanup (void *arg)
 {
-  struct cancel_ctx *ctx = arg;
+  struct new_sem *isem = arg;
-  struct __pthread *wakeup = ctx->wakeup;
-  sem_t *sem = ctx->sem;
-  int unblock;
 
-  __pthread_spin_wait (&sem->__lock);
+#if __HAVE_64B_ATOMICS
-  /* The thread only needs to be awaken if it's blocking or about to block.
+  atomic_fetch_add_relaxed (&isem->data, -((uint64_t) 1 << SEM_NWAITERS_SHIFT));
-     If it was already unblocked, it's not queued any more.  */
+#else
-  unblock = wakeup->prevp != NULL;
+  __sem_wait_32_finish (isem);
-  if (unblock)
+#endif
-    {
-      __pthread_dequeue (wakeup);
-      ctx->cancel_wake = 1;
-    }
-  __pthread_spin_unlock (&sem->__lock);
-
-  if (unblock)
-    __pthread_wakeup (wakeup);
 }
 
 int
@@ -58,123 +48,148 @@ __sem_timedwait_internal (sem_t *restrict sem,
 			  clockid_t clock_id,
 			  const struct timespec *restrict timeout)
 {
-  error_t err;
+  struct new_sem *isem = (struct new_sem *) sem;
-  int cancelled, oldtype, drain;
+  int err, ret = 0;
-  int ret = 0;
+  int flags = isem->pshared ? GSYNC_SHARED : 0;
 
-  struct __pthread *self = _pthread_self ();
+  __pthread_testcancel ();
-  struct cancel_ctx ctx;
-  ctx.wakeup = self;
-  ctx.sem = sem;
-  ctx.cancel_wake = 0;
 
-  /* Test for a pending cancellation request, switch to deferred mode for
+  if (__sem_waitfast (isem, 0) == 0)
-     safer resource handling, and prepare the hook to call in case we're
+    return 0;
-     cancelled while blocking.  Once CANCEL_LOCK is released, the cancellation
-     hook can be called by another thread at any time.  Whatever happens,
-     this function must exit with MUTEX locked.
 
-     This function contains inline implementations of pthread_testcancel and
+  int cancel_oldtype = LIBC_CANCEL_ASYNC();
-     pthread_setcanceltype to reduce locking overhead.  */
-  __pthread_mutex_lock (&self->cancel_lock);
-  cancelled = (self->cancel_state == PTHREAD_CANCEL_ENABLE)
-      && self->cancel_pending;
 
-  if (cancelled)
+#if __HAVE_64B_ATOMICS
+  uint64_t d = atomic_fetch_add_relaxed (&sem->data,
+		 (uint64_t) 1 << SEM_NWAITERS_SHIFT);
+
+  pthread_cleanup_push (__sem_wait_cleanup, isem);
+
+  for (;;)
     {
-      __pthread_mutex_unlock (&self->cancel_lock);
+      if ((d & SEM_VALUE_MASK) == 0)
-      __pthread_exit (PTHREAD_CANCELED);
+	{
+	  /* No token, sleep.  */
+	  if (timeout)
+	    err = __lll_abstimed_wait_intr (
+		      ((unsigned int *) &sem->data) + SEM_VALUE_OFFSET,
+		      0, timeout, flags, clock_id);
+	  else
+	    err = __lll_wait_intr (
+		      ((unsigned int *) &sem->data) + SEM_VALUE_OFFSET,
+		      0, flags);
+
+	  if (err != 0)
+	    {
+	      /* Error, interruption or timeout, abort.  */
+	      if (err == KERN_TIMEDOUT)
+		err = ETIMEDOUT;
+	      if (err == KERN_INTERRUPTED)
+		err = EINTR;
+	      ret = __hurd_fail (err);
+	      __sem_wait_cleanup (isem);
+	      break;
+	    }
+
+	  /* Token changed */
+	  d = atomic_load_relaxed (&sem->data);
+	}
+      else
+	{
+	  /* Try to acquire and dequeue.  */
+	  if (atomic_compare_exchange_weak_acquire (&sem->data,
+	      &d, d - 1 - ((uint64_t) 1 << SEM_NWAITERS_SHIFT)))
+	    {
+	      /* Success */
+	      ret = 0;
+	      break;
+	    }
+	}
     }
 
-  self->cancel_hook = cancel_hook;
+  pthread_cleanup_pop (0);
-  self->cancel_hook_arg = &ctx;
+#else
-  oldtype = self->cancel_type;
+  unsigned int v;
 
-  if (oldtype != PTHREAD_CANCEL_DEFERRED)
+  atomic_fetch_add_acquire (&isem->nwaiters, 1);
-    self->cancel_type = PTHREAD_CANCEL_DEFERRED;
 
-  /* Add ourselves to the list of waiters.  This is done while setting
+  pthread_cleanup_push (__sem_wait_cleanup, isem);
-     the cancellation hook to simplify the cancellation procedure, i.e.
+
-     if the thread is queued, it can be cancelled, otherwise it is
+  v = atomic_load_relaxed (&isem->value);
-     already unblocked, progressing on the return path.  */
+  do
-  __pthread_spin_wait (&sem->__lock);
-  if (sem->__value > 0)
-    /* Successful down.  */
     {
-      sem->__value--;
+      do
-      __pthread_spin_unlock (&sem->__lock);
+	{
-      goto out_locked;
+	  do
+	    {
+	      if ((v & SEM_NWAITERS_MASK) != 0)
+		break;
+	    }
+	  while (!atomic_compare_exchange_weak_release (&isem->value,
+	      &v, v | SEM_NWAITERS_MASK));
+
+	  if ((v >> SEM_VALUE_SHIFT) == 0)
+	    {
+	      /* No token, sleep.  */
+	      if (timeout)
+		err = __lll_abstimed_wait_intr (&isem->value,
+			  SEM_NWAITERS_MASK, timeout, flags, clock_id);
+	      else
+		err = __lll_wait_intr (&isem->value,
+			  SEM_NWAITERS_MASK, flags);
+
+	      if (err != 0)
+		{
+		  /* Error, interruption or timeout, abort.  */
+		  if (err == KERN_TIMEDOUT)
+		    err = ETIMEDOUT;
+		  if (err == KERN_INTERRUPTED)
+		    err = EINTR;
+		  ret = __hurd_fail (err);
+		  goto error;
+		}
+
+	      /* Token changed */
+	      v = atomic_load_relaxed (&isem->value);
+	    }
+	}
+      while ((v >> SEM_VALUE_SHIFT) == 0);
     }
+  while (!atomic_compare_exchange_weak_acquire (&isem->value,
+	  &v, v - (1 << SEM_VALUE_SHIFT)));
 
-  if (timeout != NULL && ! valid_nanoseconds (timeout->tv_nsec))
+error:
-    {
+  pthread_cleanup_pop (0);
-      errno = EINVAL;
-      ret = -1;
-      __pthread_spin_unlock (&sem->__lock);
-      goto out_locked;
-    }
 
-  /* Add ourselves to the queue.  */
+  __sem_wait_32_finish (isem);
-  __pthread_enqueue (&sem->__queue, self);
+#endif
-  __pthread_spin_unlock (&sem->__lock);
 
-  __pthread_mutex_unlock (&self->cancel_lock);
+  LIBC_CANCEL_RESET (cancel_oldtype);
-
-  /* Block the thread.  */
-  if (timeout != NULL)
-    err = __pthread_timedblock_intr (self, timeout, clock_id);
-  else
-    err = __pthread_block_intr (self);
-
-  __pthread_spin_wait (&sem->__lock);
-  if (self->prevp == NULL)
-    /* Another thread removed us from the queue, which means a wakeup message
-       has been sent.  It was either consumed while we were blocking, or
-       queued after we timed out and before we acquired the semaphore lock, in
-       which case the message queue must be drained.  */
-    drain = err ? 1 : 0;
-  else
-    {
-      /* We're still in the queue.  Noone attempted to wake us up, i.e. we
-         timed out.  */
-      __pthread_dequeue (self);
-      drain = 0;
-    }
-  __pthread_spin_unlock (&sem->__lock);
-
-  if (drain)
-    __pthread_block (self);
-
-  if (err)
-    {
-      assert (err == ETIMEDOUT || err == EINTR);
-      errno = err;
-      ret = -1;
-    }
-
-  /* We're almost done.  Remove the unblock hook, restore the previous
-     cancellation type, and check for a pending cancellation request.  */
-  __pthread_mutex_lock (&self->cancel_lock);
-out_locked:
-  self->cancel_hook = NULL;
-  self->cancel_hook_arg = NULL;
-  self->cancel_type = oldtype;
-  cancelled = (self->cancel_state == PTHREAD_CANCEL_ENABLE)
-      && self->cancel_pending;
-  __pthread_mutex_unlock (&self->cancel_lock);
-
-  if (cancelled)
-    {
-      if (ret == 0 && ctx.cancel_wake == 0)
-	/* We were cancelled while waking up with a token, put it back.  */
-	__sem_post (sem);
-
-      __pthread_exit (PTHREAD_CANCELED);
-    }
 
   return ret;
 }
 
+#if !__HAVE_64B_ATOMICS
+/* Stop being a registered waiter (non-64b-atomics code only).  */
+static void
+__sem_wait_32_finish (struct new_sem *isem)
+{
+  unsigned int wguess = atomic_load_relaxed (&isem->nwaiters);
+  if (wguess == 1)
+    atomic_fetch_and_acquire (&isem->value, ~SEM_NWAITERS_MASK);
+
+  unsigned int wfinal = atomic_fetch_add_release (&isem->nwaiters, -1);
+  if (wfinal > 1 && wguess == 1)
+    {
+      unsigned int v = atomic_fetch_or_relaxed (&isem->value,
+						SEM_NWAITERS_MASK);
+      v >>= SEM_VALUE_SHIFT;
+      while (v--)
+	__lll_wake (&isem->value, isem->pshared ? GSYNC_SHARED : 0);
+    }
+}
+#endif
+
 int
 __sem_clockwait (sem_t *sem, clockid_t clockid,
 		 const struct timespec *restrict timeout)

sysdeps/htl/sem-trywait.c

@@ -24,18 +24,13 @@
 int
 __sem_trywait (sem_t *sem)
 {
-  __pthread_spin_wait (&sem->__lock);
+  struct new_sem *isem = (struct new_sem *) sem;
-  if (sem->__value > 0)
+
-    /* Successful down.  */
+  if (__sem_waitfast (isem, 1) == 0)
-    {
+    return 0;
-      sem->__value--;
-      __pthread_spin_unlock (&sem->__lock);
-      return 0;
-    }
-  __pthread_spin_unlock (&sem->__lock);
 
   errno = EAGAIN;
   return -1;
 }
 
-strong_alias (__sem_trywait, sem_trywait);
+weak_alias (__sem_trywait, sem_trywait);

sysdeps/htl/sem-waitfast.c

@@ -0,0 +1,55 @@
+/* Lock a semaphore if it does not require blocking.  Generic version.
+   Copyright (C) 2005-2020 Free Software Foundation, Inc.
+   This file is part of the GNU C Library.
+
+   The GNU C Library is free software; you can redistribute it and/or
+   modify it under the terms of the GNU Lesser General Public
+   License as published by the Free Software Foundation; either
+   version 2.1 of the License, or (at your option) any later version.
+
+   The GNU C Library is distributed in the hope that it will be useful,
+   but WITHOUT ANY WARRANTY; without even the implied warranty of
+   MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE.  See the GNU
+   Lesser General Public License for more details.
+
+   You should have received a copy of the GNU Lesser General Public
+   License along with the GNU C Library;  if not, see
+   <https://www.gnu.org/licenses/>.  */
+
+#include <semaphore.h>
+#include <errno.h>
+
+#include <pt-internal.h>
+
+int
+__sem_waitfast (struct new_sem *isem, int definitive_result)
+{
+#if __HAVE_64B_ATOMICS
+  uint64_t d = atomic_load_relaxed (&isem->data);
+
+  do
+    {
+      if ((d & SEM_VALUE_MASK) == 0)
+	break;
+      if (atomic_compare_exchange_weak_acquire (&isem->data, &d, d - 1))
+	/* Successful down.  */
+	return 0;
+    }
+  while (definitive_result);
+  return -1;
+#else
+  unsigned v = atomic_load_relaxed (&isem->value);
+
+  do
+    {
+      if ((v >> SEM_VALUE_SHIFT) == 0)
+	break;
+      if (atomic_compare_exchange_weak_acquire (&isem->value,
+	    &v, v - (1 << SEM_VALUE_SHIFT)))
+	/* Successful down.  */
+	return 0;
+    }
+  while (definitive_result);
+  return -1;
+#endif
+}

sysdeps/mach/hurd/i386/Makefile

@@ -116,7 +116,6 @@ test-xfail-tst-cond13 = yes
 test-xfail-tst-cond23 = yes
 test-xfail-tst-rwlock4 = yes
 test-xfail-tst-rwlock12 = yes
-test-xfail-tst-sem3 = yes
 test-xfail-tst-barrier2 = yes
 test-xfail-tst-pututxline-cache = yes
 test-xfail-tst-pututxline-lockfail = yes