+2000-04-11 Ulrich Drepper <drepper@redhat.com>
+
+ * internals.h: Define MEMORY_BARRIER as empty if not defined already.
+ * spinlock.c (__pthread_lock): Add memory barriers.
+ (__pthread_unlock): Likewise.
+ * sysdeps/alpha/pt-machine.h (MEMORY_BARRIER): Define using mb
+ instruction.
+ (RELEASE): Not needed anymore.
+ (__compare_and_swap): Mark asm as modifying memory.
+ * sysdeps/powerpc/pt-machine.h (sync): Remove. Replace with definition
+ of MEMORY_BARRIER.
+ (__compare_and_swap): Use MEMORY_BARRIER instead of sync.
+ * sysdeps/sparc/sparc32/pt-machine.h (RELEASE): Not needed anymore.
+ (MEMORY_BARRIER): Define using stbar.
+ * sysdeps/sparc/sparc64/pt-machine.h (MEMORY_BARRIER): Define using
+ stbar.
+ (__compare_and_swap): Use MEMORY_BARRIER to ensure ordering.
+ Patch by Xavier Leroy <Xavier.Leroy@inria.fr> based on comments by
+ Mike Burrows <m3b@pa.dec.com>.
+
2000-04-09 Ulrich Drepper <drepper@redhat.com>
* signals.c (sigaction): Fix return value for the case SIG is one
#endif
}
+/* If MEMORY_BARRIER isn't defined in pt-machine.h, assume the architecture
+ doesn't need a memory barrier instruction (e.g. Intel x86) */
+
+#ifndef MEMORY_BARRIER
+#define MEMORY_BARRIER()
+#endif
+
/* Max number of times we must spin on a spinlock calling sched_yield().
After MAX_SPIN_COUNT iterations, we put the calling thread to sleep. */
if (self != NULL) {
ASSERT(self->p_nextlock == NULL);
THREAD_SETMEM(self, p_nextlock, (pthread_descr) oldstatus);
+ /* Make sure the store in p_nextlock completes before performing
+ the compare-and-swap */
+ MEMORY_BARRIER();
}
} while(! compare_and_swap(&lock->__status, oldstatus, newstatus,
&lock->__spinlock));
maxprio = thr->p_priority;
}
ptr = &(thr->p_nextlock);
+ /* Prevent reordering of the load of lock->__status above and the
+ load of *ptr below, as well as reordering of *ptr between
+ several iterations of the while loop. Some processors (e.g.
+ multiprocessor Alphas) could perform such reordering even though
+ the loads are dependent. */
+ MEMORY_BARRIER();
thr = *ptr;
}
+ /* Prevent reordering of the load of lock->__status above and
+ thr->p_nextlock below */
+ MEMORY_BARRIER();
/* Remove max prio thread from waiting list. */
if (maxptr == (pthread_descr *) &lock->__status) {
/* If max prio thread is at head, remove it with compare-and-swap
thr = *maxptr;
*maxptr = thr->p_nextlock;
}
+ /* Prevent reordering of store to *maxptr above and store to thr->p_nextlock
+ below */
+ MEMORY_BARRIER();
/* Wake up the selected waiting thread */
thr->p_nextlock = NULL;
restart(thr);
} else {
res = 0;
}
+ /* Prevent reordering of store to *ptr above and store to *spinlock below */
+ MEMORY_BARRIER();
*spinlock = 0;
return res;
}
register char *stack_pointer __asm__("$30");
+/* Memory barrier; default is to do nothing */
+#define MEMORY_BARRIER() __asm__ __volatile__("mb" : : : "memory")
+
+
/* Spinlock implementation; required. */
PT_EI long int
testandset (int *spinlock)
return ret;
}
-/* Spinlock release; default is just set to zero. */
-#define RELEASE(spinlock) \
- __asm__ __volatile__("mb" : : : "memory"); \
- *spinlock = 0
-
/* Begin allocating thread stacks at this address. Default is to allocate
them just below the initial program stack. */
"2:\tmb\n"
"/* End compare & swap */"
: "=&r"(ret), "=m"(*p)
- : "r"(oldval), "r"(newval), "m"(*p));
+ : "r"(oldval), "r"(newval), "m"(*p)
+ : "memory");
return ret;
}
/* For multiprocessor systems, we want to ensure all memory accesses
are completed before we reset a lock. */
-#if 0
-/* on non multiprocessor systems, you can just: */
-#define sync() /* nothing */
-#else
-#define sync() __asm__ __volatile__ ("sync")
-#endif
+#define MEMORY_BARRIER() __asm__ __volatile__ ("sync" : : : "memory")
/* Get some notion of the current stack. Need not be exactly the top
of the stack, just something somewhere in the current frame. */
: "=&r"(ret)
: "r"(p), "r"(newval), "r"(oldval)
: "cr0", "memory");
- sync();
+ MEMORY_BARRIER();
return ret == 0;
}
}
-/* Spinlock release; default is just set to zero. */
-#define RELEASE(spinlock) \
- __asm__ __volatile__("stbar; stb %1,%0" : "=m"(*(spinlock)) : "r"(0));
+/* Memory barrier; default is to do nothing */
+#define MEMORY_BARRIER() __asm__ __volatile__("stbar" : : : "memory")
/* Get some notion of the current stack. Need not be exactly the top
}
+/* Memory barrier; default is to do nothing */
+/* FIXME: is stbar OK, or should we use the more general membar instruction?
+ If so, which mode to pass to membar? */
+#define MEMORY_BARRIER() __asm__ __volatile__("stbar" : : : "memory")
+
+
/* Get some notion of the current stack. Need not be exactly the top
of the stack, just something somewhere in the current frame. */
#define CURRENT_STACK_FRAME stack_pointer
__asm__ __volatile__ ("casx [%4], %2, %0"
: "=r"(readval), "=m"(*p)
: "r"(oldval), "m"(*p), "r"(p), "0"(newval));
-
+ MEMORY_BARRIER();
return readval == oldval;
}