Changes in / [b14ec5f:84a6e70]

Location:

libcfa/src/concurrency

Files:

• kernel.cfa (modified) (5 diffs)
• kernel.hfa (modified) (1 diff)
• kernel/startup.cfa (modified) (1 diff)

libcfa/src/concurrency/kernel.cfa

@@ -126 +126 @@
 static bool mark_idle (__cluster_proc_list & idles, processor & proc);
 static void mark_awake(__cluster_proc_list & idles, processor & proc);
+static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list idles );
 
 extern void __cfa_io_start( processor * );
@@ -765 +766 @@
 
         // Check if there is a sleeping processor
-        int fd = __atomic_load_n(&this->procs.fd, __ATOMIC_SEQ_CST);
+        processor * p;
+        unsigned idle;
+        unsigned total;
+        [idle, total, p] = query_idles(this->procs);
 
         // If no one is sleeping, we are done
-        if( fd == 0 ) return;
+        if( idle == 0 ) return;
 
         // We found a processor, wake it up
         eventfd_t val;
         val = 1;
-        eventfd_write( fd, val );
+        eventfd_write( p->idle, val );
 
         #if !defined(__CFA_NO_STATISTICS__)
@@ -809 +813 @@
                 remove(proc);
                 insert_first(this.idles, proc);
-
-                __atomic_store_n(&this.fd, proc.idle, __ATOMIC_SEQ_CST);
         unlock( this );
         /* paranoid */ verify( ! __preemption_enabled() );
@@ -824 +826 @@
                 remove(proc);
                 insert_last(this.actives, proc);
-
-                __atomic_store_n(&this.fd, this.idles`first.idle, __ATOMIC_SEQ_CST);
         unlock( this );
+        /* paranoid */ verify( ! __preemption_enabled() );
+}
+
+static [unsigned idle, unsigned total, * processor] query_idles( & __cluster_proc_list this ) {
+        /* paranoid */ verify( ! __preemption_enabled() );
+        /* paranoid */ verify( ready_schedule_islocked() );
+
+        for() {
+                uint64_t l = __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST);
+                if( 1 == (l % 2) ) { Pause(); continue; }
+                unsigned idle    = this.idle;
+                unsigned total   = this.total;
+                processor * proc = &this.idles`first;
+                // Compiler fence is unnecessary, but gcc-8 and older incorrectly reorder code without it
+                asm volatile("": : :"memory");
+                if(l != __atomic_load_n(&this.lock, __ATOMIC_SEQ_CST)) { Pause(); continue; }
+                return [idle, total, proc];
+        }
+
+        /* paranoid */ verify( ready_schedule_islocked() );
         /* paranoid */ verify( ! __preemption_enabled() );
 }
@@ -888 +908 @@
                 if(head == tail) return false;
                 #if OLD_MAIN
-                        ready_schedule_lock();
-                        ret = __cfa_io_drain( proc );
-                        ready_schedule_unlock();
+                ready_schedule_lock();
+                ret = __cfa_io_drain( proc );
+                ready_schedule_unlock();
                 #else
                         ret = __cfa_io_drain( proc );
-                #endif
+        #endif
         #endif
         return ret;

libcfa/src/concurrency/kernel.hfa

@@ -195 +195 @@
 struct __cluster_proc_list {
         // Spin lock protecting the queue
-        __spinlock_t lock;
-
-        // FD to use to wake a processor
-        volatile int fd;
+        volatile uint64_t lock;
 
         // Total number of processors

libcfa/src/concurrency/kernel/startup.cfa

@@ -584 +584 @@
 // Cluster
 static void ?{}(__cluster_proc_list & this) {
-        this.fd    = 0;
+        this.lock  = 0;
         this.idle  = 0;
         this.total = 0;