Changes in libcfa/src/concurrency/kernel.cfa [c9c1c1cb:7cf3b1d]

File:

• libcfa/src/concurrency/kernel.cfa (modified) (13 diffs)

libcfa/src/concurrency/kernel.cfa

@@ -42 +42 @@
 
 #if !defined(__CFA_NO_STATISTICS__)
-        #define __STATS_DEF( ...) __VA_ARGS__
+        #define __STATS( ...) __VA_ARGS__
 #else
-        #define __STATS_DEF( ...)
+        #define __STATS( ...)
 #endif
 
@@ -122 +122 @@
 static thread$ * __next_thread(cluster * this);
 static thread$ * __next_thread_slow(cluster * this);
-static thread$ * __next_thread_search(cluster * this);
 static inline bool __must_unpark( thread$ * thrd ) __attribute((nonnull(1)));
 static void __run_thread(processor * this, thread$ * dst);
@@ -188 +187 @@
                 MAIN_LOOP:
                 for() {
+                        #define OLD_MAIN 1
+                        #if OLD_MAIN
                         // Check if there is pending io
                         __maybe_io_drain( this );
@@ -195 +196 @@
 
                         if( !readyThread ) {
-                                __IO_STATS__(true, io.flush.idle++; )
                                 __cfa_io_flush( this, 0 );
 
-                                readyThread = __next_thread( this->cltr );
-                        }
-
-                        if( !readyThread ) for(5) {
-                                __IO_STATS__(true, io.flush.idle++; )
-
                                 readyThread = __next_thread_slow( this->cltr );
-
-                                if( readyThread ) break;
-
-                                __cfa_io_flush( this, 0 );
                         }
 
@@ -220 +210 @@
 
                                 // Confirm the ready-queue is empty
-                                readyThread = __next_thread_search( this->cltr );
+                                readyThread = __next_thread_slow( this->cltr );
                                 if( readyThread ) {
                                         // A thread was found, cancel the halt
                                         mark_awake(this->cltr->procs, * this);
 
-                                        __STATS__(true, ready.sleep.cancels++; )
+                                        #if !defined(__CFA_NO_STATISTICS__)
+                                                __tls_stats()->ready.sleep.cancels++;
+                                        #endif
 
                                         // continue the mai loop
@@ -252 +244 @@
 
                         if(this->io.pending && !this->io.dirty) {
-                                __IO_STATS__(true, io.flush.dirty++; )
                                 __cfa_io_flush( this, 0 );
                         }
+
+                        #else
+                                #warning new kernel loop
+                        SEARCH: {
+                                /* paranoid */ verify( ! __preemption_enabled() );
+
+                                // First, lock the scheduler since we are searching for a thread
+                                ready_schedule_lock();
+
+                                // Try to get the next thread
+                                readyThread = pop_fast( this->cltr );
+                                if(readyThread) { ready_schedule_unlock(); break SEARCH; }
+
+                                // If we can't find a thread, might as well flush any outstanding I/O
+                                if(this->io.pending) { __cfa_io_flush( this, 0 ); }
+
+                                // Spin a little on I/O, just in case
+                                for(5) {
+                                        __maybe_io_drain( this );
+                                        readyThread = pop_fast( this->cltr );
+                                        if(readyThread) { ready_schedule_unlock(); break SEARCH; }
+                                }
+
+                                // no luck, try stealing a few times
+                                for(5) {
+                                        if( __maybe_io_drain( this ) ) {
+                                                readyThread = pop_fast( this->cltr );
+                                        } else {
+                                                readyThread = pop_slow( this->cltr );
+                                        }
+                                        if(readyThread) { ready_schedule_unlock(); break SEARCH; }
+                                }
+
+                                // still no luck, search for a thread
+                                readyThread = pop_search( this->cltr );
+                                if(readyThread) { ready_schedule_unlock(); break SEARCH; }
+
+                                // Don't block if we are done
+                                if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) {
+                                        ready_schedule_unlock();
+                                        break MAIN_LOOP;
+                                }
+
+                                __STATS( __tls_stats()->ready.sleep.halts++; )
+
+                                // Push self to idle stack
+                                ready_schedule_unlock();
+                                if(!mark_idle(this->cltr->procs, * this)) goto SEARCH;
+                                ready_schedule_lock();
+
+                                // Confirm the ready-queue is empty
+                                __maybe_io_drain( this );
+                                readyThread = pop_search( this->cltr );
+                                ready_schedule_unlock();
+
+                                if( readyThread ) {
+                                        // A thread was found, cancel the halt
+                                        mark_awake(this->cltr->procs, * this);
+
+                                        __STATS( __tls_stats()->ready.sleep.cancels++; )
+
+                                        // continue the main loop
+                                        break SEARCH;
+                                }
+
+                                __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 0\n", this->unique_id, rdtscl()); )
+                                __cfadbg_print_safe(runtime_core, "Kernel : core %p waiting on eventfd %d\n", this, this->idle_fd);
+
+                                {
+                                        eventfd_t val;
+                                        ssize_t ret = read( this->idle_fd, &val, sizeof(val) );
+                                        if(ret < 0) {
+                                                switch((int)errno) {
+                                                case EAGAIN:
+                                                #if EAGAIN != EWOULDBLOCK
+                                                        case EWOULDBLOCK:
+                                                #endif
+                                                case EINTR:
+                                                        // No need to do anything special here, just assume it's a legitimate wake-up
+                                                        break;
+                                                default:
+                                                        abort( "KERNEL : internal error, read failure on idle eventfd, error(%d) %s.", (int)errno, strerror( (int)errno ) );
+                                                }
+                                        }
+                                }
+
+                                        __STATS( if(this->print_halts) __cfaabi_bits_print_safe( STDOUT_FILENO, "PH:%d - %lld 1\n", this->unique_id, rdtscl()); )
+
+                                // We were woken up, remove self from idle
+                                mark_awake(this->cltr->procs, * this);
+
+                                // DON'T just proceed, start looking again
+                                continue MAIN_LOOP;
+                        }
+
+                RUN_THREAD:
+                        /* paranoid */ verify( ! __preemption_enabled() );
+                        /* paranoid */ verify( readyThread );
+
+                        // Reset io dirty bit
+                        this->io.dirty = false;
+
+                        // We found a thread run it
+                        __run_thread(this, readyThread);
+
+                        // Are we done?
+                        if( __atomic_load_n(&this->do_terminate, __ATOMIC_SEQ_CST) ) break MAIN_LOOP;
+
+                        if(this->io.pending && !this->io.dirty) {
+                                __cfa_io_flush( this, 0 );
+                        }
+
+                        ready_schedule_lock();
+                        __maybe_io_drain( this );
+                        ready_schedule_unlock();
+                        #endif
                 }
 
@@ -365 +472 @@
                                 break RUNNING;
                         case TICKET_UNBLOCK:
-                                __STATS__(true, ready.threads.threads++; )
+                                #if !defined(__CFA_NO_STATISTICS__)
+                                        __tls_stats()->ready.threads.threads++;
+                                #endif
                                 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
                                 // In this case, just run it again.
@@ -380 +489 @@
         __cfadbg_print_safe(runtime_core, "Kernel : core %p finished running thread %p\n", this, thrd_dst);
 
-        __STATS__(true, ready.threads.threads--; )
+        #if !defined(__CFA_NO_STATISTICS__)
+                __tls_stats()->ready.threads.threads--;
+        #endif
 
         /* paranoid */ verify( ! __preemption_enabled() );
@@ -391 +502 @@
         thread$ * thrd_src = kernelTLS().this_thread;
 
-        __STATS_DEF( thrd_src->last_proc = kernelTLS().this_processor; )
+        __STATS( thrd_src->last_proc = kernelTLS().this_processor; )
 
         // Run the thread on this processor
@@ -443 +554 @@
         // Dereference the thread now because once we push it, there is not guaranteed it's still valid.
         struct cluster * cl = thrd->curr_cluster;
-        __STATS_DEF(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
+        __STATS(bool outside = hint == UNPARK_LOCAL && thrd->last_proc && thrd->last_proc != kernelTLS().this_processor; )
 
         // push the thread to the cluster ready-queue
@@ -494 +605 @@
 
         ready_schedule_lock();
-                thread$ * thrd = pop_slow( this );
-        ready_schedule_unlock();
-
-        /* paranoid */ verify( ! __preemption_enabled() );
-        return thrd;
-}
-
-// KERNEL ONLY
-static inline thread$ * __next_thread_search(cluster * this) with( *this ) {
-        /* paranoid */ verify( ! __preemption_enabled() );
-
-        ready_schedule_lock();
-                thread$ * thrd = pop_search( this );
+                thread$ * thrd;
+                for(25) {
+                        thrd = pop_slow( this );
+                        if(thrd) goto RET;
+                }
+                thrd = pop_search( this );
+
+                RET:
         ready_schedule_unlock();
 
@@ -748 +854 @@
 
 static bool mark_idle(__cluster_proc_list & this, processor & proc) {
-        __STATS__(true, ready.sleep.halts++; )
+        #if !defined(__CFA_NO_STATISTICS__)
+                __tls_stats()->ready.sleep.halts++;
+        #endif
 
         proc.idle_wctx.fd = 0;
@@ -841 +949 @@
                 unsigned tail = *ctx->cq.tail;
                 if(head == tail) return false;
-                ready_schedule_lock();
-                ret = __cfa_io_drain( proc );
-                ready_schedule_unlock();
+                #if OLD_MAIN
+                        ready_schedule_lock();
+                        ret = __cfa_io_drain( proc );
+                        ready_schedule_unlock();
+                #else
+                        ret = __cfa_io_drain( proc );
+                #endif
         #endif
         return ret;