Changes in libcfa/src/concurrency/kernel.cfa [ada0246d:29cb302]

File:

• libcfa/src/concurrency/kernel.cfa (modified) (36 diffs)

libcfa/src/concurrency/kernel.cfa

@@ -118 +118 @@
 // Kernel Scheduling logic
 static $thread * __next_thread(cluster * this);
+static bool __has_next_thread(cluster * this);
 static void __run_thread(processor * this, $thread * dst);
-static $thread * __halt(processor * this);
-static bool __wake_one(cluster * cltr, bool was_empty);
 static bool __wake_proc(processor *);
+static bool __wake_one(struct __processor_id_t * id, cluster * cltr);
+static void __halt(processor * this);
 
 //-----------------------------------------------------------------------------
 // Kernel storage
-KERNEL_STORAGE(cluster,         mainCluster);
-KERNEL_STORAGE(processor,       mainProcessor);
-KERNEL_STORAGE($thread, mainThread);
-KERNEL_STORAGE(__stack_t,       mainThreadCtx);
-
-cluster     * mainCluster;
-processor   * mainProcessor;
-$thread * mainThread;
+KERNEL_STORAGE(cluster,              mainCluster);
+KERNEL_STORAGE(processor,            mainProcessor);
+KERNEL_STORAGE($thread,              mainThread);
+KERNEL_STORAGE(__stack_t,            mainThreadCtx);
+KERNEL_STORAGE(__scheduler_RWLock_t, __scheduler_lock);
+#if !defined(__CFA_NO_STATISTICS__)
+KERNEL_STORAGE(__stats_t, mainProcStats);
+#endif
+
+cluster              * mainCluster;
+processor            * mainProcessor;
+$thread              * mainThread;
+__scheduler_RWLock_t * __scheduler_lock;
 
 extern "C" {
@@ -144 +150 @@
 thread_local struct KernelThreadData kernelTLS __attribute__ ((tls_model ( "initial-exec" ))) = {
         NULL,                                                                                           // cannot use 0p
+        NULL,
         NULL,
         { 1, false, false },
@@ -190 +197 @@
 
 void ?{}( $thread & this, current_stack_info_t * info) with( this ) {
+        ticket = 1;
         state = Start;
         self_cor{ info };
@@ -197 +205 @@
         self_mon.recursion = 1;
         self_mon_p = &self_mon;
-        next = 0p;
+        link.next = 0p;
+        link.prev = 0p;
 
         node.next = 0p;
@@ -220 +229 @@
 static void * __invoke_processor(void * arg);
 
-void ?{}(processor & this, const char name[], cluster & cltr) with( this ) {
+void ?{}(processor & this, const char name[], cluster & _cltr) with( this ) {
         this.name = name;
-        this.cltr = &cltr;
+        this.cltr = &_cltr;
+        id = -1u;
         terminated{ 0 };
         destroyer = 0p;
@@ -235 +245 @@
 
         this.stack = __create_pthread( &this.kernel_thread, __invoke_processor, (void *)&this );
+        __atomic_fetch_add( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
 
         __cfadbg_print_safe(runtime_core, "Kernel : core %p created\n", &this);
@@ -254 +265 @@
 
         free( this.stack );
+
+        __atomic_fetch_sub( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
 }
 
@@ -259 +272 @@
         this.name = name;
         this.preemption_rate = preemption_rate;
+        this.nprocessors = 0;
         ready_queue{};
-        ready_queue_lock{};
 
         #if !defined(__CFA_NO_STATISTICS__)
                 print_stats = false;
+                stats = alloc();
+                __init_stats( stats );
         #endif
 
-        procs{ __get };
-        idles{ __get };
         threads{ __get };
 
@@ -277 +290 @@
 void ^?{}(cluster & this) {
         __kernel_io_shutdown( this, &this == mainCluster );
+
+        #if !defined(__CFA_NO_STATISTICS__)
+                if(this.print_stats) {
+                        __print_stats( this.stats );
+                }
+                free( this.stats );
+        #endif
 
         unregister(this);
@@ -295 +315 @@
         __cfadbg_print_safe(runtime_core, "Kernel : core %p starting\n", this);
 
-        doregister(this->cltr, this);
+        // register the processor unless it's the main thread which is handled in the boot sequence
+        if(this != mainProcessor) {
+                this->id = doregister((__processor_id_t*)this);
+                // Lock the RWlock so no-one pushes/pops while we are changing the queue
+                uint_fast32_t last_size = ready_mutate_lock();
+
+                        // Adjust the ready queue size
+                        ready_queue_grow( this->cltr );
+
+                // Unlock the RWlock
+                ready_mutate_unlock( last_size );
+        }
 
         {
@@ -308 +339 @@
                         readyThread = __next_thread( this->cltr );
 
-                        // If no ready thread
-                        if( readyThread == 0p ) {
-                                // Block until a thread is ready
-                                readyThread = __halt(this);
-                        }
-
                         // Check if we actually found a thread
                         if( readyThread ) {
                                 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
                                 /* paranoid */ verifyf( readyThread->state == Ready || readyThread->preempted != __NO_PREEMPTION, "state : %d, preempted %d\n", readyThread->state, readyThread->preempted);
-                                /* paranoid */ verifyf( readyThread->next == 0p, "Expected null got %p", readyThread->next );
+                                /* paranoid */ verifyf( readyThread->link.next == 0p, "Expected null got %p", readyThread->link.next );
+                                __builtin_prefetch( readyThread->context.SP );
 
                                 // We found a thread run it
@@ -325 +351 @@
                                 /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
                         }
+                        else {
+                                // Block until a thread is ready
+                                __halt(this);
+                        }
                 }
 
@@ -330 +360 @@
         }
 
-        unregister(this->cltr, this);
-
         V( this->terminated );
 
+        // unregister the processor unless it's the main thread which is handled in the boot sequence
+        if(this != mainProcessor) {
+                // Lock the RWlock so no-one pushes/pops while we are changing the queue
+                uint_fast32_t last_size = ready_mutate_lock();
+
+                        // Adjust the ready queue size
+                        ready_queue_shrink( this->cltr );
+
+                        // Make sure we aren't on the idle queue
+                        #if !defined(__CFA_NO_STATISTICS__)
+                                bool removed =
+                        #endif
+                        unsafe_remove( this->cltr->idles, this );
+
+                        #if !defined(__CFA_NO_STATISTICS__)
+                                if(removed) __tls_stats()->ready.sleep.exits++;
+                        #endif
+
+                // Unlock the RWlock
+                ready_mutate_unlock( last_size );
+
+                // Finally we don't need the read_lock any more
+                unregister((__processor_id_t*)this);
+        }
+        else {
+                // HACK : the coroutine context switch expects this_thread to be set
+                // and it make sense for it to be set in all other cases except here
+                // fake it
+                kernelTLS.this_thread = mainThread;
+        }
+
         __cfadbg_print_safe(runtime_core, "Kernel : core %p terminated\n", this);
-
-        // HACK : the coroutine context switch expects this_thread to be set
-        // and it make sense for it to be set in all other cases except here
-        // fake it
-        if( this == mainProcessor ) kernelTLS.this_thread = mainThread;
 }
 
@@ -360 +414 @@
         // Actually run the thread
         RUNNING:  while(true) {
-                if(unlikely(thrd_dst->preempted)) {
-                        thrd_dst->preempted = __NO_PREEMPTION;
-                        verify(thrd_dst->state == Active  || thrd_dst->state == Rerun);
-                } else {
-                        verify(thrd_dst->state == Blocked || thrd_dst->state == Ready); // Ready means scheduled normally, blocked means rerun
-                        thrd_dst->state = Active;
-                }
+                thrd_dst->preempted = __NO_PREEMPTION;
+                thrd_dst->state = Active;
 
                 __cfaabi_dbg_debug_do(
@@ -398 +447 @@
                 if(unlikely(thrd_dst->preempted != __NO_PREEMPTION)) {
                         // The thread was preempted, reschedule it and reset the flag
-                        __schedule_thread( thrd_dst );
+                        __schedule_thread( (__processor_id_t*)this, thrd_dst );
                         break RUNNING;
                 }
 
+                if(unlikely(thrd_dst->state == Halted)) {
+                        // The thread has halted, it should never be scheduled/run again
+                        // We may need to wake someone up here since
+                        unpark( this->destroyer __cfaabi_dbg_ctx2 );
+                        this->destroyer = 0p;
+                        break RUNNING;
+                }
+
+                /* paranoid */ verify( thrd_dst->state == Active );
+                thrd_dst->state = Blocked;
+
                 // set state of processor coroutine to active and the thread to inactive
-                static_assert(sizeof(thrd_dst->state) == sizeof(int));
-                enum coroutine_state old_state = __atomic_exchange_n(&thrd_dst->state, Blocked, __ATOMIC_SEQ_CST);
-                __cfaabi_dbg_debug_do( thrd_dst->park_result = old_state; )
-                switch(old_state) {
-                        case Halted:
-                                // The thread has halted, it should never be scheduled/run again, leave it back to Halted and move on
-                                thrd_dst->state = Halted;
-
-                                // We may need to wake someone up here since
-                                unpark( this->destroyer __cfaabi_dbg_ctx2 );
-                                this->destroyer = 0p;
-                                break RUNNING;
-                        case Active:
+                int old_ticket = __atomic_fetch_sub(&thrd_dst->ticket, 1, __ATOMIC_SEQ_CST);
+                __cfaabi_dbg_debug_do( thrd_dst->park_result = old_ticket; )
+                switch(old_ticket) {
+                        case 1:
                                 // This is case 1, the regular case, nothing more is needed
                                 break RUNNING;
-                        case Rerun:
+                        case 2:
                                 // This is case 2, the racy case, someone tried to run this thread before it finished blocking
                                 // In this case, just run it again.
@@ -424 +475 @@
                         default:
                                 // This makes no sense, something is wrong abort
-                                abort("Finished running a thread that was Blocked/Start/Primed %d\n", old_state);
+                                abort();
                 }
         }
@@ -438 +489 @@
         $coroutine * proc_cor = get_coroutine(kernelTLS.this_processor->runner);
         $thread * thrd_src = kernelTLS.this_thread;
+
+        #if !defined(__CFA_NO_STATISTICS__)
+                struct processor * last_proc = kernelTLS.this_processor;
+        #endif
 
         // Run the thread on this processor
@@ -453 +508 @@
         }
 
+        #if !defined(__CFA_NO_STATISTICS__)
+                if(last_proc != kernelTLS.this_processor) {
+                        __tls_stats()->ready.threads.migration++;
+                }
+        #endif
+
         /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
         /* paranoid */ verifyf( ((uintptr_t)thrd_src->context.SP) < ((uintptr_t)__get_stack(thrd_src->curr_cor)->base ), "ERROR : Returning $thread %p has been corrupted.\n StackPointer too small.\n", thrd_src );
@@ -463 +524 @@
 // It effectively constructs a coroutine by stealing the pthread stack
 static void * __invoke_processor(void * arg) {
+        #if !defined( __CFA_NO_STATISTICS__ )
+                __stats_t local_stats;
+                __init_stats( &local_stats );
+                kernelTLS.this_stats = &local_stats;
+        #endif
+
         processor * proc = (processor *) arg;
         kernelTLS.this_processor = proc;
@@ -494 +561 @@
         __cfadbg_print_safe(runtime_core, "Kernel : core %p main ended (%p)\n", proc, &proc->runner);
 
+        #if !defined(__CFA_NO_STATISTICS__)
+                __tally_stats(proc->cltr->stats, &local_stats);
+        #endif
+
         return 0p;
 }
@@ -591 +662 @@
 // Scheduler routines
 // KERNEL ONLY
-void __schedule_thread( $thread * thrd ) with( *thrd->curr_cluster ) {
+void __schedule_thread( struct __processor_id_t * id, $thread * thrd ) {
+        /* paranoid */ verify( thrd );
+        /* paranoid */ verify( thrd->state != Halted );
         /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
         /* paranoid */ #if defined( __CFA_WITH_VERIFY__ )
-        /* paranoid */ if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
-                          "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
-        /* paranoid */ if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active || thrd->state == Rerun,
-                          "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
+        /* paranoid */  if( thrd->state == Blocked || thrd->state == Start ) assertf( thrd->preempted == __NO_PREEMPTION,
+                                        "Error inactive thread marked as preempted, state %d, preemption %d\n", thrd->state, thrd->preempted );
+        /* paranoid */  if( thrd->preempted != __NO_PREEMPTION ) assertf(thrd->state == Active,
+                                        "Error preempted thread marked as not currently running, state %d, preemption %d\n", thrd->state, thrd->preempted );
         /* paranoid */ #endif
-        /* paranoid */ verifyf( thrd->next == 0p, "Expected null got %p", thrd->next );
+        /* paranoid */ verifyf( thrd->link.next == 0p, "Expected null got %p", thrd->link.next );
 
         if (thrd->preempted == __NO_PREEMPTION) thrd->state = Ready;
 
-        lock  ( ready_queue_lock __cfaabi_dbg_ctx2 );
-        bool was_empty = !(ready_queue != 0);
-        append( ready_queue, thrd );
-        unlock( ready_queue_lock );
-
-        __wake_one(thrd->curr_cluster, was_empty);
+        ready_schedule_lock  ( id );
+                push( thrd->curr_cluster, thrd );
+
+                #if !defined(__CFA_NO_STATISTICS__)
+                        bool woke =
+                #endif
+                        __wake_one(id, thrd->curr_cluster);
+
+                #if !defined(__CFA_NO_STATISTICS__)
+                        if(woke) __tls_stats()->ready.sleep.wakes++;
+                #endif
+        ready_schedule_unlock( id );
 
         /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
@@ -617 +696 @@
         /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
 
-        lock( ready_queue_lock __cfaabi_dbg_ctx2 );
-        $thread * head = pop_head( ready_queue );
-        unlock( ready_queue_lock );
+        ready_schedule_lock  ( (__processor_id_t*)kernelTLS.this_processor );
+                $thread * head = pop( this );
+        ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );
 
         /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
@@ -625 +704 @@
 }
 
+// KERNEL ONLY
+static bool __has_next_thread(cluster * this) with( *this ) {
+        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
+
+        ready_schedule_lock  ( (__processor_id_t*)kernelTLS.this_processor );
+                bool not_empty = query( this );
+        ready_schedule_unlock( (__processor_id_t*)kernelTLS.this_processor );
+
+        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
+        return not_empty;
+}
+
 // KERNEL ONLY unpark with out disabling interrupts
-void __unpark( $thread * thrd __cfaabi_dbg_ctx_param2 ) {
-        static_assert(sizeof(thrd->state) == sizeof(int));
-
+void __unpark(  struct __processor_id_t * id, $thread * thrd __cfaabi_dbg_ctx_param2 ) {
         // record activity
         __cfaabi_dbg_debug_do( char * old_caller = thrd->unpark_caller; )
         __cfaabi_dbg_record_thrd( *thrd, false, caller );
 
-        enum coroutine_state old_state = __atomic_exchange_n(&thrd->state, Rerun, __ATOMIC_SEQ_CST);
-        __cfaabi_dbg_debug_do( thrd->unpark_result = old_state; )
-        switch(old_state) {
-                case Active:
+        int old_ticket = __atomic_fetch_add(&thrd->ticket, 1, __ATOMIC_SEQ_CST);
+        __cfaabi_dbg_debug_do( thrd->unpark_result = old_ticket; thrd->unpark_state = thrd->state; )
+        switch(old_ticket) {
+                case 1:
                         // Wake won the race, the thread will reschedule/rerun itself
                         break;
-                case Blocked:
+                case 0:
                         /* paranoid */ verify( ! thrd->preempted != __NO_PREEMPTION );
+                        /* paranoid */ verify( thrd->state == Blocked );
 
                         // Wake lost the race,
-                        thrd->state = Blocked;
-                        __schedule_thread( thrd );
+                        __schedule_thread( id, thrd );
                         break;
-                case Rerun:
-                        abort("More than one thread attempted to schedule thread %p\n", thrd);
-                        break;
-                case Halted:
-                case Start:
-                case Primed:
                 default:
                         // This makes no sense, something is wrong abort
@@ -662 +745 @@
 
         disable_interrupts();
-        __unpark( thrd __cfaabi_dbg_ctx_fwd2 );
+        __unpark( (__processor_id_t*)kernelTLS.this_processor, thrd __cfaabi_dbg_ctx_fwd2 );
         enable_interrupts( __cfaabi_dbg_ctx );
 }
@@ -697 +780 @@
 
         $thread * thrd = kernelTLS.this_thread;
-        /* paranoid */ verify(thrd->state == Active || thrd->state == Rerun);
+        /* paranoid */ verify(thrd->state == Active);
 
         // SKULLDUGGERY: It is possible that we are preempting this thread just before
@@ -704 +787 @@
         // If that is the case, abandon the preemption.
         bool preempted = false;
-        if(thrd->next == 0p) {
+        if(thrd->link.next == 0p) {
                 preempted = true;
                 thrd->preempted = reason;
@@ -730 +813 @@
         __cfa_dbg_global_clusters.list{ __get };
         __cfa_dbg_global_clusters.lock{};
+
+        // Initialize the global scheduler lock
+        __scheduler_lock = (__scheduler_RWLock_t*)&storage___scheduler_lock;
+        (*__scheduler_lock){};
 
         // Initialize the main cluster
@@ -764 +851 @@
                 pending_preemption = false;
                 kernel_thread = pthread_self();
+                id = -1u;
 
                 runner{ &this };
                 __cfadbg_print_safe(runtime_core, "Kernel : constructed main processor context %p\n", &runner);
+
+                __atomic_fetch_add( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
         }
 
@@ -774 +864 @@
         (*mainProcessor){};
 
+        mainProcessor->id = doregister( (__processor_id_t*)mainProcessor);
+
         //initialize the global state variables
         kernelTLS.this_processor = mainProcessor;
         kernelTLS.this_thread    = mainThread;
 
+        #if !defined( __CFA_NO_STATISTICS__ )
+                kernelTLS.this_stats = (__stats_t *)& storage_mainProcStats;
+                __init_stats( kernelTLS.this_stats );
+        #endif
+
         // Enable preemption
         kernel_start_preemption();
@@ -783 +880 @@
         // Add the main thread to the ready queue
         // once resume is called on mainProcessor->runner the mainThread needs to be scheduled like any normal thread
-        __schedule_thread(mainThread);
+        __schedule_thread((__processor_id_t *)mainProcessor, mainThread);
 
         // SKULLDUGGERY: Force a context switch to the main processor to set the main thread's context to the current UNIX
@@ -827 +924 @@
         kernel_stop_preemption();
 
+        unregister((__processor_id_t*)mainProcessor);
+
         // Destroy the main processor and its context in reverse order of construction
         // These were manually constructed so we need manually destroy them
         void ^?{}(processor & this) with( this ){
                 /* paranoid */ verify( this.do_terminate == true );
+                __atomic_fetch_sub( &cltr->nprocessors, 1u, __ATOMIC_SEQ_CST );
+                __cfaabi_dbg_print_safe("Kernel : destroyed main processor context %p\n", &runner);
         }
 
@@ -836 +937 @@
 
         // Final step, destroy the main thread since it is no longer needed
+
         // Since we provided a stack to this taxk it will not destroy anything
         /* paranoid */ verify(mainThread->self_cor.stack.storage == (__stack_t*)(((uintptr_t)&storage_mainThreadCtx)| 0x1));
@@ -842 +944 @@
         ^(*mainCluster){};
 
+        ^(*__scheduler_lock){};
+
         ^(__cfa_dbg_global_clusters.list){};
         ^(__cfa_dbg_global_clusters.lock){};
@@ -851 +955 @@
 // Kernel Idle Sleep
 //=============================================================================================
-static $thread * __halt(processor * this) with( *this ) {
-        if( do_terminate ) return 0p;
-
-        // First, lock the cluster idle
-        lock( cltr->idle_lock __cfaabi_dbg_ctx2 );
-
-        // Check if we can find a thread
-        if( $thread * found = __next_thread( cltr ) ) {
-                unlock( cltr->idle_lock );
-                return found;
-        }
-
-        // Move this processor from the active list to the idle list
-        move_to_front(cltr->procs, cltr->idles, *this);
-
-        // Unlock the idle lock so we don't go to sleep with a lock
-        unlock    (cltr->idle_lock);
-
-        // We are ready to sleep
-        __cfadbg_print_safe(runtime_core, "Kernel : Processor %p ready to sleep\n", this);
-        wait( idle );
-
-        // We have woken up
-        __cfadbg_print_safe(runtime_core, "Kernel : Processor %p woke up and ready to run\n", this);
-
-        // Get ourself off the idle list
-        with( *cltr ) {
-                lock  (idle_lock __cfaabi_dbg_ctx2);
-                move_to_front(idles, procs, *this);
-                unlock(idle_lock);
-        }
-
-        // Don't check the ready queue again, we may not be in a position to run a thread
-        return 0p;
-}
-
 // Wake a thread from the front if there are any
-static bool __wake_one(cluster * this, __attribute__((unused)) bool force) {
-        // if we don't want to force check if we know it's false
-        // if( !this->idles.head && !force ) return false;
-
-        // First, lock the cluster idle
-        lock( this->idle_lock __cfaabi_dbg_ctx2 );
-
-        // Check if there is someone to wake up
-        if( !this->idles.head ) {
-                // Nope unlock and return false
-                unlock( this->idle_lock );
-                return false;
-        }
-
-        // Wake them up
-        __cfadbg_print_safe(runtime_core, "Kernel : waking Processor %p\n", this->idles.head);
-        /* paranoid */ verify( ! kernelTLS.preemption_state.enabled );
-        post( this->idles.head->idle );
-
-        // Unlock and return true
-        unlock( this->idle_lock );
+static bool __wake_one(struct __processor_id_t * id, cluster * this) {
+        /* paranoid */ verify( ready_schedule_islocked( id ) );
+
+        // Check if there is a sleeping processor
+        processor * p = pop(this->idles);
+
+        // If no one is sleeping, we are done
+        if( 0p == p ) return false;
+
+        // We found a processor, wake it up
+        post( p->idle );
+
         return true;
 }
@@ -922 +981 @@
 
         return ret;
+}
+
+static void __halt(processor * this) with( *this ) {
+        if( do_terminate ) return;
+
+        #if !defined(__CFA_NO_STATISTICS__)
+                __tls_stats()->ready.sleep.halts++;
+        #endif
+        // Push self to queue
+        push(cltr->idles, *this);
+
+        // Makre sure we don't miss a thread
+        if( __has_next_thread(cltr) ) {
+                // A thread was posted, make sure a processor is woken up
+                struct __processor_id_t *id = (struct __processor_id_t *) this;
+                ready_schedule_lock  ( id );
+                        __wake_one( id, cltr );
+                ready_schedule_unlock( id );
+                #if !defined(__CFA_NO_STATISTICS__)
+                        __tls_stats()->ready.sleep.cancels++;
+                #endif
+        }
+
+        wait( idle );
 }
 
@@ -1078 +1161 @@
         cltr->nthreads -= 1;
         unlock(cltr->thread_list_lock);
-}
-
-void doregister( cluster * cltr, processor * proc ) {
-        lock      (cltr->idle_lock __cfaabi_dbg_ctx2);
-        cltr->nprocessors += 1;
-        push_front(cltr->procs, *proc);
-        unlock    (cltr->idle_lock);
-}
-
-void unregister( cluster * cltr, processor * proc ) {
-        lock  (cltr->idle_lock __cfaabi_dbg_ctx2);
-        remove(cltr->procs, *proc );
-        cltr->nprocessors -= 1;
-        unlock(cltr->idle_lock);
 }