X-Git-Url: http://git.megacz.com/?a=blobdiff_plain;f=ghc%2Frts%2FSchedule.c;h=5451cc5b33ff28b8527e40af9445d28c9ba61e85;hb=50027272414438955dbc41696541cbd25da55883;hp=ccb6b74dcc80771b71cf4d7a60450ba106523402;hpb=892b901751581f2e73199e8c5e24731eebd67406;p=ghc-hetmet.git diff --git a/ghc/rts/Schedule.c b/ghc/rts/Schedule.c index ccb6b74..5451cc5 100644 --- a/ghc/rts/Schedule.c +++ b/ghc/rts/Schedule.c @@ -1,11 +1,78 @@ -/* ----------------------------------------------------------------------------- - * $Id: Schedule.c,v 1.11 1999/02/26 16:44:13 simonm Exp $ +/* --------------------------------------------------------------------------- + * $Id: Schedule.c,v 1.95 2001/03/23 16:36:21 simonmar Exp $ * - * (c) The GHC Team, 1998-1999 + * (c) The GHC Team, 1998-2000 * * Scheduler * - * ---------------------------------------------------------------------------*/ + * Different GHC ways use this scheduler quite differently (see comments below) + * Here is the global picture: + * + * WAY Name CPP flag What's it for + * -------------------------------------- + * mp GUM PAR Parallel execution on a distributed memory machine + * s SMP SMP Parallel execution on a shared memory machine + * mg GranSim GRAN Simulation of parallel execution + * md GUM/GdH DIST Distributed execution (based on GUM) + * --------------------------------------------------------------------------*/ + +//@node Main scheduling code, , , +//@section Main scheduling code + +/* + * Version with scheduler monitor support for SMPs (WAY=s): + + This design provides a high-level API to create and schedule threads etc. + as documented in the SMP design document. + + It uses a monitor design controlled by a single mutex to exercise control + over accesses to shared data structures, and builds on the Posix threads + library. + + The majority of state is shared. In order to keep essential per-task state, + there is a Capability structure, which contains all the information + needed to run a thread: its STG registers, a pointer to its TSO, a + nursery etc. During STG execution, a pointer to the capability is + kept in a register (BaseReg). + + In a non-SMP build, there is one global capability, namely MainRegTable. + + SDM & KH, 10/99 + + * Version with support for distributed memory parallelism aka GUM (WAY=mp): + + The main scheduling loop in GUM iterates until a finish message is received. + In that case a global flag @receivedFinish@ is set and this instance of + the RTS shuts down. See ghc/rts/parallel/HLComms.c:processMessages() + for the handling of incoming messages, such as PP_FINISH. + Note that in the parallel case we have a system manager that coordinates + different PEs, each of which are running one instance of the RTS. + See ghc/rts/parallel/SysMan.c for the main routine of the parallel program. + From this routine processes executing ghc/rts/Main.c are spawned. -- HWL + + * Version with support for simulating parallel execution aka GranSim (WAY=mg): + + The main scheduling code in GranSim is quite different from that in std + (concurrent) Haskell: while concurrent Haskell just iterates over the + threads in the runnable queue, GranSim is event driven, i.e. it iterates + over the events in the global event queue. -- HWL +*/ + +//@menu +//* Includes:: +//* Variables and Data structures:: +//* Main scheduling loop:: +//* Suspend and Resume:: +//* Run queue code:: +//* Garbage Collextion Routines:: +//* Blocking Queue Routines:: +//* Exception Handling Routines:: +//* Debugging Routines:: +//* Index:: +//@end menu + +//@node Includes, Variables and Data structures, Main scheduling code, Main scheduling code +//@subsection Includes #include "Rts.h" #include "SchedAPI.h" @@ -19,30 +86,121 @@ #include "Schedule.h" #include "StgMiscClosures.h" #include "Storage.h" -#include "Evaluator.h" +#include "Interpreter.h" +#include "Exception.h" #include "Printer.h" #include "Main.h" #include "Signals.h" -#include "Profiling.h" #include "Sanity.h" +#include "Stats.h" +#include "Itimer.h" +#include "Prelude.h" +#if defined(GRAN) || defined(PAR) +# include "GranSimRts.h" +# include "GranSim.h" +# include "ParallelRts.h" +# include "Parallel.h" +# include "ParallelDebug.h" +# include "FetchMe.h" +# include "HLC.h" +#endif +#include "Sparks.h" + +#include + +//@node Variables and Data structures, Prototypes, Includes, Main scheduling code +//@subsection Variables and Data structures + +/* Main threads: + * + * These are the threads which clients have requested that we run. + * + * In an SMP build, we might have several concurrent clients all + * waiting for results, and each one will wait on a condition variable + * until the result is available. + * + * In non-SMP, clients are strictly nested: the first client calls + * into the RTS, which might call out again to C with a _ccall_GC, and + * eventually re-enter the RTS. + * + * Main threads information is kept in a linked list: + */ +//@cindex StgMainThread +typedef struct StgMainThread_ { + StgTSO * tso; + SchedulerStatus stat; + StgClosure ** ret; +#ifdef SMP + pthread_cond_t wakeup; +#endif + struct StgMainThread_ *link; +} StgMainThread; + +/* Main thread queue. + * Locks required: sched_mutex. + */ +static StgMainThread *main_threads; + +/* Thread queues. + * Locks required: sched_mutex. + */ +#if defined(GRAN) + +StgTSO* ActiveTSO = NULL; /* for assigning system costs; GranSim-Light only */ +/* rtsTime TimeOfNextEvent, EndOfTimeSlice; now in GranSim.c */ + +/* + In GranSim we have a runable and a blocked queue for each processor. + In order to minimise code changes new arrays run_queue_hds/tls + are created. run_queue_hd is then a short cut (macro) for + run_queue_hds[CurrentProc] (see GranSim.h). + -- HWL +*/ +StgTSO *run_queue_hds[MAX_PROC], *run_queue_tls[MAX_PROC]; +StgTSO *blocked_queue_hds[MAX_PROC], *blocked_queue_tls[MAX_PROC]; +StgTSO *ccalling_threadss[MAX_PROC]; +/* We use the same global list of threads (all_threads) in GranSim as in + the std RTS (i.e. we are cheating). However, we don't use this list in + the GranSim specific code at the moment (so we are only potentially + cheating). */ + +#else /* !GRAN */ StgTSO *run_queue_hd, *run_queue_tl; StgTSO *blocked_queue_hd, *blocked_queue_tl; -StgTSO *ccalling_threads; +StgTSO *sleeping_queue; /* perhaps replace with a hash table? */ -#define MAX_SCHEDULE_NESTING 256 -nat next_main_thread; -StgTSO *main_threads[MAX_SCHEDULE_NESTING]; +#endif + +/* Linked list of all threads. + * Used for detecting garbage collected threads. + */ +StgTSO *all_threads; + +/* Threads suspended in _ccall_GC. + */ +static StgTSO *suspended_ccalling_threads; static void GetRoots(void); static StgTSO *threadStackOverflow(StgTSO *tso); +/* KH: The following two flags are shared memory locations. There is no need + to lock them, since they are only unset at the end of a scheduler + operation. +*/ + /* flag set by signal handler to precipitate a context switch */ +//@cindex context_switch nat context_switch; + /* if this flag is set as well, give up execution */ -static nat interrupted; +//@cindex interrupted +rtsBool interrupted; -/* Next thread ID to allocate */ +/* Next thread ID to allocate. + * Locks required: sched_mutex + */ +//@cindex next_thread_id StgThreadID next_thread_id = 1; /* @@ -50,14 +208,7 @@ StgThreadID next_thread_id = 1; * Rule of thumb: if CurrentTSO != NULL, then we're running a Haskell * thread. If CurrentTSO == NULL, then we're at the scheduler level. */ -StgTSO *CurrentTSO; -StgRegTable MainRegTable; - -/* - * The thread state for the main thread. - */ -StgTSO *MainTSO; - + /* The smallest stack size that makes any sense is: * RESERVED_STACK_WORDS (so we can get back from the stack overflow) * + sizeofW(StgStopFrame) (the stg_stop_thread_info frame) @@ -70,691 +221,3287 @@ StgTSO *MainTSO; #define MIN_STACK_WORDS (RESERVED_STACK_WORDS + sizeofW(StgStopFrame) + 2) -/* ----------------------------------------------------------------------------- - Create a new thread. - - The new thread starts with the given stack size. Before the - scheduler can run, however, this thread needs to have a closure - (and possibly some arguments) pushed on its stack. See - pushClosure() in Schedule.h. - - createGenThread() and createIOThread() (in Schedule.h) are - convenient packaged versions of this function. - -------------------------------------------------------------------------- */ - -StgTSO * -createThread(nat stack_size) -{ - StgTSO *tso; - - /* catch ridiculously small stack sizes */ - if (stack_size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) { - stack_size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW; - } - - tso = (StgTSO *)allocate(stack_size); - TICK_ALLOC_TSO(stack_size-sizeofW(StgTSO),0); - - initThread(tso, stack_size - TSO_STRUCT_SIZEW); - return tso; -} - -void -initThread(StgTSO *tso, nat stack_size) -{ - SET_INFO(tso,&TSO_info); - tso->whatNext = ThreadEnterGHC; - tso->state = tso_state_runnable; - tso->id = next_thread_id++; - - tso->splim = (P_)&(tso->stack) + RESERVED_STACK_WORDS; - tso->stack_size = stack_size; - tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize) - - TSO_STRUCT_SIZEW; - tso->sp = (P_)&(tso->stack) + stack_size; - -#ifdef PROFILING - tso->prof.CCCS = CCS_MAIN; +/* Free capability list. + * Locks required: sched_mutex. + */ +#ifdef SMP +//@cindex free_capabilities +//@cindex n_free_capabilities +Capability *free_capabilities; /* Available capabilities for running threads */ +nat n_free_capabilities; /* total number of available capabilities */ +#else +//@cindex MainRegTable +Capability MainRegTable; /* for non-SMP, we have one global capability */ #endif - /* put a stop frame on the stack */ - tso->sp -= sizeofW(StgStopFrame); - SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_MAIN); - tso->su = (StgUpdateFrame*)tso->sp; - - IF_DEBUG(scheduler,belch("Initialised thread %ld, stack size = %lx words\n", - tso->id, tso->stack_size)); - - /* Put the new thread on the head of the runnable queue. - * The caller of createThread better push an appropriate closure - * on this thread's stack before the scheduler is invoked. - */ - tso->link = run_queue_hd; - run_queue_hd = tso; - if (run_queue_tl == END_TSO_QUEUE) { - run_queue_tl = tso; - } - - IF_DEBUG(scheduler,printTSO(tso)); -} +#if defined(GRAN) +StgTSO *CurrentTSO; +#endif -/* ----------------------------------------------------------------------------- - Delete a thread - reverting all blackholes to (something - equivalent to) their former state. - - We create an AP_UPD for every UpdateFrame on the stack. - Entering one of these AP_UPDs pushes everything from the corresponding - update frame upwards onto the stack. (Actually, it pushes everything - up to the next update frame plus a pointer to the next AP_UPD - object. Entering the next AP_UPD object pushes more onto the - stack until we reach the last AP_UPD object - at which point - the stack should look exactly as it did when we killed the TSO - and we can continue execution by entering the closure on top of - the stack. - -------------------------------------------------------------------------- */ +/* This is used in `TSO.h' and gcc 2.96 insists that this variable actually + * exists - earlier gccs apparently didn't. + * -= chak + */ +StgTSO dummy_tso; -void deleteThread(StgTSO *tso) -{ - StgUpdateFrame* su = tso->su; - StgPtr sp = tso->sp; +rtsBool ready_to_gc; - /* Thread already dead? */ - if (tso->whatNext == ThreadComplete || tso->whatNext == ThreadKilled) { - return; - } +/* All our current task ids, saved in case we need to kill them later. + */ +#ifdef SMP +//@cindex task_ids +task_info *task_ids; +#endif - IF_DEBUG(scheduler, belch("Killing thread %ld.", tso->id)); +void addToBlockedQueue ( StgTSO *tso ); - tso->whatNext = ThreadKilled; /* changed to ThreadComplete in schedule() */ - tso->link = END_TSO_QUEUE; /* Just to be on the safe side... */ +static void schedule ( void ); + void interruptStgRts ( void ); +#if defined(GRAN) +static StgTSO * createThread_ ( nat size, rtsBool have_lock, StgInt pri ); +#else +static StgTSO * createThread_ ( nat size, rtsBool have_lock ); +#endif - /* Threads that finish normally leave Su pointing to the word - * beyond the top of the stack, and Sp pointing to the last word - * on the stack, which is the return value of the thread. - */ - if ((P_)tso->su >= tso->stack + tso->stack_size - || get_itbl(tso->su)->type == STOP_FRAME) { - return; - } - - IF_DEBUG(scheduler, - fprintf(stderr, "Freezing TSO stack\n"); - printTSO(tso); - ); - - /* The stack freezing code assumes there's a closure pointer on - * the top of the stack. This isn't always the case with compiled - * code, so we have to push a dummy closure on the top which just - * returns to the next return address on the stack. - */ - if (LOOKS_LIKE_GHC_INFO(*sp)) { - *(--sp) = (W_)&dummy_ret_closure; - } +static void detectBlackHoles ( void ); - while (1) { - int words = (stgCast(StgPtr,su) - stgCast(StgPtr,sp)) - 1; - nat i; - StgAP_UPD* ap = stgCast(StgAP_UPD*,allocate(AP_sizeW(words))); - TICK_ALLOC_THK(words+1,0); +#ifdef DEBUG +static void sched_belch(char *s, ...); +#endif - /* First build an AP_UPD consisting of the stack chunk above the - * current update frame, with the top word on the stack as the - * fun field. - */ - ASSERT(words >= 0); - - /* if (words == 0) { -- optimisation - ap = stgCast(StgAP_UPD*,*stgCast(StgPtr*,sp)++); - } else */ { - ap->n_args = words; - ap->fun = stgCast(StgClosure*,*stgCast(StgPtr*,sp)++); - for(i=0; i < (nat)words; ++i) { - payloadWord(ap,i) = *sp++; - } - } +#ifdef SMP +//@cindex sched_mutex +//@cindex term_mutex +//@cindex thread_ready_cond +//@cindex gc_pending_cond +pthread_mutex_t sched_mutex = PTHREAD_MUTEX_INITIALIZER; +pthread_mutex_t term_mutex = PTHREAD_MUTEX_INITIALIZER; +pthread_cond_t thread_ready_cond = PTHREAD_COND_INITIALIZER; +pthread_cond_t gc_pending_cond = PTHREAD_COND_INITIALIZER; + +nat await_death; +#endif - switch (get_itbl(su)->type) { - - case UPDATE_FRAME: - { - SET_HDR(ap,&AP_UPD_info,su->header.prof.ccs /* ToDo */); - - IF_DEBUG(scheduler, - fprintf(stderr, "Updating "); - printPtr(stgCast(StgPtr,su->updatee)); - fprintf(stderr, " with "); - printObj(stgCast(StgClosure*,ap)); - ); - - /* Replace the updatee with an indirection - happily - * this will also wake up any threads currently - * waiting on the result. - */ - UPD_IND(su->updatee,ap); /* revert the black hole */ - su = su->link; - sp += sizeofW(StgUpdateFrame) -1; - sp[0] = stgCast(StgWord,ap); /* push onto stack */ - break; - } - - case CATCH_FRAME: - { - StgCatchFrame *cf = (StgCatchFrame *)su; - StgClosure* o; - - /* We want a PAP, not an AP_UPD. Fortunately, the - * layout's the same. - */ - SET_HDR(ap,&PAP_info,su->header.prof.ccs /* ToDo */); - - /* now build o = FUN(catch,ap,handler) */ - o = stgCast(StgClosure*, allocate(sizeofW(StgClosure)+2)); - TICK_ALLOC_THK(2,0); - SET_HDR(o,&catch_info,su->header.prof.ccs /* ToDo */); - payloadCPtr(o,0) = stgCast(StgClosure*,ap); - payloadCPtr(o,1) = cf->handler; - - IF_DEBUG(scheduler, - fprintf(stderr, "Built "); - printObj(stgCast(StgClosure*,o)); - ); - - /* pop the old handler and put o on the stack */ - su = cf->link; - sp += sizeofW(StgCatchFrame) - 1; - sp[0] = (W_)o; - break; - } - - case SEQ_FRAME: - { - StgSeqFrame *sf = (StgSeqFrame *)su; - StgClosure* o; - - SET_HDR(ap,&PAP_info,su->header.prof.ccs /* ToDo */); - - /* now build o = FUN(seq,ap) */ - o = stgCast(StgClosure*, allocate(sizeofW(StgClosure)+1)); - TICK_ALLOC_THK(1,0); - SET_HDR(o,&seq_info,su->header.prof.ccs /* ToDo */); - payloadCPtr(o,0) = stgCast(StgClosure*,ap); - - IF_DEBUG(scheduler, - fprintf(stderr, "Built "); - printObj(stgCast(StgClosure*,o)); - ); - - /* pop the old handler and put o on the stack */ - su = sf->link; - sp += sizeofW(StgSeqFrame) - 1; - sp[0] = (W_)o; - break; - } - - case STOP_FRAME: - return; - - default: - barf("freezeTSO"); - } - } -} +#if defined(PAR) +StgTSO *LastTSO; +rtsTime TimeOfLastYield; +rtsBool emitSchedule = rtsTrue; +#endif -void initScheduler(void) -{ - run_queue_hd = END_TSO_QUEUE; - run_queue_tl = END_TSO_QUEUE; - blocked_queue_hd = END_TSO_QUEUE; - blocked_queue_tl = END_TSO_QUEUE; - ccalling_threads = END_TSO_QUEUE; - next_main_thread = 0; +#if DEBUG +char *whatNext_strs[] = { + "ThreadEnterGHC", + "ThreadRunGHC", + "ThreadEnterInterp", + "ThreadKilled", + "ThreadComplete" +}; + +char *threadReturnCode_strs[] = { + "HeapOverflow", /* might also be StackOverflow */ + "StackOverflow", + "ThreadYielding", + "ThreadBlocked", + "ThreadFinished" +}; +#endif - context_switch = 0; - interrupted = 0; +#ifdef PAR +StgTSO * createSparkThread(rtsSpark spark); +StgTSO * activateSpark (rtsSpark spark); +#endif - enteredCAFs = END_CAF_LIST; -} +/* + * The thread state for the main thread. +// ToDo: check whether not needed any more +StgTSO *MainTSO; + */ -void -run_all_threads ( void ) -{ - while (run_queue_hd != END_TSO_QUEUE) { - schedule(run_queue_hd, NULL); - } -} +//@node Main scheduling loop, Suspend and Resume, Prototypes, Main scheduling code +//@subsection Main scheduling loop -/* ----------------------------------------------------------------------------- +/* --------------------------------------------------------------------------- Main scheduling loop. We use round-robin scheduling, each thread returning to the scheduler loop when one of these conditions is detected: - * stack overflow * out of heap space * timer expires (thread yields) * thread blocks * thread ends - -------------------------------------------------------------------------- */ + * stack overflow -SchedulerStatus schedule(StgTSO *main, StgClosure **ret_val) + Locking notes: we acquire the scheduler lock once at the beginning + of the scheduler loop, and release it when + + * running a thread, or + * waiting for work, or + * waiting for a GC to complete. + + GRAN version: + In a GranSim setup this loop iterates over the global event queue. + This revolves around the global event queue, which determines what + to do next. Therefore, it's more complicated than either the + concurrent or the parallel (GUM) setup. + + GUM version: + GUM iterates over incoming messages. + It starts with nothing to do (thus CurrentTSO == END_TSO_QUEUE), + and sends out a fish whenever it has nothing to do; in-between + doing the actual reductions (shared code below) it processes the + incoming messages and deals with delayed operations + (see PendingFetches). + This is not the ugliest code you could imagine, but it's bloody close. + + ------------------------------------------------------------------------ */ +//@cindex schedule +static void +schedule( void ) { StgTSO *t; + Capability *cap; StgThreadReturnCode ret; - StgTSO **MainTSO; - rtsBool in_ccall_gc; +#if defined(GRAN) + rtsEvent *event; +#elif defined(PAR) + StgSparkPool *pool; + rtsSpark spark; + StgTSO *tso; + GlobalTaskId pe; + rtsBool receivedFinish = rtsFalse; +# if defined(DEBUG) + nat tp_size, sp_size; // stats only +# endif +#endif + rtsBool was_interrupted = rtsFalse; + + ACQUIRE_LOCK(&sched_mutex); - /* Return value is NULL by default, it is only filled in if the - * main thread completes successfully. - */ - if (ret_val) { *ret_val = NULL; } +#if defined(GRAN) - /* Save away a pointer to the main thread so that we can keep track - * of it should a garbage collection happen. We keep a stack of - * main threads in order to support scheduler re-entry. We can't - * use the normal TSO linkage for this stack, because the main TSO - * may need to be linked onto other queues. - */ - main_threads[next_main_thread] = main; - MainTSO = &main_threads[next_main_thread]; - next_main_thread++; - IF_DEBUG(scheduler, - fprintf(stderr, "Scheduler entered: nesting = %d\n", - next_main_thread);); + /* set up first event to get things going */ + /* ToDo: assign costs for system setup and init MainTSO ! */ + new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc], + ContinueThread, + CurrentTSO, (StgClosure*)NULL, (rtsSpark*)NULL); - /* Are we being re-entered? - */ - if (CurrentTSO != NULL) { - /* This happens when a _ccall_gc from Haskell ends up re-entering - * the scheduler. - * - * Block the current thread (put it on the ccalling_queue) and - * continue executing. The calling thread better have stashed - * away its state properly and left its stack with a proper stack - * frame on the top. - */ - threadPaused(CurrentTSO); - CurrentTSO->link = ccalling_threads; - ccalling_threads = CurrentTSO; - in_ccall_gc = rtsTrue; - IF_DEBUG(scheduler, - fprintf(stderr, "Re-entry, thread %d did a _ccall_gc\n", - CurrentTSO->id);); - } else { - in_ccall_gc = rtsFalse; - } + IF_DEBUG(gran, + fprintf(stderr, "GRAN: Init CurrentTSO (in schedule) = %p\n", CurrentTSO); + G_TSO(CurrentTSO, 5)); - /* Take a thread from the run queue. - */ - t = run_queue_hd; - if (t != END_TSO_QUEUE) { - run_queue_hd = t->link; - t->link = END_TSO_QUEUE; - if (run_queue_hd == END_TSO_QUEUE) { - run_queue_tl = END_TSO_QUEUE; - } - } + if (RtsFlags.GranFlags.Light) { + /* Save current time; GranSim Light only */ + CurrentTSO->gran.clock = CurrentTime[CurrentProc]; + } - while (t != END_TSO_QUEUE) { - CurrentTSO = t; + event = get_next_event(); - /* If we have more threads on the run queue, set up a context - * switch at some point in the future. - */ - if (run_queue_hd != END_TSO_QUEUE) { - context_switch = 1; - } else { - context_switch = 0; - } - IF_DEBUG(scheduler, belch("Running thread %ld...\n", t->id)); + while (event!=(rtsEvent*)NULL) { + /* Choose the processor with the next event */ + CurrentProc = event->proc; + CurrentTSO = event->tso; - /* Be friendly to the storage manager: we're about to *run* this - * thread, so we better make sure the TSO is mutable. - */ - if (t->mut_link == NULL) { - recordMutable((StgMutClosure *)t); - } +#elif defined(PAR) - /* Run the current thread */ - switch (t->whatNext) { - case ThreadKilled: - case ThreadComplete: - /* thread already killed. Drop it and carry on. */ - goto next_thread; - case ThreadEnterGHC: - ret = StgRun((StgFunPtr) stg_enterStackTop); - break; - case ThreadRunGHC: - ret = StgRun((StgFunPtr) stg_returnToStackTop); - break; - case ThreadEnterHugs: -#ifdef INTERPRETER - { - IF_DEBUG(scheduler,belch("entering Hugs")); - LoadThreadState(); - /* CHECK_SENSIBLE_REGS(); */ - { - StgClosure* c = stgCast(StgClosure*,*Sp); - Sp += 1; - ret = enter(c); - } - SaveThreadState(); - break; - } + while (!receivedFinish) { /* set by processMessages */ + /* when receiving PP_FINISH message */ #else - barf("Panic: entered a BCO but no bytecode interpreter in this build"); -#endif - default: - barf("schedule: invalid whatNext field"); - } - /* We may have garbage collected while running the thread - * (eg. something nefarious like _ccall_GC_ performGC), and hence - * CurrentTSO may have moved. Update t to reflect this. - */ - t = CurrentTSO; - CurrentTSO = NULL; + while (1) { - /* Costs for the scheduler are assigned to CCS_SYSTEM */ -#ifdef PROFILING - CCCS = CCS_SYSTEM; #endif - switch (ret) { + IF_DEBUG(scheduler, printAllThreads()); - case HeapOverflow: - IF_DEBUG(scheduler,belch("Thread %ld stopped: HeapOverflow\n", t->id)); - threadPaused(t); - PUSH_ON_RUN_QUEUE(t); - GarbageCollect(GetRoots); - break; + /* If we're interrupted (the user pressed ^C, or some other + * termination condition occurred), kill all the currently running + * threads. + */ + if (interrupted) { + IF_DEBUG(scheduler, sched_belch("interrupted")); + deleteAllThreads(); + interrupted = rtsFalse; + was_interrupted = rtsTrue; + } - case StackOverflow: - IF_DEBUG(scheduler,belch("Thread %ld stopped, StackOverflow\n", t->id)); - { - nat i; - /* enlarge the stack */ - StgTSO *new_t = threadStackOverflow(t); - - /* This TSO has moved, so update any pointers to it from the - * main thread stack. It better not be on any other queues... - * (it shouldn't be) - */ - for (i = 0; i < next_main_thread; i++) { - if (main_threads[i] == t) { - main_threads[i] = new_t; + /* Go through the list of main threads and wake up any + * clients whose computations have finished. ToDo: this + * should be done more efficiently without a linear scan + * of the main threads list, somehow... + */ +#ifdef SMP + { + StgMainThread *m, **prev; + prev = &main_threads; + for (m = main_threads; m != NULL; m = m->link) { + switch (m->tso->what_next) { + case ThreadComplete: + if (m->ret) { + *(m->ret) = (StgClosure *)m->tso->sp[0]; } + *prev = m->link; + m->stat = Success; + pthread_cond_broadcast(&m->wakeup); + break; + case ThreadKilled: + *prev = m->link; + if (was_interrupted) { + m->stat = Interrupted; + } else { + m->stat = Killed; + } + pthread_cond_broadcast(&m->wakeup); + break; + default: + break; } - t = new_t; } - PUSH_ON_RUN_QUEUE(t); - break; + } - case ThreadYielding: - IF_DEBUG(scheduler, - if (t->whatNext == ThreadEnterHugs) { - /* ToDo: or maybe a timer expired when we were in Hugs? - * or maybe someone hit ctrl-C - */ - belch("Thread %ld stopped to switch to Hugs\n", t->id); - } else { - belch("Thread %ld stopped, timer expired\n", t->id); - } - ); - threadPaused(t); - if (interrupted) { - IF_DEBUG(scheduler,belch("Scheduler interrupted - returning")); - deleteThread(t); - while (run_queue_hd != END_TSO_QUEUE) { - run_queue_hd = t->link; - deleteThread(t); +#else +# if defined(PAR) + /* in GUM do this only on the Main PE */ + if (IAmMainThread) +# endif + /* If our main thread has finished or been killed, return. + */ + { + StgMainThread *m = main_threads; + if (m->tso->what_next == ThreadComplete + || m->tso->what_next == ThreadKilled) { + main_threads = main_threads->link; + if (m->tso->what_next == ThreadComplete) { + /* we finished successfully, fill in the return value */ + if (m->ret) { *(m->ret) = (StgClosure *)m->tso->sp[0]; }; + m->stat = Success; + return; + } else { + if (was_interrupted) { + m->stat = Interrupted; + } else { + m->stat = Killed; } - run_queue_tl = END_TSO_QUEUE; - /* ToDo: should I do the same with blocked queues? */ - return Interrupted; + return; + } } + } +#endif - /* Put the thread back on the run queue, at the end. - * t->link is already set to END_TSO_QUEUE. - */ - ASSERT(t->link == END_TSO_QUEUE); - if (run_queue_tl == END_TSO_QUEUE) { - run_queue_hd = run_queue_tl = t; - } else { - ASSERT(get_itbl(run_queue_tl)->type == TSO); - if (run_queue_hd == run_queue_tl) { - run_queue_hd->link = t; - run_queue_tl = t; + /* Top up the run queue from our spark pool. We try to make the + * number of threads in the run queue equal to the number of + * free capabilities. + */ +#if defined(SMP) + { + nat n = n_free_capabilities; + StgTSO *tso = run_queue_hd; + + /* Count the run queue */ + while (n > 0 && tso != END_TSO_QUEUE) { + tso = tso->link; + n--; + } + + for (; n > 0; n--) { + StgClosure *spark; + spark = findSpark(rtsFalse); + if (spark == NULL) { + break; /* no more sparks in the pool */ } else { - run_queue_tl->link = t; - run_queue_tl = t; + /* I'd prefer this to be done in activateSpark -- HWL */ + /* tricky - it needs to hold the scheduler lock and + * not try to re-acquire it -- SDM */ + createSparkThread(spark); + IF_DEBUG(scheduler, + sched_belch("==^^ turning spark of closure %p into a thread", + (StgClosure *)spark)); } } - break; - - case ThreadBlocked: - IF_DEBUG(scheduler,belch("Thread %ld stopped, blocking\n", t->id)); - threadPaused(t); - /* assume the thread has put itself on some blocked queue - * somewhere. + /* We need to wake up the other tasks if we just created some + * work for them. */ - break; - - case ThreadFinished: - IF_DEBUG(scheduler,belch("Thread %ld finished\n", t->id)); - deleteThread(t); - t->whatNext = ThreadComplete; - break; + if (n_free_capabilities - n > 1) { + pthread_cond_signal(&thread_ready_cond); + } + } +#endif /* SMP */ - default: - barf("schedule: invalid thread return code"); + /* Check whether any waiting threads need to be woken up. If the + * run queue is empty, and there are no other tasks running, we + * can wait indefinitely for something to happen. + * ToDo: what if another client comes along & requests another + * main thread? + */ + if (blocked_queue_hd != END_TSO_QUEUE || sleeping_queue != END_TSO_QUEUE) { + awaitEvent( + (run_queue_hd == END_TSO_QUEUE) +#ifdef SMP + && (n_free_capabilities == RtsFlags.ParFlags.nNodes) +#endif + ); } + /* we can be interrupted while waiting for I/O... */ + if (interrupted) continue; /* check for signals each time around the scheduler */ +#ifndef mingw32_TARGET_OS if (signals_pending()) { start_signal_handlers(); } +#endif - /* If our main thread has finished or been killed, return. - * If we were re-entered as a result of a _ccall_gc, then - * pop the blocked thread off the ccalling_threads stack back - * into CurrentTSO. + /* + * Detect deadlock: when we have no threads to run, there are no + * threads waiting on I/O or sleeping, and all the other tasks are + * waiting for work, we must have a deadlock of some description. + * + * We first try to find threads blocked on themselves (ie. black + * holes), and generate NonTermination exceptions where necessary. + * + * If no threads are black holed, we have a deadlock situation, so + * inform all the main threads. */ - if ((*MainTSO)->whatNext == ThreadComplete - || (*MainTSO)->whatNext == ThreadKilled) { - next_main_thread--; - if (in_ccall_gc) { - CurrentTSO = ccalling_threads; - ccalling_threads = ccalling_threads->link; - /* remember to stub the link field of CurrentTSO */ - CurrentTSO->link = END_TSO_QUEUE; - } - if ((*MainTSO)->whatNext == ThreadComplete) { - /* we finished successfully, fill in the return value */ - if (ret_val) { *ret_val = (StgClosure *)(*MainTSO)->sp[0]; }; - return Success; - } else { - return Killed; - } +#ifdef SMP + if (blocked_queue_hd == END_TSO_QUEUE + && run_queue_hd == END_TSO_QUEUE + && sleeping_queue == END_TSO_QUEUE + && (n_free_capabilities == RtsFlags.ParFlags.nNodes)) + { + IF_DEBUG(scheduler, sched_belch("deadlocked, checking for black holes...")); + detectBlackHoles(); + if (run_queue_hd == END_TSO_QUEUE) { + StgMainThread *m; + for (m = main_threads; m != NULL; m = m->link) { + m->ret = NULL; + m->stat = Deadlock; + pthread_cond_broadcast(&m->wakeup); + } + main_threads = NULL; + } + } +#elif defined(PAR) + /* ToDo: add deadlock detection in GUM (similar to SMP) -- HWL */ +#else /* ! SMP */ + if (blocked_queue_hd == END_TSO_QUEUE + && run_queue_hd == END_TSO_QUEUE + && sleeping_queue == END_TSO_QUEUE) + { + IF_DEBUG(scheduler, sched_belch("deadlocked, checking for black holes...")); + detectBlackHoles(); + if (run_queue_hd == END_TSO_QUEUE) { + StgMainThread *m = main_threads; + m->ret = NULL; + m->stat = Deadlock; + main_threads = m->link; + return; + } } +#endif - next_thread: - t = run_queue_hd; - if (t != END_TSO_QUEUE) { - run_queue_hd = t->link; - t->link = END_TSO_QUEUE; - if (run_queue_hd == END_TSO_QUEUE) { +#ifdef SMP + /* If there's a GC pending, don't do anything until it has + * completed. + */ + if (ready_to_gc) { + IF_DEBUG(scheduler,sched_belch("waiting for GC")); + pthread_cond_wait(&gc_pending_cond, &sched_mutex); + } + + /* block until we've got a thread on the run queue and a free + * capability. + */ + while (run_queue_hd == END_TSO_QUEUE || free_capabilities == NULL) { + IF_DEBUG(scheduler, sched_belch("waiting for work")); + pthread_cond_wait(&thread_ready_cond, &sched_mutex); + IF_DEBUG(scheduler, sched_belch("work now available")); + } +#endif + +#if defined(GRAN) + + if (RtsFlags.GranFlags.Light) + GranSimLight_enter_system(event, &ActiveTSO); // adjust ActiveTSO etc + + /* adjust time based on time-stamp */ + if (event->time > CurrentTime[CurrentProc] && + event->evttype != ContinueThread) + CurrentTime[CurrentProc] = event->time; + + /* Deal with the idle PEs (may issue FindWork or MoveSpark events) */ + if (!RtsFlags.GranFlags.Light) + handleIdlePEs(); + + IF_DEBUG(gran, fprintf(stderr, "GRAN: switch by event-type\n")); + + /* main event dispatcher in GranSim */ + switch (event->evttype) { + /* Should just be continuing execution */ + case ContinueThread: + IF_DEBUG(gran, fprintf(stderr, "GRAN: doing ContinueThread\n")); + /* ToDo: check assertion + ASSERT(run_queue_hd != (StgTSO*)NULL && + run_queue_hd != END_TSO_QUEUE); + */ + /* Ignore ContinueThreads for fetching threads (if synchr comm) */ + if (!RtsFlags.GranFlags.DoAsyncFetch && + procStatus[CurrentProc]==Fetching) { + belch("ghuH: Spurious ContinueThread while Fetching ignored; TSO %d (%p) [PE %d]", + CurrentTSO->id, CurrentTSO, CurrentProc); + goto next_thread; + } + /* Ignore ContinueThreads for completed threads */ + if (CurrentTSO->what_next == ThreadComplete) { + belch("ghuH: found a ContinueThread event for completed thread %d (%p) [PE %d] (ignoring ContinueThread)", + CurrentTSO->id, CurrentTSO, CurrentProc); + goto next_thread; + } + /* Ignore ContinueThreads for threads that are being migrated */ + if (PROCS(CurrentTSO)==Nowhere) { + belch("ghuH: trying to run the migrating TSO %d (%p) [PE %d] (ignoring ContinueThread)", + CurrentTSO->id, CurrentTSO, CurrentProc); + goto next_thread; + } + /* The thread should be at the beginning of the run queue */ + if (CurrentTSO!=run_queue_hds[CurrentProc]) { + belch("ghuH: TSO %d (%p) [PE %d] is not at the start of the run_queue when doing a ContinueThread", + CurrentTSO->id, CurrentTSO, CurrentProc); + break; // run the thread anyway + } + /* + new_event(proc, proc, CurrentTime[proc], + FindWork, + (StgTSO*)NULL, (StgClosure*)NULL, (rtsSpark*)NULL); + goto next_thread; + */ /* Catches superfluous CONTINUEs -- should be unnecessary */ + break; // now actually run the thread; DaH Qu'vam yImuHbej + + case FetchNode: + do_the_fetchnode(event); + goto next_thread; /* handle next event in event queue */ + + case GlobalBlock: + do_the_globalblock(event); + goto next_thread; /* handle next event in event queue */ + + case FetchReply: + do_the_fetchreply(event); + goto next_thread; /* handle next event in event queue */ + + case UnblockThread: /* Move from the blocked queue to the tail of */ + do_the_unblock(event); + goto next_thread; /* handle next event in event queue */ + + case ResumeThread: /* Move from the blocked queue to the tail of */ + /* the runnable queue ( i.e. Qu' SImqa'lu') */ + event->tso->gran.blocktime += + CurrentTime[CurrentProc] - event->tso->gran.blockedat; + do_the_startthread(event); + goto next_thread; /* handle next event in event queue */ + + case StartThread: + do_the_startthread(event); + goto next_thread; /* handle next event in event queue */ + + case MoveThread: + do_the_movethread(event); + goto next_thread; /* handle next event in event queue */ + + case MoveSpark: + do_the_movespark(event); + goto next_thread; /* handle next event in event queue */ + + case FindWork: + do_the_findwork(event); + goto next_thread; /* handle next event in event queue */ + + default: + barf("Illegal event type %u\n", event->evttype); + } /* switch */ + + /* This point was scheduler_loop in the old RTS */ + + IF_DEBUG(gran, belch("GRAN: after main switch")); + + TimeOfLastEvent = CurrentTime[CurrentProc]; + TimeOfNextEvent = get_time_of_next_event(); + IgnoreEvents=(TimeOfNextEvent==0); // HWL HACK + // CurrentTSO = ThreadQueueHd; + + IF_DEBUG(gran, belch("GRAN: time of next event is: %ld", + TimeOfNextEvent)); + + if (RtsFlags.GranFlags.Light) + GranSimLight_leave_system(event, &ActiveTSO); + + EndOfTimeSlice = CurrentTime[CurrentProc]+RtsFlags.GranFlags.time_slice; + + IF_DEBUG(gran, + belch("GRAN: end of time-slice is %#lx", EndOfTimeSlice)); + + /* in a GranSim setup the TSO stays on the run queue */ + t = CurrentTSO; + /* Take a thread from the run queue. */ + t = POP_RUN_QUEUE(); // take_off_run_queue(t); + + IF_DEBUG(gran, + fprintf(stderr, "GRAN: About to run current thread, which is\n"); + G_TSO(t,5)); + + context_switch = 0; // turned on via GranYield, checking events and time slice + + IF_DEBUG(gran, + DumpGranEvent(GR_SCHEDULE, t)); + + procStatus[CurrentProc] = Busy; + +#elif defined(PAR) + if (PendingFetches != END_BF_QUEUE) { + processFetches(); + } + + /* ToDo: phps merge with spark activation above */ + /* check whether we have local work and send requests if we have none */ + if (EMPTY_RUN_QUEUE()) { /* no runnable threads */ + /* :-[ no local threads => look out for local sparks */ + /* the spark pool for the current PE */ + pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable + if (advisory_thread_count < RtsFlags.ParFlags.maxThreads && + pool->hd < pool->tl) { + /* + * ToDo: add GC code check that we really have enough heap afterwards!! + * Old comment: + * If we're here (no runnable threads) and we have pending + * sparks, we must have a space problem. Get enough space + * to turn one of those pending sparks into a + * thread... + */ + + spark = findSpark(rtsFalse); /* get a spark */ + if (spark != (rtsSpark) NULL) { + tso = activateSpark(spark); /* turn the spark into a thread */ + IF_PAR_DEBUG(schedule, + belch("==== schedule: Created TSO %d (%p); %d threads active", + tso->id, tso, advisory_thread_count)); + + if (tso==END_TSO_QUEUE) { /* failed to activate spark->back to loop */ + belch("==^^ failed to activate spark"); + goto next_thread; + } /* otherwise fall through & pick-up new tso */ + } else { + IF_PAR_DEBUG(verbose, + belch("==^^ no local sparks (spark pool contains only NFs: %d)", + spark_queue_len(pool))); + goto next_thread; + } + } + + /* If we still have no work we need to send a FISH to get a spark + from another PE + */ + if (EMPTY_RUN_QUEUE()) { + /* =8-[ no local sparks => look for work on other PEs */ + /* + * We really have absolutely no work. Send out a fish + * (there may be some out there already), and wait for + * something to arrive. We clearly can't run any threads + * until a SCHEDULE or RESUME arrives, and so that's what + * we're hoping to see. (Of course, we still have to + * respond to other types of messages.) + */ + TIME now = msTime() /*CURRENT_TIME*/; + IF_PAR_DEBUG(verbose, + belch("-- now=%ld", now)); + IF_PAR_DEBUG(verbose, + if (outstandingFishes < RtsFlags.ParFlags.maxFishes && + (last_fish_arrived_at!=0 && + last_fish_arrived_at+RtsFlags.ParFlags.fishDelay > now)) { + belch("--$$ delaying FISH until %ld (last fish %ld, delay %ld, now %ld)", + last_fish_arrived_at+RtsFlags.ParFlags.fishDelay, + last_fish_arrived_at, + RtsFlags.ParFlags.fishDelay, now); + }); + + if (outstandingFishes < RtsFlags.ParFlags.maxFishes && + (last_fish_arrived_at==0 || + (last_fish_arrived_at+RtsFlags.ParFlags.fishDelay <= now))) { + /* outstandingFishes is set in sendFish, processFish; + avoid flooding system with fishes via delay */ + pe = choosePE(); + sendFish(pe, mytid, NEW_FISH_AGE, NEW_FISH_HISTORY, + NEW_FISH_HUNGER); + + // Global statistics: count no. of fishes + if (RtsFlags.ParFlags.ParStats.Global && + RtsFlags.GcFlags.giveStats > NO_GC_STATS) { + globalParStats.tot_fish_mess++; + } + } + + receivedFinish = processMessages(); + goto next_thread; + } + } else if (PacketsWaiting()) { /* Look for incoming messages */ + receivedFinish = processMessages(); + } + + /* Now we are sure that we have some work available */ + ASSERT(run_queue_hd != END_TSO_QUEUE); + + /* Take a thread from the run queue, if we have work */ + t = POP_RUN_QUEUE(); // take_off_run_queue(END_TSO_QUEUE); + IF_DEBUG(sanity,checkTSO(t)); + + /* ToDo: write something to the log-file + if (RTSflags.ParFlags.granSimStats && !sameThread) + DumpGranEvent(GR_SCHEDULE, RunnableThreadsHd); + + CurrentTSO = t; + */ + /* the spark pool for the current PE */ + pool = &(MainRegTable.rSparks); // generalise to cap = &MainRegTable + + IF_DEBUG(scheduler, + belch("--=^ %d threads, %d sparks on [%#x]", + run_queue_len(), spark_queue_len(pool), CURRENT_PROC)); + +#if 1 + if (0 && RtsFlags.ParFlags.ParStats.Full && + t && LastTSO && t->id != LastTSO->id && + LastTSO->why_blocked == NotBlocked && + LastTSO->what_next != ThreadComplete) { + // if previously scheduled TSO not blocked we have to record the context switch + DumpVeryRawGranEvent(TimeOfLastYield, CURRENT_PROC, CURRENT_PROC, + GR_DESCHEDULE, LastTSO, (StgClosure *)NULL, 0, 0); + } + + if (RtsFlags.ParFlags.ParStats.Full && + (emitSchedule /* forced emit */ || + (t && LastTSO && t->id != LastTSO->id))) { + /* + we are running a different TSO, so write a schedule event to log file + NB: If we use fair scheduling we also have to write a deschedule + event for LastTSO; with unfair scheduling we know that the + previous tso has blocked whenever we switch to another tso, so + we don't need it in GUM for now + */ + DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC, + GR_SCHEDULE, t, (StgClosure *)NULL, 0, 0); + emitSchedule = rtsFalse; + } + +#endif +#else /* !GRAN && !PAR */ + + /* grab a thread from the run queue + */ + ASSERT(run_queue_hd != END_TSO_QUEUE); + t = POP_RUN_QUEUE(); + IF_DEBUG(sanity,checkTSO(t)); + +#endif + + /* grab a capability + */ +#ifdef SMP + cap = free_capabilities; + free_capabilities = cap->link; + n_free_capabilities--; +#else + cap = &MainRegTable; +#endif + + cap->rCurrentTSO = t; + + /* context switches are now initiated by the timer signal, unless + * the user specified "context switch as often as possible", with + * +RTS -C0 + */ + if (RtsFlags.ConcFlags.ctxtSwitchTicks == 0 + && (run_queue_hd != END_TSO_QUEUE + || blocked_queue_hd != END_TSO_QUEUE + || sleeping_queue != END_TSO_QUEUE)) + context_switch = 1; + else + context_switch = 0; + + RELEASE_LOCK(&sched_mutex); + + IF_DEBUG(scheduler, sched_belch("-->> Running TSO %ld (%p) %s ...", + t->id, t, whatNext_strs[t->what_next])); + + /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */ + /* Run the current thread + */ + switch (cap->rCurrentTSO->what_next) { + case ThreadKilled: + case ThreadComplete: + /* Thread already finished, return to scheduler. */ + ret = ThreadFinished; + break; + case ThreadEnterGHC: + ret = StgRun((StgFunPtr) stg_enterStackTop, cap); + break; + case ThreadRunGHC: + ret = StgRun((StgFunPtr) stg_returnToStackTop, cap); + break; + case ThreadEnterInterp: + ret = interpretBCO(cap); + break; + default: + barf("schedule: invalid what_next field"); + } + /* +++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++++ */ + + /* Costs for the scheduler are assigned to CCS_SYSTEM */ +#ifdef PROFILING + CCCS = CCS_SYSTEM; +#endif + + ACQUIRE_LOCK(&sched_mutex); + +#ifdef SMP + IF_DEBUG(scheduler,fprintf(stderr,"scheduler (task %ld): ", pthread_self());); +#elif !defined(GRAN) && !defined(PAR) + IF_DEBUG(scheduler,fprintf(stderr,"scheduler: ");); +#endif + t = cap->rCurrentTSO; + +#if defined(PAR) + /* HACK 675: if the last thread didn't yield, make sure to print a + SCHEDULE event to the log file when StgRunning the next thread, even + if it is the same one as before */ + LastTSO = t; + TimeOfLastYield = CURRENT_TIME; +#endif + + switch (ret) { + case HeapOverflow: +#if defined(GRAN) + IF_DEBUG(gran, + DumpGranEvent(GR_DESCHEDULE, t)); + globalGranStats.tot_heapover++; +#elif defined(PAR) + // IF_DEBUG(par, + //DumpGranEvent(GR_DESCHEDULE, t); + globalParStats.tot_heapover++; +#endif + /* make all the running tasks block on a condition variable, + * maybe set context_switch and wait till they all pile in, + * then have them wait on a GC condition variable. + */ + IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped: HeapOverflow", + t->id, t, whatNext_strs[t->what_next])); + threadPaused(t); +#if defined(GRAN) + ASSERT(!is_on_queue(t,CurrentProc)); +#elif defined(PAR) + /* Currently we emit a DESCHEDULE event before GC in GUM. + ToDo: either add separate event to distinguish SYSTEM time from rest + or just nuke this DESCHEDULE (and the following SCHEDULE) */ + if (0 && RtsFlags.ParFlags.ParStats.Full) { + DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC, + GR_DESCHEDULE, t, (StgClosure *)NULL, 0, 0); + emitSchedule = rtsTrue; + } +#endif + + ready_to_gc = rtsTrue; + context_switch = 1; /* stop other threads ASAP */ + PUSH_ON_RUN_QUEUE(t); + /* actual GC is done at the end of the while loop */ + break; + + case StackOverflow: +#if defined(GRAN) + IF_DEBUG(gran, + DumpGranEvent(GR_DESCHEDULE, t)); + globalGranStats.tot_stackover++; +#elif defined(PAR) + // IF_DEBUG(par, + // DumpGranEvent(GR_DESCHEDULE, t); + globalParStats.tot_stackover++; +#endif + IF_DEBUG(scheduler,belch("--<< thread %ld (%p; %s) stopped, StackOverflow", + t->id, t, whatNext_strs[t->what_next])); + /* just adjust the stack for this thread, then pop it back + * on the run queue. + */ + threadPaused(t); + { + StgMainThread *m; + /* enlarge the stack */ + StgTSO *new_t = threadStackOverflow(t); + + /* This TSO has moved, so update any pointers to it from the + * main thread stack. It better not be on any other queues... + * (it shouldn't be). + */ + for (m = main_threads; m != NULL; m = m->link) { + if (m->tso == t) { + m->tso = new_t; + } + } + threadPaused(new_t); + PUSH_ON_RUN_QUEUE(new_t); + } + break; + + case ThreadYielding: +#if defined(GRAN) + IF_DEBUG(gran, + DumpGranEvent(GR_DESCHEDULE, t)); + globalGranStats.tot_yields++; +#elif defined(PAR) + // IF_DEBUG(par, + // DumpGranEvent(GR_DESCHEDULE, t); + globalParStats.tot_yields++; +#endif + /* put the thread back on the run queue. Then, if we're ready to + * GC, check whether this is the last task to stop. If so, wake + * up the GC thread. getThread will block during a GC until the + * GC is finished. + */ + IF_DEBUG(scheduler, + if (t->what_next == ThreadEnterInterp) { + /* ToDo: or maybe a timer expired when we were in Hugs? + * or maybe someone hit ctrl-C + */ + belch("--<< thread %ld (%p; %s) stopped to switch to Hugs", + t->id, t, whatNext_strs[t->what_next]); + } else { + belch("--<< thread %ld (%p; %s) stopped, yielding", + t->id, t, whatNext_strs[t->what_next]); + } + ); + + threadPaused(t); + + IF_DEBUG(sanity, + //belch("&& Doing sanity check on yielding TSO %ld.", t->id); + checkTSO(t)); + ASSERT(t->link == END_TSO_QUEUE); +#if defined(GRAN) + ASSERT(!is_on_queue(t,CurrentProc)); + + IF_DEBUG(sanity, + //belch("&& Doing sanity check on all ThreadQueues (and their TSOs)."); + checkThreadQsSanity(rtsTrue)); +#endif +#if defined(PAR) + if (RtsFlags.ParFlags.doFairScheduling) { + /* this does round-robin scheduling; good for concurrency */ + APPEND_TO_RUN_QUEUE(t); + } else { + /* this does unfair scheduling; good for parallelism */ + PUSH_ON_RUN_QUEUE(t); + } +#else + /* this does round-robin scheduling; good for concurrency */ + APPEND_TO_RUN_QUEUE(t); +#endif +#if defined(GRAN) + /* add a ContinueThread event to actually process the thread */ + new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc], + ContinueThread, + t, (StgClosure*)NULL, (rtsSpark*)NULL); + IF_GRAN_DEBUG(bq, + belch("GRAN: eventq and runnableq after adding yielded thread to queue again:"); + G_EVENTQ(0); + G_CURR_THREADQ(0)); +#endif /* GRAN */ + break; + + case ThreadBlocked: +#if defined(GRAN) + IF_DEBUG(scheduler, + belch("--<< thread %ld (%p; %s) stopped, blocking on node %p [PE %d] with BQ: ", + t->id, t, whatNext_strs[t->what_next], t->block_info.closure, (t->block_info.closure==(StgClosure*)NULL ? 99 : where_is(t->block_info.closure))); + if (t->block_info.closure!=(StgClosure*)NULL) print_bq(t->block_info.closure)); + + // ??? needed; should emit block before + IF_DEBUG(gran, + DumpGranEvent(GR_DESCHEDULE, t)); + prune_eventq(t, (StgClosure *)NULL); // prune ContinueThreads for t + /* + ngoq Dogh! + ASSERT(procStatus[CurrentProc]==Busy || + ((procStatus[CurrentProc]==Fetching) && + (t->block_info.closure!=(StgClosure*)NULL))); + if (run_queue_hds[CurrentProc] == END_TSO_QUEUE && + !(!RtsFlags.GranFlags.DoAsyncFetch && + procStatus[CurrentProc]==Fetching)) + procStatus[CurrentProc] = Idle; + */ +#elif defined(PAR) + IF_DEBUG(scheduler, + belch("--<< thread %ld (%p; %s) stopped, blocking on node %p with BQ: ", + t->id, t, whatNext_strs[t->what_next], t->block_info.closure)); + IF_PAR_DEBUG(bq, + + if (t->block_info.closure!=(StgClosure*)NULL) + print_bq(t->block_info.closure)); + + /* Send a fetch (if BlockedOnGA) and dump event to log file */ + blockThread(t); + + /* whatever we schedule next, we must log that schedule */ + emitSchedule = rtsTrue; + +#else /* !GRAN */ + /* don't need to do anything. Either the thread is blocked on + * I/O, in which case we'll have called addToBlockedQueue + * previously, or it's blocked on an MVar or Blackhole, in which + * case it'll be on the relevant queue already. + */ + IF_DEBUG(scheduler, + fprintf(stderr, "--<< thread %d (%p) stopped: ", t->id, t); + printThreadBlockage(t); + fprintf(stderr, "\n")); + + /* Only for dumping event to log file + ToDo: do I need this in GranSim, too? + blockThread(t); + */ +#endif + threadPaused(t); + break; + + case ThreadFinished: + /* Need to check whether this was a main thread, and if so, signal + * the task that started it with the return value. If we have no + * more main threads, we probably need to stop all the tasks until + * we get a new one. + */ + /* We also end up here if the thread kills itself with an + * uncaught exception, see Exception.hc. + */ + IF_DEBUG(scheduler,belch("--++ thread %d (%p) finished", t->id, t)); +#if defined(GRAN) + endThread(t, CurrentProc); // clean-up the thread +#elif defined(PAR) + /* For now all are advisory -- HWL */ + //if(t->priority==AdvisoryPriority) ?? + advisory_thread_count--; + +# ifdef DIST + if(t->dist.priority==RevalPriority) + FinishReval(t); +# endif + + if (RtsFlags.ParFlags.ParStats.Full && + !RtsFlags.ParFlags.ParStats.Suppressed) + DumpEndEvent(CURRENT_PROC, t, rtsFalse /* not mandatory */); +#endif + break; + + default: + barf("schedule: invalid thread return code %d", (int)ret); + } + +#ifdef SMP + cap->link = free_capabilities; + free_capabilities = cap; + n_free_capabilities++; +#endif + +#ifdef SMP + if (ready_to_gc && n_free_capabilities == RtsFlags.ParFlags.nNodes) +#else + if (ready_to_gc) +#endif + { + /* everybody back, start the GC. + * Could do it in this thread, or signal a condition var + * to do it in another thread. Either way, we need to + * broadcast on gc_pending_cond afterward. + */ +#ifdef SMP + IF_DEBUG(scheduler,sched_belch("doing GC")); +#endif + GarbageCollect(GetRoots,rtsFalse); + ready_to_gc = rtsFalse; +#ifdef SMP + pthread_cond_broadcast(&gc_pending_cond); +#endif +#if defined(GRAN) + /* add a ContinueThread event to continue execution of current thread */ + new_event(CurrentProc, CurrentProc, CurrentTime[CurrentProc], + ContinueThread, + t, (StgClosure*)NULL, (rtsSpark*)NULL); + IF_GRAN_DEBUG(bq, + fprintf(stderr, "GRAN: eventq and runnableq after Garbage collection:\n"); + G_EVENTQ(0); + G_CURR_THREADQ(0)); +#endif /* GRAN */ + } +#if defined(GRAN) + next_thread: + IF_GRAN_DEBUG(unused, + print_eventq(EventHd)); + + event = get_next_event(); + +#elif defined(PAR) + next_thread: + /* ToDo: wait for next message to arrive rather than busy wait */ + +#else /* GRAN */ + /* not any more + next_thread: + t = take_off_run_queue(END_TSO_QUEUE); + */ +#endif /* GRAN */ + } /* end of while(1) */ + IF_PAR_DEBUG(verbose, + belch("== Leaving schedule() after having received Finish")); +} + +/* --------------------------------------------------------------------------- + * deleteAllThreads(): kill all the live threads. + * + * This is used when we catch a user interrupt (^C), before performing + * any necessary cleanups and running finalizers. + * ------------------------------------------------------------------------- */ + +void deleteAllThreads ( void ) +{ + StgTSO* t; + IF_DEBUG(scheduler,sched_belch("deleting all threads")); + for (t = run_queue_hd; t != END_TSO_QUEUE; t = t->link) { + deleteThread(t); + } + for (t = blocked_queue_hd; t != END_TSO_QUEUE; t = t->link) { + deleteThread(t); + } + for (t = sleeping_queue; t != END_TSO_QUEUE; t = t->link) { + deleteThread(t); + } + run_queue_hd = run_queue_tl = END_TSO_QUEUE; + blocked_queue_hd = blocked_queue_tl = END_TSO_QUEUE; + sleeping_queue = END_TSO_QUEUE; +} + +/* startThread and insertThread are now in GranSim.c -- HWL */ + +//@node Suspend and Resume, Run queue code, Main scheduling loop, Main scheduling code +//@subsection Suspend and Resume + +/* --------------------------------------------------------------------------- + * Suspending & resuming Haskell threads. + * + * When making a "safe" call to C (aka _ccall_GC), the task gives back + * its capability before calling the C function. This allows another + * task to pick up the capability and carry on running Haskell + * threads. It also means that if the C call blocks, it won't lock + * the whole system. + * + * The Haskell thread making the C call is put to sleep for the + * duration of the call, on the susepended_ccalling_threads queue. We + * give out a token to the task, which it can use to resume the thread + * on return from the C function. + * ------------------------------------------------------------------------- */ + +StgInt +suspendThread( Capability *cap ) +{ + nat tok; + + ACQUIRE_LOCK(&sched_mutex); + + IF_DEBUG(scheduler, + sched_belch("thread %d did a _ccall_gc", cap->rCurrentTSO->id)); + + threadPaused(cap->rCurrentTSO); + cap->rCurrentTSO->link = suspended_ccalling_threads; + suspended_ccalling_threads = cap->rCurrentTSO; + + /* Use the thread ID as the token; it should be unique */ + tok = cap->rCurrentTSO->id; + +#ifdef SMP + cap->link = free_capabilities; + free_capabilities = cap; + n_free_capabilities++; +#endif + + RELEASE_LOCK(&sched_mutex); + return tok; +} + +Capability * +resumeThread( StgInt tok ) +{ + StgTSO *tso, **prev; + Capability *cap; + + ACQUIRE_LOCK(&sched_mutex); + + prev = &suspended_ccalling_threads; + for (tso = suspended_ccalling_threads; + tso != END_TSO_QUEUE; + prev = &tso->link, tso = tso->link) { + if (tso->id == (StgThreadID)tok) { + *prev = tso->link; + break; + } + } + if (tso == END_TSO_QUEUE) { + barf("resumeThread: thread not found"); + } + tso->link = END_TSO_QUEUE; + +#ifdef SMP + while (free_capabilities == NULL) { + IF_DEBUG(scheduler, sched_belch("waiting to resume")); + pthread_cond_wait(&thread_ready_cond, &sched_mutex); + IF_DEBUG(scheduler, sched_belch("resuming thread %d", tso->id)); + } + cap = free_capabilities; + free_capabilities = cap->link; + n_free_capabilities--; +#else + cap = &MainRegTable; +#endif + + cap->rCurrentTSO = tso; + + RELEASE_LOCK(&sched_mutex); + return cap; +} + + +/* --------------------------------------------------------------------------- + * Static functions + * ------------------------------------------------------------------------ */ +static void unblockThread(StgTSO *tso); + +/* --------------------------------------------------------------------------- + * Comparing Thread ids. + * + * This is used from STG land in the implementation of the + * instances of Eq/Ord for ThreadIds. + * ------------------------------------------------------------------------ */ + +int cmp_thread(const StgTSO *tso1, const StgTSO *tso2) +{ + StgThreadID id1 = tso1->id; + StgThreadID id2 = tso2->id; + + if (id1 < id2) return (-1); + if (id1 > id2) return 1; + return 0; +} + +/* --------------------------------------------------------------------------- + Create a new thread. + + The new thread starts with the given stack size. Before the + scheduler can run, however, this thread needs to have a closure + (and possibly some arguments) pushed on its stack. See + pushClosure() in Schedule.h. + + createGenThread() and createIOThread() (in SchedAPI.h) are + convenient packaged versions of this function. + + currently pri (priority) is only used in a GRAN setup -- HWL + ------------------------------------------------------------------------ */ +//@cindex createThread +#if defined(GRAN) +/* currently pri (priority) is only used in a GRAN setup -- HWL */ +StgTSO * +createThread(nat stack_size, StgInt pri) +{ + return createThread_(stack_size, rtsFalse, pri); +} + +static StgTSO * +createThread_(nat size, rtsBool have_lock, StgInt pri) +{ +#else +StgTSO * +createThread(nat stack_size) +{ + return createThread_(stack_size, rtsFalse); +} + +static StgTSO * +createThread_(nat size, rtsBool have_lock) +{ +#endif + + StgTSO *tso; + nat stack_size; + + /* First check whether we should create a thread at all */ +#if defined(PAR) + /* check that no more than RtsFlags.ParFlags.maxThreads threads are created */ + if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) { + threadsIgnored++; + belch("{createThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)", + RtsFlags.ParFlags.maxThreads, advisory_thread_count); + return END_TSO_QUEUE; + } + threadsCreated++; +#endif + +#if defined(GRAN) + ASSERT(!RtsFlags.GranFlags.Light || CurrentProc==0); +#endif + + // ToDo: check whether size = stack_size - TSO_STRUCT_SIZEW + + /* catch ridiculously small stack sizes */ + if (size < MIN_STACK_WORDS + TSO_STRUCT_SIZEW) { + size = MIN_STACK_WORDS + TSO_STRUCT_SIZEW; + } + + stack_size = size - TSO_STRUCT_SIZEW; + + tso = (StgTSO *)allocate(size); + TICK_ALLOC_TSO(size-TSO_STRUCT_SIZEW, 0); + + SET_HDR(tso, &stg_TSO_info, CCS_SYSTEM); +#if defined(GRAN) + SET_GRAN_HDR(tso, ThisPE); +#endif + tso->what_next = ThreadEnterGHC; + + /* tso->id needs to be unique. For now we use a heavyweight mutex to + * protect the increment operation on next_thread_id. + * In future, we could use an atomic increment instead. + */ + if (!have_lock) { ACQUIRE_LOCK(&sched_mutex); } + tso->id = next_thread_id++; + if (!have_lock) { RELEASE_LOCK(&sched_mutex); } + + tso->why_blocked = NotBlocked; + tso->blocked_exceptions = NULL; + + tso->stack_size = stack_size; + tso->max_stack_size = round_to_mblocks(RtsFlags.GcFlags.maxStkSize) + - TSO_STRUCT_SIZEW; + tso->sp = (P_)&(tso->stack) + stack_size; + +#ifdef PROFILING + tso->prof.CCCS = CCS_MAIN; +#endif + + /* put a stop frame on the stack */ + tso->sp -= sizeofW(StgStopFrame); + SET_HDR((StgClosure*)tso->sp,(StgInfoTable *)&stg_stop_thread_info,CCS_SYSTEM); + tso->su = (StgUpdateFrame*)tso->sp; + + // ToDo: check this +#if defined(GRAN) + tso->link = END_TSO_QUEUE; + /* uses more flexible routine in GranSim */ + insertThread(tso, CurrentProc); +#else + /* In a non-GranSim setup the pushing of a TSO onto the runq is separated + * from its creation + */ +#endif + +#if defined(GRAN) + if (RtsFlags.GranFlags.GranSimStats.Full) + DumpGranEvent(GR_START,tso); +#elif defined(PAR) + if (RtsFlags.ParFlags.ParStats.Full) + DumpGranEvent(GR_STARTQ,tso); + /* HACk to avoid SCHEDULE + LastTSO = tso; */ +#endif + + /* Link the new thread on the global thread list. + */ + tso->global_link = all_threads; + all_threads = tso; + +#if defined(DIST) + tso->dist.priority = MandatoryPriority; //by default that is... +#endif + +#if defined(GRAN) + tso->gran.pri = pri; +# if defined(DEBUG) + tso->gran.magic = TSO_MAGIC; // debugging only +# endif + tso->gran.sparkname = 0; + tso->gran.startedat = CURRENT_TIME; + tso->gran.exported = 0; + tso->gran.basicblocks = 0; + tso->gran.allocs = 0; + tso->gran.exectime = 0; + tso->gran.fetchtime = 0; + tso->gran.fetchcount = 0; + tso->gran.blocktime = 0; + tso->gran.blockcount = 0; + tso->gran.blockedat = 0; + tso->gran.globalsparks = 0; + tso->gran.localsparks = 0; + if (RtsFlags.GranFlags.Light) + tso->gran.clock = Now; /* local clock */ + else + tso->gran.clock = 0; + + IF_DEBUG(gran,printTSO(tso)); +#elif defined(PAR) +# if defined(DEBUG) + tso->par.magic = TSO_MAGIC; // debugging only +# endif + tso->par.sparkname = 0; + tso->par.startedat = CURRENT_TIME; + tso->par.exported = 0; + tso->par.basicblocks = 0; + tso->par.allocs = 0; + tso->par.exectime = 0; + tso->par.fetchtime = 0; + tso->par.fetchcount = 0; + tso->par.blocktime = 0; + tso->par.blockcount = 0; + tso->par.blockedat = 0; + tso->par.globalsparks = 0; + tso->par.localsparks = 0; +#endif + +#if defined(GRAN) + globalGranStats.tot_threads_created++; + globalGranStats.threads_created_on_PE[CurrentProc]++; + globalGranStats.tot_sq_len += spark_queue_len(CurrentProc); + globalGranStats.tot_sq_probes++; +#elif defined(PAR) + // collect parallel global statistics (currently done together with GC stats) + if (RtsFlags.ParFlags.ParStats.Global && + RtsFlags.GcFlags.giveStats > NO_GC_STATS) { + //fprintf(stderr, "Creating thread %d @ %11.2f\n", tso->id, usertime()); + globalParStats.tot_threads_created++; + } +#endif + +#if defined(GRAN) + IF_GRAN_DEBUG(pri, + belch("==__ schedule: Created TSO %d (%p);", + CurrentProc, tso, tso->id)); +#elif defined(PAR) + IF_PAR_DEBUG(verbose, + belch("==__ schedule: Created TSO %d (%p); %d threads active", + tso->id, tso, advisory_thread_count)); +#else + IF_DEBUG(scheduler,sched_belch("created thread %ld, stack size = %lx words", + tso->id, tso->stack_size)); +#endif + return tso; +} + +#if defined(PAR) +/* RFP: + all parallel thread creation calls should fall through the following routine. +*/ +StgTSO * +createSparkThread(rtsSpark spark) +{ StgTSO *tso; + ASSERT(spark != (rtsSpark)NULL); + if (advisory_thread_count >= RtsFlags.ParFlags.maxThreads) + { threadsIgnored++; + barf("{createSparkThread}Daq ghuH: refusing to create another thread; no more than %d threads allowed (currently %d)", + RtsFlags.ParFlags.maxThreads, advisory_thread_count); + return END_TSO_QUEUE; + } + else + { threadsCreated++; + tso = createThread_(RtsFlags.GcFlags.initialStkSize, rtsTrue); + if (tso==END_TSO_QUEUE) + barf("createSparkThread: Cannot create TSO"); +#if defined(DIST) + tso->priority = AdvisoryPriority; +#endif + pushClosure(tso,spark); + PUSH_ON_RUN_QUEUE(tso); + advisory_thread_count++; + } + return tso; +} +#endif + +/* + Turn a spark into a thread. + ToDo: fix for SMP (needs to acquire SCHED_MUTEX!) +*/ +#if defined(PAR) +//@cindex activateSpark +StgTSO * +activateSpark (rtsSpark spark) +{ + StgTSO *tso; + + tso = createSparkThread(spark); + if (RtsFlags.ParFlags.ParStats.Full) { + //ASSERT(run_queue_hd == END_TSO_QUEUE); // I think ... + IF_PAR_DEBUG(verbose, + belch("==^^ activateSpark: turning spark of closure %p (%s) into a thread", + (StgClosure *)spark, info_type((StgClosure *)spark))); + } + // ToDo: fwd info on local/global spark to thread -- HWL + // tso->gran.exported = spark->exported; + // tso->gran.locked = !spark->global; + // tso->gran.sparkname = spark->name; + + return tso; +} +#endif + +/* --------------------------------------------------------------------------- + * scheduleThread() + * + * scheduleThread puts a thread on the head of the runnable queue. + * This will usually be done immediately after a thread is created. + * The caller of scheduleThread must create the thread using e.g. + * createThread and push an appropriate closure + * on this thread's stack before the scheduler is invoked. + * ------------------------------------------------------------------------ */ + +void +scheduleThread(StgTSO *tso) +{ + if (tso==END_TSO_QUEUE){ + schedule(); + return; + } + + ACQUIRE_LOCK(&sched_mutex); + + /* Put the new thread on the head of the runnable queue. The caller + * better push an appropriate closure on this thread's stack + * beforehand. In the SMP case, the thread may start running as + * soon as we release the scheduler lock below. + */ + PUSH_ON_RUN_QUEUE(tso); + THREAD_RUNNABLE(); + +#if 0 + IF_DEBUG(scheduler,printTSO(tso)); +#endif + RELEASE_LOCK(&sched_mutex); +} + +/* --------------------------------------------------------------------------- + * startTasks() + * + * Start up Posix threads to run each of the scheduler tasks. + * I believe the task ids are not needed in the system as defined. + * KH @ 25/10/99 + * ------------------------------------------------------------------------ */ + +#if defined(PAR) || defined(SMP) +void +taskStart(void) /* ( void *arg STG_UNUSED) */ +{ + scheduleThread(END_TSO_QUEUE); +} +#endif + +/* --------------------------------------------------------------------------- + * initScheduler() + * + * Initialise the scheduler. This resets all the queues - if the + * queues contained any threads, they'll be garbage collected at the + * next pass. + * + * This now calls startTasks(), so should only be called once! KH @ 25/10/99 + * ------------------------------------------------------------------------ */ + +#ifdef SMP +static void +term_handler(int sig STG_UNUSED) +{ + stat_workerStop(); + ACQUIRE_LOCK(&term_mutex); + await_death--; + RELEASE_LOCK(&term_mutex); + pthread_exit(NULL); +} +#endif + +//@cindex initScheduler +void +initScheduler(void) +{ +#if defined(GRAN) + nat i; + + for (i=0; i<=MAX_PROC; i++) { + run_queue_hds[i] = END_TSO_QUEUE; + run_queue_tls[i] = END_TSO_QUEUE; + blocked_queue_hds[i] = END_TSO_QUEUE; + blocked_queue_tls[i] = END_TSO_QUEUE; + ccalling_threadss[i] = END_TSO_QUEUE; + sleeping_queue = END_TSO_QUEUE; + } +#else + run_queue_hd = END_TSO_QUEUE; + run_queue_tl = END_TSO_QUEUE; + blocked_queue_hd = END_TSO_QUEUE; + blocked_queue_tl = END_TSO_QUEUE; + sleeping_queue = END_TSO_QUEUE; +#endif + + suspended_ccalling_threads = END_TSO_QUEUE; + + main_threads = NULL; + all_threads = END_TSO_QUEUE; + + context_switch = 0; + interrupted = 0; + + RtsFlags.ConcFlags.ctxtSwitchTicks = + RtsFlags.ConcFlags.ctxtSwitchTime / TICK_MILLISECS; + + /* Install the SIGHUP handler */ +#ifdef SMP + { + struct sigaction action,oact; + + action.sa_handler = term_handler; + sigemptyset(&action.sa_mask); + action.sa_flags = 0; + if (sigaction(SIGTERM, &action, &oact) != 0) { + barf("can't install TERM handler"); + } + } +#endif + +#ifdef SMP + /* Allocate N Capabilities */ + { + nat i; + Capability *cap, *prev; + cap = NULL; + prev = NULL; + for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) { + cap = stgMallocBytes(sizeof(Capability), "initScheduler:capabilities"); + cap->link = prev; + prev = cap; + } + free_capabilities = cap; + n_free_capabilities = RtsFlags.ParFlags.nNodes; + } + IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Allocated %d capabilities\n", + n_free_capabilities);); +#endif + +#if defined(SMP) || defined(PAR) + initSparkPools(); +#endif +} + +#ifdef SMP +void +startTasks( void ) +{ + nat i; + int r; + pthread_t tid; + + /* make some space for saving all the thread ids */ + task_ids = stgMallocBytes(RtsFlags.ParFlags.nNodes * sizeof(task_info), + "initScheduler:task_ids"); + + /* and create all the threads */ + for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) { + r = pthread_create(&tid,NULL,taskStart,NULL); + if (r != 0) { + barf("startTasks: Can't create new Posix thread"); + } + task_ids[i].id = tid; + task_ids[i].mut_time = 0.0; + task_ids[i].mut_etime = 0.0; + task_ids[i].gc_time = 0.0; + task_ids[i].gc_etime = 0.0; + task_ids[i].elapsedtimestart = elapsedtime(); + IF_DEBUG(scheduler,fprintf(stderr,"scheduler: Started task: %ld\n",tid);); + } +} +#endif + +void +exitScheduler( void ) +{ +#ifdef SMP + nat i; + + /* Don't want to use pthread_cancel, since we'd have to install + * these silly exception handlers (pthread_cleanup_{push,pop}) around + * all our locks. + */ +#if 0 + /* Cancel all our tasks */ + for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) { + pthread_cancel(task_ids[i].id); + } + + /* Wait for all the tasks to terminate */ + for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) { + IF_DEBUG(scheduler,fprintf(stderr,"scheduler: waiting for task %ld\n", + task_ids[i].id)); + pthread_join(task_ids[i].id, NULL); + } +#endif + + /* Send 'em all a SIGHUP. That should shut 'em up. + */ + await_death = RtsFlags.ParFlags.nNodes; + for (i = 0; i < RtsFlags.ParFlags.nNodes; i++) { + pthread_kill(task_ids[i].id,SIGTERM); + } + while (await_death > 0) { + sched_yield(); + } +#endif +} + +/* ----------------------------------------------------------------------------- + Managing the per-task allocation areas. + + Each capability comes with an allocation area. These are + fixed-length block lists into which allocation can be done. + + ToDo: no support for two-space collection at the moment??? + -------------------------------------------------------------------------- */ + +/* ----------------------------------------------------------------------------- + * waitThread is the external interface for running a new computation + * and waiting for the result. + * + * In the non-SMP case, we create a new main thread, push it on the + * main-thread stack, and invoke the scheduler to run it. The + * scheduler will return when the top main thread on the stack has + * completed or died, and fill in the necessary fields of the + * main_thread structure. + * + * In the SMP case, we create a main thread as before, but we then + * create a new condition variable and sleep on it. When our new + * main thread has completed, we'll be woken up and the status/result + * will be in the main_thread struct. + * -------------------------------------------------------------------------- */ + +int +howManyThreadsAvail ( void ) +{ + int i = 0; + StgTSO* q; + for (q = run_queue_hd; q != END_TSO_QUEUE; q = q->link) + i++; + for (q = blocked_queue_hd; q != END_TSO_QUEUE; q = q->link) + i++; + for (q = sleeping_queue; q != END_TSO_QUEUE; q = q->link) + i++; + return i; +} + +void +finishAllThreads ( void ) +{ + do { + while (run_queue_hd != END_TSO_QUEUE) { + waitThread ( run_queue_hd, NULL ); + } + while (blocked_queue_hd != END_TSO_QUEUE) { + waitThread ( blocked_queue_hd, NULL ); + } + while (sleeping_queue != END_TSO_QUEUE) { + waitThread ( blocked_queue_hd, NULL ); + } + } while + (blocked_queue_hd != END_TSO_QUEUE || + run_queue_hd != END_TSO_QUEUE || + sleeping_queue != END_TSO_QUEUE); +} + +SchedulerStatus +waitThread(StgTSO *tso, /*out*/StgClosure **ret) +{ + StgMainThread *m; + SchedulerStatus stat; + + ACQUIRE_LOCK(&sched_mutex); + + m = stgMallocBytes(sizeof(StgMainThread), "waitThread"); + + m->tso = tso; + m->ret = ret; + m->stat = NoStatus; +#ifdef SMP + pthread_cond_init(&m->wakeup, NULL); +#endif + + m->link = main_threads; + main_threads = m; + + IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: new main thread (%d)\n", + m->tso->id)); + +#ifdef SMP + do { + pthread_cond_wait(&m->wakeup, &sched_mutex); + } while (m->stat == NoStatus); +#elif defined(GRAN) + /* GranSim specific init */ + CurrentTSO = m->tso; // the TSO to run + procStatus[MainProc] = Busy; // status of main PE + CurrentProc = MainProc; // PE to run it on + + schedule(); +#else + schedule(); + ASSERT(m->stat != NoStatus); +#endif + + stat = m->stat; + +#ifdef SMP + pthread_cond_destroy(&m->wakeup); +#endif + + IF_DEBUG(scheduler, fprintf(stderr, "== scheduler: main thread (%d) finished\n", + m->tso->id)); + free(m); + + RELEASE_LOCK(&sched_mutex); + + return stat; +} + +//@node Run queue code, Garbage Collextion Routines, Suspend and Resume, Main scheduling code +//@subsection Run queue code + +#if 0 +/* + NB: In GranSim we have many run queues; run_queue_hd is actually a macro + unfolding to run_queue_hds[CurrentProc], thus CurrentProc is an + implicit global variable that has to be correct when calling these + fcts -- HWL +*/ + +/* Put the new thread on the head of the runnable queue. + * The caller of createThread better push an appropriate closure + * on this thread's stack before the scheduler is invoked. + */ +static /* inline */ void +add_to_run_queue(tso) +StgTSO* tso; +{ + ASSERT(tso!=run_queue_hd && tso!=run_queue_tl); + tso->link = run_queue_hd; + run_queue_hd = tso; + if (run_queue_tl == END_TSO_QUEUE) { + run_queue_tl = tso; + } +} + +/* Put the new thread at the end of the runnable queue. */ +static /* inline */ void +push_on_run_queue(tso) +StgTSO* tso; +{ + ASSERT(get_itbl((StgClosure *)tso)->type == TSO); + ASSERT(run_queue_hd!=NULL && run_queue_tl!=NULL); + ASSERT(tso!=run_queue_hd && tso!=run_queue_tl); + if (run_queue_hd == END_TSO_QUEUE) { + run_queue_hd = tso; + } else { + run_queue_tl->link = tso; + } + run_queue_tl = tso; +} + +/* + Should be inlined because it's used very often in schedule. The tso + argument is actually only needed in GranSim, where we want to have the + possibility to schedule *any* TSO on the run queue, irrespective of the + actual ordering. Therefore, if tso is not the nil TSO then we traverse + the run queue and dequeue the tso, adjusting the links in the queue. +*/ +//@cindex take_off_run_queue +static /* inline */ StgTSO* +take_off_run_queue(StgTSO *tso) { + StgTSO *t, *prev; + + /* + qetlaHbogh Qu' ngaSbogh ghomDaQ {tso} yIteq! + + if tso is specified, unlink that tso from the run_queue (doesn't have + to be at the beginning of the queue); GranSim only + */ + if (tso!=END_TSO_QUEUE) { + /* find tso in queue */ + for (t=run_queue_hd, prev=END_TSO_QUEUE; + t!=END_TSO_QUEUE && t!=tso; + prev=t, t=t->link) + /* nothing */ ; + ASSERT(t==tso); + /* now actually dequeue the tso */ + if (prev!=END_TSO_QUEUE) { + ASSERT(run_queue_hd!=t); + prev->link = t->link; + } else { + /* t is at beginning of thread queue */ + ASSERT(run_queue_hd==t); + run_queue_hd = t->link; + } + /* t is at end of thread queue */ + if (t->link==END_TSO_QUEUE) { + ASSERT(t==run_queue_tl); + run_queue_tl = prev; + } else { + ASSERT(run_queue_tl!=t); + } + t->link = END_TSO_QUEUE; + } else { + /* take tso from the beginning of the queue; std concurrent code */ + t = run_queue_hd; + if (t != END_TSO_QUEUE) { + run_queue_hd = t->link; + t->link = END_TSO_QUEUE; + if (run_queue_hd == END_TSO_QUEUE) { run_queue_tl = END_TSO_QUEUE; } } - } + } + return t; +} + +#endif /* 0 */ + +//@node Garbage Collextion Routines, Blocking Queue Routines, Run queue code, Main scheduling code +//@subsection Garbage Collextion Routines + +/* --------------------------------------------------------------------------- + Where are the roots that we know about? + + - all the threads on the runnable queue + - all the threads on the blocked queue + - all the threads on the sleeping queue + - all the thread currently executing a _ccall_GC + - all the "main threads" + + ------------------------------------------------------------------------ */ + +/* This has to be protected either by the scheduler monitor, or by the + garbage collection monitor (probably the latter). + KH @ 25/10/99 +*/ + +static void GetRoots(void) +{ + StgMainThread *m; + +#if defined(GRAN) + { + nat i; + for (i=0; i<=RtsFlags.GranFlags.proc; i++) { + if ((run_queue_hds[i] != END_TSO_QUEUE) && ((run_queue_hds[i] != NULL))) + run_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_hds[i]); + if ((run_queue_tls[i] != END_TSO_QUEUE) && ((run_queue_tls[i] != NULL))) + run_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)run_queue_tls[i]); + + if ((blocked_queue_hds[i] != END_TSO_QUEUE) && ((blocked_queue_hds[i] != NULL))) + blocked_queue_hds[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hds[i]); + if ((blocked_queue_tls[i] != END_TSO_QUEUE) && ((blocked_queue_tls[i] != NULL))) + blocked_queue_tls[i] = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tls[i]); + if ((ccalling_threadss[i] != END_TSO_QUEUE) && ((ccalling_threadss[i] != NULL))) + ccalling_threadss[i] = (StgTSO *)MarkRoot((StgClosure *)ccalling_threadss[i]); + } + } + + markEventQueue(); + +#else /* !GRAN */ + if (run_queue_hd != END_TSO_QUEUE) { + ASSERT(run_queue_tl != END_TSO_QUEUE); + run_queue_hd = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd); + run_queue_tl = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl); + } + + if (blocked_queue_hd != END_TSO_QUEUE) { + ASSERT(blocked_queue_tl != END_TSO_QUEUE); + blocked_queue_hd = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd); + blocked_queue_tl = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl); + } + + if (sleeping_queue != END_TSO_QUEUE) { + sleeping_queue = (StgTSO *)MarkRoot((StgClosure *)sleeping_queue); + } +#endif + + for (m = main_threads; m != NULL; m = m->link) { + m->tso = (StgTSO *)MarkRoot((StgClosure *)m->tso); + } + if (suspended_ccalling_threads != END_TSO_QUEUE) + suspended_ccalling_threads = + (StgTSO *)MarkRoot((StgClosure *)suspended_ccalling_threads); + +#if defined(SMP) || defined(PAR) || defined(GRAN) + markSparkQueue(); +#endif +} + +/* ----------------------------------------------------------------------------- + performGC + + This is the interface to the garbage collector from Haskell land. + We provide this so that external C code can allocate and garbage + collect when called from Haskell via _ccall_GC. + + It might be useful to provide an interface whereby the programmer + can specify more roots (ToDo). + + This needs to be protected by the GC condition variable above. KH. + -------------------------------------------------------------------------- */ + +void (*extra_roots)(void); + +void +performGC(void) +{ + GarbageCollect(GetRoots,rtsFalse); +} + +void +performMajorGC(void) +{ + GarbageCollect(GetRoots,rtsTrue); +} + +static void +AllRoots(void) +{ + GetRoots(); /* the scheduler's roots */ + extra_roots(); /* the user's roots */ +} + +void +performGCWithRoots(void (*get_roots)(void)) +{ + extra_roots = get_roots; + + GarbageCollect(AllRoots,rtsFalse); +} + +/* ----------------------------------------------------------------------------- + Stack overflow + + If the thread has reached its maximum stack size, then raise the + StackOverflow exception in the offending thread. Otherwise + relocate the TSO into a larger chunk of memory and adjust its stack + size appropriately. + -------------------------------------------------------------------------- */ + +static StgTSO * +threadStackOverflow(StgTSO *tso) +{ + nat new_stack_size, new_tso_size, diff, stack_words; + StgPtr new_sp; + StgTSO *dest; + + IF_DEBUG(sanity,checkTSO(tso)); + if (tso->stack_size >= tso->max_stack_size) { + + IF_DEBUG(gc, + belch("@@ threadStackOverflow of TSO %d (%p): stack too large (now %ld; max is %ld", + tso->id, tso, tso->stack_size, tso->max_stack_size); + /* If we're debugging, just print out the top of the stack */ + printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size, + tso->sp+64))); + + /* Send this thread the StackOverflow exception */ + raiseAsync(tso, (StgClosure *)stackOverflow_closure); + return tso; + } + + /* Try to double the current stack size. If that takes us over the + * maximum stack size for this thread, then use the maximum instead. + * Finally round up so the TSO ends up as a whole number of blocks. + */ + new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size); + new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) + + TSO_STRUCT_SIZE)/sizeof(W_); + new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */ + new_stack_size = new_tso_size - TSO_STRUCT_SIZEW; + + IF_DEBUG(scheduler, fprintf(stderr,"== scheduler: increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size)); + + dest = (StgTSO *)allocate(new_tso_size); + TICK_ALLOC_TSO(new_tso_size-sizeofW(StgTSO),0); + + /* copy the TSO block and the old stack into the new area */ + memcpy(dest,tso,TSO_STRUCT_SIZE); + stack_words = tso->stack + tso->stack_size - tso->sp; + new_sp = (P_)dest + new_tso_size - stack_words; + memcpy(new_sp, tso->sp, stack_words * sizeof(W_)); + + /* relocate the stack pointers... */ + diff = (P_)new_sp - (P_)tso->sp; /* In *words* */ + dest->su = (StgUpdateFrame *) ((P_)dest->su + diff); + dest->sp = new_sp; + dest->stack_size = new_stack_size; + + /* and relocate the update frame list */ + relocate_TSO(tso, dest); + + /* Mark the old TSO as relocated. We have to check for relocated + * TSOs in the garbage collector and any primops that deal with TSOs. + * + * It's important to set the sp and su values to just beyond the end + * of the stack, so we don't attempt to scavenge any part of the + * dead TSO's stack. + */ + tso->what_next = ThreadRelocated; + tso->link = dest; + tso->sp = (P_)&(tso->stack[tso->stack_size]); + tso->su = (StgUpdateFrame *)tso->sp; + tso->why_blocked = NotBlocked; + dest->mut_link = NULL; + + IF_PAR_DEBUG(verbose, + belch("@@ threadStackOverflow of TSO %d (now at %p): stack size increased to %ld", + tso->id, tso, tso->stack_size); + /* If we're debugging, just print out the top of the stack */ + printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size, + tso->sp+64))); + + IF_DEBUG(sanity,checkTSO(tso)); +#if 0 + IF_DEBUG(scheduler,printTSO(dest)); +#endif + + return dest; +} + +//@node Blocking Queue Routines, Exception Handling Routines, Garbage Collextion Routines, Main scheduling code +//@subsection Blocking Queue Routines + +/* --------------------------------------------------------------------------- + Wake up a queue that was blocked on some resource. + ------------------------------------------------------------------------ */ + +#if defined(GRAN) +static inline void +unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node ) +{ +} +#elif defined(PAR) +static inline void +unblockCount ( StgBlockingQueueElement *bqe, StgClosure *node ) +{ + /* write RESUME events to log file and + update blocked and fetch time (depending on type of the orig closure) */ + if (RtsFlags.ParFlags.ParStats.Full) { + DumpRawGranEvent(CURRENT_PROC, CURRENT_PROC, + GR_RESUMEQ, ((StgTSO *)bqe), ((StgTSO *)bqe)->block_info.closure, + 0, 0 /* spark_queue_len(ADVISORY_POOL) */); + if (EMPTY_RUN_QUEUE()) + emitSchedule = rtsTrue; + + switch (get_itbl(node)->type) { + case FETCH_ME_BQ: + ((StgTSO *)bqe)->par.fetchtime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat; + break; + case RBH: + case FETCH_ME: + case BLACKHOLE_BQ: + ((StgTSO *)bqe)->par.blocktime += CURRENT_TIME-((StgTSO *)bqe)->par.blockedat; + break; +#ifdef DIST + case MVAR: + break; +#endif + default: + barf("{unblockOneLocked}Daq Qagh: unexpected closure in blocking queue"); + } + } +} +#endif + +#if defined(GRAN) +static StgBlockingQueueElement * +unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node) +{ + StgTSO *tso; + PEs node_loc, tso_loc; + + node_loc = where_is(node); // should be lifted out of loop + tso = (StgTSO *)bqe; // wastes an assignment to get the type right + tso_loc = where_is((StgClosure *)tso); + if (IS_LOCAL_TO(PROCS(node),tso_loc)) { // TSO is local + /* !fake_fetch => TSO is on CurrentProc is same as IS_LOCAL_TO */ + ASSERT(CurrentProc!=node_loc || tso_loc==CurrentProc); + CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.lunblocktime; + // insertThread(tso, node_loc); + new_event(tso_loc, tso_loc, CurrentTime[CurrentProc], + ResumeThread, + tso, node, (rtsSpark*)NULL); + tso->link = END_TSO_QUEUE; // overwrite link just to be sure + // len_local++; + // len++; + } else { // TSO is remote (actually should be FMBQ) + CurrentTime[CurrentProc] += RtsFlags.GranFlags.Costs.mpacktime + + RtsFlags.GranFlags.Costs.gunblocktime + + RtsFlags.GranFlags.Costs.latency; + new_event(tso_loc, CurrentProc, CurrentTime[CurrentProc], + UnblockThread, + tso, node, (rtsSpark*)NULL); + tso->link = END_TSO_QUEUE; // overwrite link just to be sure + // len++; + } + /* the thread-queue-overhead is accounted for in either Resume or UnblockThread */ + IF_GRAN_DEBUG(bq, + fprintf(stderr," %s TSO %d (%p) [PE %d] (block_info.closure=%p) (next=%p) ,", + (node_loc==tso_loc ? "Local" : "Global"), + tso->id, tso, CurrentProc, tso->block_info.closure, tso->link)); + tso->block_info.closure = NULL; + IF_DEBUG(scheduler,belch("-- Waking up thread %ld (%p)", + tso->id, tso)); +} +#elif defined(PAR) +static StgBlockingQueueElement * +unblockOneLocked(StgBlockingQueueElement *bqe, StgClosure *node) +{ + StgBlockingQueueElement *next; + + switch (get_itbl(bqe)->type) { + case TSO: + ASSERT(((StgTSO *)bqe)->why_blocked != NotBlocked); + /* if it's a TSO just push it onto the run_queue */ + next = bqe->link; + // ((StgTSO *)bqe)->link = END_TSO_QUEUE; // debugging? + PUSH_ON_RUN_QUEUE((StgTSO *)bqe); + THREAD_RUNNABLE(); + unblockCount(bqe, node); + /* reset blocking status after dumping event */ + ((StgTSO *)bqe)->why_blocked = NotBlocked; + break; + + case BLOCKED_FETCH: + /* if it's a BLOCKED_FETCH put it on the PendingFetches list */ + next = bqe->link; + bqe->link = (StgBlockingQueueElement *)PendingFetches; + PendingFetches = (StgBlockedFetch *)bqe; + break; + +# if defined(DEBUG) + /* can ignore this case in a non-debugging setup; + see comments on RBHSave closures above */ + case CONSTR: + /* check that the closure is an RBHSave closure */ + ASSERT(get_itbl((StgClosure *)bqe) == &stg_RBH_Save_0_info || + get_itbl((StgClosure *)bqe) == &stg_RBH_Save_1_info || + get_itbl((StgClosure *)bqe) == &stg_RBH_Save_2_info); + break; + + default: + barf("{unblockOneLocked}Daq Qagh: Unexpected IP (%#lx; %s) in blocking queue at %#lx\n", + get_itbl((StgClosure *)bqe), info_type((StgClosure *)bqe), + (StgClosure *)bqe); +# endif + } + IF_PAR_DEBUG(bq, fprintf(stderr, ", %p (%s)", bqe, info_type((StgClosure*)bqe))); + return next; +} + +#else /* !GRAN && !PAR */ +static StgTSO * +unblockOneLocked(StgTSO *tso) +{ + StgTSO *next; + + ASSERT(get_itbl(tso)->type == TSO); + ASSERT(tso->why_blocked != NotBlocked); + tso->why_blocked = NotBlocked; + next = tso->link; + PUSH_ON_RUN_QUEUE(tso); + THREAD_RUNNABLE(); + IF_DEBUG(scheduler,sched_belch("waking up thread %ld", tso->id)); + return next; +} +#endif + +#if defined(GRAN) || defined(PAR) +inline StgBlockingQueueElement * +unblockOne(StgBlockingQueueElement *bqe, StgClosure *node) +{ + ACQUIRE_LOCK(&sched_mutex); + bqe = unblockOneLocked(bqe, node); + RELEASE_LOCK(&sched_mutex); + return bqe; +} +#else +inline StgTSO * +unblockOne(StgTSO *tso) +{ + ACQUIRE_LOCK(&sched_mutex); + tso = unblockOneLocked(tso); + RELEASE_LOCK(&sched_mutex); + return tso; +} +#endif + +#if defined(GRAN) +void +awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node) +{ + StgBlockingQueueElement *bqe; + PEs node_loc; + nat len = 0; + + IF_GRAN_DEBUG(bq, + belch("##-_ AwBQ for node %p on PE %d @ %ld by TSO %d (%p): ", \ + node, CurrentProc, CurrentTime[CurrentProc], + CurrentTSO->id, CurrentTSO)); + + node_loc = where_is(node); + + ASSERT(q == END_BQ_QUEUE || + get_itbl(q)->type == TSO || // q is either a TSO or an RBHSave + get_itbl(q)->type == CONSTR); // closure (type constructor) + ASSERT(is_unique(node)); + + /* FAKE FETCH: magically copy the node to the tso's proc; + no Fetch necessary because in reality the node should not have been + moved to the other PE in the first place + */ + if (CurrentProc!=node_loc) { + IF_GRAN_DEBUG(bq, + belch("## node %p is on PE %d but CurrentProc is %d (TSO %d); assuming fake fetch and adjusting bitmask (old: %#x)", + node, node_loc, CurrentProc, CurrentTSO->id, + // CurrentTSO, where_is(CurrentTSO), + node->header.gran.procs)); + node->header.gran.procs = (node->header.gran.procs) | PE_NUMBER(CurrentProc); + IF_GRAN_DEBUG(bq, + belch("## new bitmask of node %p is %#x", + node, node->header.gran.procs)); + if (RtsFlags.GranFlags.GranSimStats.Global) { + globalGranStats.tot_fake_fetches++; + } + } + + bqe = q; + // ToDo: check: ASSERT(CurrentProc==node_loc); + while (get_itbl(bqe)->type==TSO) { // q != END_TSO_QUEUE) { + //next = bqe->link; + /* + bqe points to the current element in the queue + next points to the next element in the queue + */ + //tso = (StgTSO *)bqe; // wastes an assignment to get the type right + //tso_loc = where_is(tso); + len++; + bqe = unblockOneLocked(bqe, node); + } + + /* if this is the BQ of an RBH, we have to put back the info ripped out of + the closure to make room for the anchor of the BQ */ + if (bqe!=END_BQ_QUEUE) { + ASSERT(get_itbl(node)->type == RBH && get_itbl(bqe)->type == CONSTR); + /* + ASSERT((info_ptr==&RBH_Save_0_info) || + (info_ptr==&RBH_Save_1_info) || + (info_ptr==&RBH_Save_2_info)); + */ + /* cf. convertToRBH in RBH.c for writing the RBHSave closure */ + ((StgRBH *)node)->blocking_queue = (StgBlockingQueueElement *)((StgRBHSave *)bqe)->payload[0]; + ((StgRBH *)node)->mut_link = (StgMutClosure *)((StgRBHSave *)bqe)->payload[1]; + + IF_GRAN_DEBUG(bq, + belch("## Filled in RBH_Save for %p (%s) at end of AwBQ", + node, info_type(node))); + } + + /* statistics gathering */ + if (RtsFlags.GranFlags.GranSimStats.Global) { + // globalGranStats.tot_bq_processing_time += bq_processing_time; + globalGranStats.tot_bq_len += len; // total length of all bqs awakened + // globalGranStats.tot_bq_len_local += len_local; // same for local TSOs only + globalGranStats.tot_awbq++; // total no. of bqs awakened + } + IF_GRAN_DEBUG(bq, + fprintf(stderr,"## BQ Stats of %p: [%d entries] %s\n", + node, len, (bqe!=END_BQ_QUEUE) ? "RBH" : "")); +} +#elif defined(PAR) +void +awakenBlockedQueue(StgBlockingQueueElement *q, StgClosure *node) +{ + StgBlockingQueueElement *bqe; + + ACQUIRE_LOCK(&sched_mutex); + + IF_PAR_DEBUG(verbose, + belch("##-_ AwBQ for node %p on [%x]: ", + node, mytid)); +#ifdef DIST + //RFP + if(get_itbl(q)->type == CONSTR || q==END_BQ_QUEUE) { + IF_PAR_DEBUG(verbose, belch("## ... nothing to unblock so lets just return. RFP (BUG?)")); + return; + } +#endif + + ASSERT(q == END_BQ_QUEUE || + get_itbl(q)->type == TSO || + get_itbl(q)->type == BLOCKED_FETCH || + get_itbl(q)->type == CONSTR); + + bqe = q; + while (get_itbl(bqe)->type==TSO || + get_itbl(bqe)->type==BLOCKED_FETCH) { + bqe = unblockOneLocked(bqe, node); + } + RELEASE_LOCK(&sched_mutex); +} + +#else /* !GRAN && !PAR */ +void +awakenBlockedQueue(StgTSO *tso) +{ + ACQUIRE_LOCK(&sched_mutex); + while (tso != END_TSO_QUEUE) { + tso = unblockOneLocked(tso); + } + RELEASE_LOCK(&sched_mutex); +} +#endif + +//@node Exception Handling Routines, Debugging Routines, Blocking Queue Routines, Main scheduling code +//@subsection Exception Handling Routines + +/* --------------------------------------------------------------------------- + Interrupt execution + - usually called inside a signal handler so it mustn't do anything fancy. + ------------------------------------------------------------------------ */ + +void +interruptStgRts(void) +{ + interrupted = 1; + context_switch = 1; +} + +/* ----------------------------------------------------------------------------- + Unblock a thread + + This is for use when we raise an exception in another thread, which + may be blocked. + This has nothing to do with the UnblockThread event in GranSim. -- HWL + -------------------------------------------------------------------------- */ + +#if defined(GRAN) || defined(PAR) +/* + NB: only the type of the blocking queue is different in GranSim and GUM + the operations on the queue-elements are the same + long live polymorphism! +*/ +static void +unblockThread(StgTSO *tso) +{ + StgBlockingQueueElement *t, **last; + + ACQUIRE_LOCK(&sched_mutex); + switch (tso->why_blocked) { + + case NotBlocked: + return; /* not blocked */ + + case BlockedOnMVar: + ASSERT(get_itbl(tso->block_info.closure)->type == MVAR); + { + StgBlockingQueueElement *last_tso = END_BQ_QUEUE; + StgMVar *mvar = (StgMVar *)(tso->block_info.closure); + + last = (StgBlockingQueueElement **)&mvar->head; + for (t = (StgBlockingQueueElement *)mvar->head; + t != END_BQ_QUEUE; + last = &t->link, last_tso = t, t = t->link) { + if (t == (StgBlockingQueueElement *)tso) { + *last = (StgBlockingQueueElement *)tso->link; + if (mvar->tail == tso) { + mvar->tail = (StgTSO *)last_tso; + } + goto done; + } + } + barf("unblockThread (MVAR): TSO not found"); + } - if (blocked_queue_hd != END_TSO_QUEUE) { - return AllBlocked; - } else { - return Deadlock; - } -} + case BlockedOnBlackHole: + ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ); + { + StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure); + + last = &bq->blocking_queue; + for (t = bq->blocking_queue; + t != END_BQ_QUEUE; + last = &t->link, t = t->link) { + if (t == (StgBlockingQueueElement *)tso) { + *last = (StgBlockingQueueElement *)tso->link; + goto done; + } + } + barf("unblockThread (BLACKHOLE): TSO not found"); + } -/* ----------------------------------------------------------------------------- - Where are the roots that we know about? + case BlockedOnException: + { + StgTSO *target = tso->block_info.tso; - - all the threads on the runnable queue - - all the threads on the blocked queue - - all the thread currently executing a _ccall_GC - - all the "main threads" - - -------------------------------------------------------------------------- */ + ASSERT(get_itbl(target)->type == TSO); -static void GetRoots(void) -{ - nat i; + if (target->what_next == ThreadRelocated) { + target = target->link; + ASSERT(get_itbl(target)->type == TSO); + } + + ASSERT(target->blocked_exceptions != NULL); - run_queue_hd = (StgTSO *)MarkRoot((StgClosure *)run_queue_hd); - run_queue_tl = (StgTSO *)MarkRoot((StgClosure *)run_queue_tl); + last = (StgBlockingQueueElement **)&target->blocked_exceptions; + for (t = (StgBlockingQueueElement *)target->blocked_exceptions; + t != END_BQ_QUEUE; + last = &t->link, t = t->link) { + ASSERT(get_itbl(t)->type == TSO); + if (t == (StgBlockingQueueElement *)tso) { + *last = (StgBlockingQueueElement *)tso->link; + goto done; + } + } + barf("unblockThread (Exception): TSO not found"); + } - blocked_queue_hd = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_hd); - blocked_queue_tl = (StgTSO *)MarkRoot((StgClosure *)blocked_queue_tl); + case BlockedOnRead: + case BlockedOnWrite: + { + /* take TSO off blocked_queue */ + StgBlockingQueueElement *prev = NULL; + for (t = (StgBlockingQueueElement *)blocked_queue_hd; t != END_BQ_QUEUE; + prev = t, t = t->link) { + if (t == (StgBlockingQueueElement *)tso) { + if (prev == NULL) { + blocked_queue_hd = (StgTSO *)t->link; + if ((StgBlockingQueueElement *)blocked_queue_tl == t) { + blocked_queue_tl = END_TSO_QUEUE; + } + } else { + prev->link = t->link; + if ((StgBlockingQueueElement *)blocked_queue_tl == t) { + blocked_queue_tl = (StgTSO *)prev; + } + } + goto done; + } + } + barf("unblockThread (I/O): TSO not found"); + } - ccalling_threads = (StgTSO *)MarkRoot((StgClosure *)ccalling_threads); + case BlockedOnDelay: + { + /* take TSO off sleeping_queue */ + StgBlockingQueueElement *prev = NULL; + for (t = (StgBlockingQueueElement *)sleeping_queue; t != END_BQ_QUEUE; + prev = t, t = t->link) { + if (t == (StgBlockingQueueElement *)tso) { + if (prev == NULL) { + sleeping_queue = (StgTSO *)t->link; + } else { + prev->link = t->link; + } + goto done; + } + } + barf("unblockThread (I/O): TSO not found"); + } - for (i = 0; i < next_main_thread; i++) { - main_threads[i] = (StgTSO *)MarkRoot((StgClosure *)main_threads[i]); + default: + barf("unblockThread"); } + + done: + tso->link = END_TSO_QUEUE; + tso->why_blocked = NotBlocked; + tso->block_info.closure = NULL; + PUSH_ON_RUN_QUEUE(tso); + RELEASE_LOCK(&sched_mutex); } +#else +static void +unblockThread(StgTSO *tso) +{ + StgTSO *t, **last; + + ACQUIRE_LOCK(&sched_mutex); + switch (tso->why_blocked) { + + case NotBlocked: + return; /* not blocked */ + + case BlockedOnMVar: + ASSERT(get_itbl(tso->block_info.closure)->type == MVAR); + { + StgTSO *last_tso = END_TSO_QUEUE; + StgMVar *mvar = (StgMVar *)(tso->block_info.closure); + + last = &mvar->head; + for (t = mvar->head; t != END_TSO_QUEUE; + last = &t->link, last_tso = t, t = t->link) { + if (t == tso) { + *last = tso->link; + if (mvar->tail == tso) { + mvar->tail = last_tso; + } + goto done; + } + } + barf("unblockThread (MVAR): TSO not found"); + } -/* ----------------------------------------------------------------------------- - performGC + case BlockedOnBlackHole: + ASSERT(get_itbl(tso->block_info.closure)->type == BLACKHOLE_BQ); + { + StgBlockingQueue *bq = (StgBlockingQueue *)(tso->block_info.closure); + + last = &bq->blocking_queue; + for (t = bq->blocking_queue; t != END_TSO_QUEUE; + last = &t->link, t = t->link) { + if (t == tso) { + *last = tso->link; + goto done; + } + } + barf("unblockThread (BLACKHOLE): TSO not found"); + } - This is the interface to the garbage collector from Haskell land. - We provide this so that external C code can allocate and garbage - collect when called from Haskell via _ccall_GC. + case BlockedOnException: + { + StgTSO *target = tso->block_info.tso; - It might be useful to provide an interface whereby the programmer - can specify more roots (ToDo). - -------------------------------------------------------------------------- */ + ASSERT(get_itbl(target)->type == TSO); -void (*extra_roots)(void); + while (target->what_next == ThreadRelocated) { + target = target->link; + ASSERT(get_itbl(target)->type == TSO); + } + + ASSERT(target->blocked_exceptions != NULL); + + last = &target->blocked_exceptions; + for (t = target->blocked_exceptions; t != END_TSO_QUEUE; + last = &t->link, t = t->link) { + ASSERT(get_itbl(t)->type == TSO); + if (t == tso) { + *last = tso->link; + goto done; + } + } + barf("unblockThread (Exception): TSO not found"); + } -void -performGC(void) -{ - GarbageCollect(GetRoots); + case BlockedOnRead: + case BlockedOnWrite: + { + StgTSO *prev = NULL; + for (t = blocked_queue_hd; t != END_TSO_QUEUE; + prev = t, t = t->link) { + if (t == tso) { + if (prev == NULL) { + blocked_queue_hd = t->link; + if (blocked_queue_tl == t) { + blocked_queue_tl = END_TSO_QUEUE; + } + } else { + prev->link = t->link; + if (blocked_queue_tl == t) { + blocked_queue_tl = prev; + } + } + goto done; + } + } + barf("unblockThread (I/O): TSO not found"); + } + + case BlockedOnDelay: + { + StgTSO *prev = NULL; + for (t = sleeping_queue; t != END_TSO_QUEUE; + prev = t, t = t->link) { + if (t == tso) { + if (prev == NULL) { + sleeping_queue = t->link; + } else { + prev->link = t->link; + } + goto done; + } + } + barf("unblockThread (I/O): TSO not found"); + } + + default: + barf("unblockThread"); + } + + done: + tso->link = END_TSO_QUEUE; + tso->why_blocked = NotBlocked; + tso->block_info.closure = NULL; + PUSH_ON_RUN_QUEUE(tso); + RELEASE_LOCK(&sched_mutex); } +#endif -static void -AllRoots(void) +/* ----------------------------------------------------------------------------- + * raiseAsync() + * + * The following function implements the magic for raising an + * asynchronous exception in an existing thread. + * + * We first remove the thread from any queue on which it might be + * blocked. The possible blockages are MVARs and BLACKHOLE_BQs. + * + * We strip the stack down to the innermost CATCH_FRAME, building + * thunks in the heap for all the active computations, so they can + * be restarted if necessary. When we reach a CATCH_FRAME, we build + * an application of the handler to the exception, and push it on + * the top of the stack. + * + * How exactly do we save all the active computations? We create an + * AP_UPD for every UpdateFrame on the stack. Entering one of these + * AP_UPDs pushes everything from the corresponding update frame + * upwards onto the stack. (Actually, it pushes everything up to the + * next update frame plus a pointer to the next AP_UPD object. + * Entering the next AP_UPD object pushes more onto the stack until we + * reach the last AP_UPD object - at which point the stack should look + * exactly as it did when we killed the TSO and we can continue + * execution by entering the closure on top of the stack. + * + * We can also kill a thread entirely - this happens if either (a) the + * exception passed to raiseAsync is NULL, or (b) there's no + * CATCH_FRAME on the stack. In either case, we strip the entire + * stack and replace the thread with a zombie. + * + * -------------------------------------------------------------------------- */ + +void +deleteThread(StgTSO *tso) { - GetRoots(); /* the scheduler's roots */ - extra_roots(); /* the user's roots */ + raiseAsync(tso,NULL); } void -performGCWithRoots(void (*get_roots)(void)) +raiseAsync(StgTSO *tso, StgClosure *exception) { - extra_roots = get_roots; + StgUpdateFrame* su = tso->su; + StgPtr sp = tso->sp; + + /* Thread already dead? */ + if (tso->what_next == ThreadComplete || tso->what_next == ThreadKilled) { + return; + } + + IF_DEBUG(scheduler, sched_belch("raising exception in thread %ld.", tso->id)); + + /* Remove it from any blocking queues */ + unblockThread(tso); + + /* The stack freezing code assumes there's a closure pointer on + * the top of the stack. This isn't always the case with compiled + * code, so we have to push a dummy closure on the top which just + * returns to the next return address on the stack. + */ + if ( LOOKS_LIKE_GHC_INFO((void*)*sp) ) { + *(--sp) = (W_)&stg_dummy_ret_closure; + } + + while (1) { + int words = ((P_)su - (P_)sp) - 1; + nat i; + StgAP_UPD * ap; + + /* If we find a CATCH_FRAME, and we've got an exception to raise, + * then build PAP(handler,exception,realworld#), and leave it on + * top of the stack ready to enter. + */ + if (get_itbl(su)->type == CATCH_FRAME && exception != NULL) { + StgCatchFrame *cf = (StgCatchFrame *)su; + /* we've got an exception to raise, so let's pass it to the + * handler in this frame. + */ + ap = (StgAP_UPD *)allocate(sizeofW(StgPAP) + 2); + TICK_ALLOC_UPD_PAP(3,0); + SET_HDR(ap,&stg_PAP_info,cf->header.prof.ccs); + + ap->n_args = 2; + ap->fun = cf->handler; /* :: Exception -> IO a */ + ap->payload[0] = exception; + ap->payload[1] = ARG_TAG(0); /* realworld token */ + + /* throw away the stack from Sp up to and including the + * CATCH_FRAME. + */ + sp = (P_)su + sizeofW(StgCatchFrame) - 1; + tso->su = cf->link; + + /* Restore the blocked/unblocked state for asynchronous exceptions + * at the CATCH_FRAME. + * + * If exceptions were unblocked at the catch, arrange that they + * are unblocked again after executing the handler by pushing an + * unblockAsyncExceptions_ret stack frame. + */ + if (!cf->exceptions_blocked) { + *(sp--) = (W_)&stg_unblockAsyncExceptionszh_ret_info; + } + + /* Ensure that async exceptions are blocked when running the handler. + */ + if (tso->blocked_exceptions == NULL) { + tso->blocked_exceptions = END_TSO_QUEUE; + } + + /* Put the newly-built PAP on top of the stack, ready to execute + * when the thread restarts. + */ + sp[0] = (W_)ap; + tso->sp = sp; + tso->what_next = ThreadEnterGHC; + IF_DEBUG(sanity, checkTSO(tso)); + return; + } + + /* First build an AP_UPD consisting of the stack chunk above the + * current update frame, with the top word on the stack as the + * fun field. + */ + ap = (StgAP_UPD *)allocate(AP_sizeW(words)); + + ASSERT(words >= 0); + + ap->n_args = words; + ap->fun = (StgClosure *)sp[0]; + sp++; + for(i=0; i < (nat)words; ++i) { + ap->payload[i] = (StgClosure *)*sp++; + } + + switch (get_itbl(su)->type) { + + case UPDATE_FRAME: + { + SET_HDR(ap,&stg_AP_UPD_info,su->header.prof.ccs /* ToDo */); + TICK_ALLOC_UP_THK(words+1,0); + + IF_DEBUG(scheduler, + fprintf(stderr, "scheduler: Updating "); + printPtr((P_)su->updatee); + fprintf(stderr, " with "); + printObj((StgClosure *)ap); + ); + + /* Replace the updatee with an indirection - happily + * this will also wake up any threads currently + * waiting on the result. + * + * Warning: if we're in a loop, more than one update frame on + * the stack may point to the same object. Be careful not to + * overwrite an IND_OLDGEN in this case, because we'll screw + * up the mutable lists. To be on the safe side, don't + * overwrite any kind of indirection at all. See also + * threadSqueezeStack in GC.c, where we have to make a similar + * check. + */ + if (!closure_IND(su->updatee)) { + UPD_IND_NOLOCK(su->updatee,ap); /* revert the black hole */ + } + su = su->link; + sp += sizeofW(StgUpdateFrame) -1; + sp[0] = (W_)ap; /* push onto stack */ + break; + } + + case CATCH_FRAME: + { + StgCatchFrame *cf = (StgCatchFrame *)su; + StgClosure* o; + + /* We want a PAP, not an AP_UPD. Fortunately, the + * layout's the same. + */ + SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */); + TICK_ALLOC_UPD_PAP(words+1,0); + + /* now build o = FUN(catch,ap,handler) */ + o = (StgClosure *)allocate(sizeofW(StgClosure)+2); + TICK_ALLOC_FUN(2,0); + SET_HDR(o,&stg_catch_info,su->header.prof.ccs /* ToDo */); + o->payload[0] = (StgClosure *)ap; + o->payload[1] = cf->handler; + + IF_DEBUG(scheduler, + fprintf(stderr, "scheduler: Built "); + printObj((StgClosure *)o); + ); + + /* pop the old handler and put o on the stack */ + su = cf->link; + sp += sizeofW(StgCatchFrame) - 1; + sp[0] = (W_)o; + break; + } + + case SEQ_FRAME: + { + StgSeqFrame *sf = (StgSeqFrame *)su; + StgClosure* o; + + SET_HDR(ap,&stg_PAP_info,su->header.prof.ccs /* ToDo */); + TICK_ALLOC_UPD_PAP(words+1,0); + + /* now build o = FUN(seq,ap) */ + o = (StgClosure *)allocate(sizeofW(StgClosure)+1); + TICK_ALLOC_SE_THK(1,0); + SET_HDR(o,&stg_seq_info,su->header.prof.ccs /* ToDo */); + o->payload[0] = (StgClosure *)ap; + + IF_DEBUG(scheduler, + fprintf(stderr, "scheduler: Built "); + printObj((StgClosure *)o); + ); + + /* pop the old handler and put o on the stack */ + su = sf->link; + sp += sizeofW(StgSeqFrame) - 1; + sp[0] = (W_)o; + break; + } + + case STOP_FRAME: + /* We've stripped the entire stack, the thread is now dead. */ + sp += sizeofW(StgStopFrame) - 1; + sp[0] = (W_)exception; /* save the exception */ + tso->what_next = ThreadKilled; + tso->su = (StgUpdateFrame *)(sp+1); + tso->sp = sp; + return; - GarbageCollect(AllRoots); + default: + barf("raiseAsync"); + } + } + barf("raiseAsync"); } /* ----------------------------------------------------------------------------- - Stack overflow - - If the thread has reached its maximum stack size, - then bomb out. Otherwise relocate the TSO into a larger chunk of - memory and adjust its stack size appropriately. + resurrectThreads is called after garbage collection on the list of + threads found to be garbage. Each of these threads will be woken + up and sent a signal: BlockedOnDeadMVar if the thread was blocked + on an MVar, or NonTermination if the thread was blocked on a Black + Hole. -------------------------------------------------------------------------- */ -static StgTSO * -threadStackOverflow(StgTSO *tso) +void +resurrectThreads( StgTSO *threads ) { - nat new_stack_size, new_tso_size, diff, stack_words; - StgPtr new_sp; - StgTSO *dest; - - if (tso->stack_size >= tso->max_stack_size) { - /* ToDo: just kill this thread? */ -#ifdef DEBUG - /* If we're debugging, just print out the top of the stack */ - printStackChunk(tso->sp, stg_min(tso->stack+tso->stack_size, - tso->sp+64)); -#endif - stackOverflow(tso->max_stack_size); + StgTSO *tso, *next; + + for (tso = threads; tso != END_TSO_QUEUE; tso = next) { + next = tso->global_link; + tso->global_link = all_threads; + all_threads = tso; + IF_DEBUG(scheduler, sched_belch("resurrecting thread %d", tso->id)); + + switch (tso->why_blocked) { + case BlockedOnMVar: + case BlockedOnException: + raiseAsync(tso,(StgClosure *)BlockedOnDeadMVar_closure); + break; + case BlockedOnBlackHole: + raiseAsync(tso,(StgClosure *)NonTermination_closure); + break; + case NotBlocked: + /* This might happen if the thread was blocked on a black hole + * belonging to a thread that we've just woken up (raiseAsync + * can wake up threads, remember...). + */ + continue; + default: + barf("resurrectThreads: thread blocked in a strange way"); + } } +} - /* Try to double the current stack size. If that takes us over the - * maximum stack size for this thread, then use the maximum instead. - * Finally round up so the TSO ends up as a whole number of blocks. - */ - new_stack_size = stg_min(tso->stack_size * 2, tso->max_stack_size); - new_tso_size = (nat)BLOCK_ROUND_UP(new_stack_size * sizeof(W_) + - TSO_STRUCT_SIZE)/sizeof(W_); - new_tso_size = round_to_mblocks(new_tso_size); /* Be MBLOCK-friendly */ - new_stack_size = new_tso_size - TSO_STRUCT_SIZEW; +/* ----------------------------------------------------------------------------- + * Blackhole detection: if we reach a deadlock, test whether any + * threads are blocked on themselves. Any threads which are found to + * be self-blocked get sent a NonTermination exception. + * + * This is only done in a deadlock situation in order to avoid + * performance overhead in the normal case. + * -------------------------------------------------------------------------- */ - IF_DEBUG(scheduler, fprintf(stderr,"increasing stack size from %d words to %d.\n", tso->stack_size, new_stack_size)); +static void +detectBlackHoles( void ) +{ + StgTSO *t = all_threads; + StgUpdateFrame *frame; + StgClosure *blocked_on; - dest = (StgTSO *)allocate(new_tso_size); - TICK_ALLOC_TSO(new_tso_size-sizeofW(StgTSO),0); + for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) { - /* copy the TSO block and the old stack into the new area */ - memcpy(dest,tso,TSO_STRUCT_SIZE); - stack_words = tso->stack + tso->stack_size - tso->sp; - new_sp = (P_)dest + new_tso_size - stack_words; - memcpy(new_sp, tso->sp, stack_words * sizeof(W_)); + while (t->what_next == ThreadRelocated) { + t = t->link; + ASSERT(get_itbl(t)->type == TSO); + } + + if (t->why_blocked != BlockedOnBlackHole) { + continue; + } - /* relocate the stack pointers... */ - diff = (P_)new_sp - (P_)tso->sp; /* In *words* */ - dest->su = (StgUpdateFrame *) ((P_)dest->su + diff); - dest->sp = new_sp; - dest->splim = (P_)dest->splim + (nat)((P_)dest - (P_)tso); - dest->stack_size = new_stack_size; - - /* and relocate the update frame list */ - relocate_TSO(tso, dest); + blocked_on = t->block_info.closure; + + for (frame = t->su; ; frame = frame->link) { + switch (get_itbl(frame)->type) { + + case UPDATE_FRAME: + if (frame->updatee == blocked_on) { + /* We are blocking on one of our own computations, so + * send this thread the NonTermination exception. + */ + IF_DEBUG(scheduler, + sched_belch("thread %d is blocked on itself", t->id)); + raiseAsync(t, (StgClosure *)NonTermination_closure); + goto done; + } + else { + continue; + } + + case CATCH_FRAME: + case SEQ_FRAME: + continue; + + case STOP_FRAME: + break; + } + break; + } - /* Mark the old one as dead so we don't try to scavenge it during - * garbage collection (the TSO will likely be on a mutables list in - * some generation, but it'll get collected soon enough). - */ - tso->whatNext = ThreadKilled; - dest->mut_link = NULL; + done: ; + } +} - IF_DEBUG(sanity,checkTSO(tso)); -#if 0 - IF_DEBUG(scheduler,printTSO(dest)); +//@node Debugging Routines, Index, Exception Handling Routines, Main scheduling code +//@subsection Debugging Routines + +/* ----------------------------------------------------------------------------- + Debugging: why is a thread blocked + -------------------------------------------------------------------------- */ + +#ifdef DEBUG + +void +printThreadBlockage(StgTSO *tso) +{ + switch (tso->why_blocked) { + case BlockedOnRead: + fprintf(stderr,"is blocked on read from fd %d", tso->block_info.fd); + break; + case BlockedOnWrite: + fprintf(stderr,"is blocked on write to fd %d", tso->block_info.fd); + break; + case BlockedOnDelay: + fprintf(stderr,"is blocked until %d", tso->block_info.target); + break; + case BlockedOnMVar: + fprintf(stderr,"is blocked on an MVar"); + break; + case BlockedOnException: + fprintf(stderr,"is blocked on delivering an exception to thread %d", + tso->block_info.tso->id); + break; + case BlockedOnBlackHole: + fprintf(stderr,"is blocked on a black hole"); + break; + case NotBlocked: + fprintf(stderr,"is not blocked"); + break; +#if defined(PAR) + case BlockedOnGA: + fprintf(stderr,"is blocked on global address; local FM_BQ is %p (%s)", + tso->block_info.closure, info_type(tso->block_info.closure)); + break; + case BlockedOnGA_NoSend: + fprintf(stderr,"is blocked on global address (no send); local FM_BQ is %p (%s)", + tso->block_info.closure, info_type(tso->block_info.closure)); + break; #endif - if (tso == MainTSO) { /* hack */ - MainTSO = dest; + default: + barf("printThreadBlockage: strange tso->why_blocked: %d for TSO %d (%d)", + tso->why_blocked, tso->id, tso); } - return dest; } -/* ----------------------------------------------------------------------------- - Wake up a queue that was blocked on some resource (usually a - computation in progress). - -------------------------------------------------------------------------- */ +void +printThreadStatus(StgTSO *tso) +{ + switch (tso->what_next) { + case ThreadKilled: + fprintf(stderr,"has been killed"); + break; + case ThreadComplete: + fprintf(stderr,"has completed"); + break; + default: + printThreadBlockage(tso); + } +} + +void +printAllThreads(void) +{ + StgTSO *t; -void awaken_blocked_queue(StgTSO *q) +# if defined(GRAN) + char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN]; + ullong_format_string(TIME_ON_PROC(CurrentProc), + time_string, rtsFalse/*no commas!*/); + + sched_belch("all threads at [%s]:", time_string); +# elif defined(PAR) + char time_string[TIME_STR_LEN], node_str[NODE_STR_LEN]; + ullong_format_string(CURRENT_TIME, + time_string, rtsFalse/*no commas!*/); + + sched_belch("all threads at [%s]:", time_string); +# else + sched_belch("all threads:"); +# endif + + for (t = all_threads; t != END_TSO_QUEUE; t = t->global_link) { + fprintf(stderr, "\tthread %d ", t->id); + printThreadStatus(t); + fprintf(stderr,"\n"); + } +} + +/* + Print a whole blocking queue attached to node (debugging only). +*/ +//@cindex print_bq +# if defined(PAR) +void +print_bq (StgClosure *node) { + StgBlockingQueueElement *bqe; StgTSO *tso; + rtsBool end; - while (q != END_TSO_QUEUE) { - ASSERT(get_itbl(q)->type == TSO); - tso = q; - q = tso->link; - PUSH_ON_RUN_QUEUE(tso); - IF_DEBUG(scheduler,belch("Waking up thread %ld", tso->id)); + fprintf(stderr,"## BQ of closure %p (%s): ", + node, info_type(node)); + + /* should cover all closures that may have a blocking queue */ + ASSERT(get_itbl(node)->type == BLACKHOLE_BQ || + get_itbl(node)->type == FETCH_ME_BQ || + get_itbl(node)->type == RBH || + get_itbl(node)->type == MVAR); + + ASSERT(node!=(StgClosure*)NULL); // sanity check + + print_bqe(((StgBlockingQueue*)node)->blocking_queue); +} + +/* + Print a whole blocking queue starting with the element bqe. +*/ +void +print_bqe (StgBlockingQueueElement *bqe) +{ + rtsBool end; + + /* + NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure; + */ + for (end = (bqe==END_BQ_QUEUE); + !end; // iterate until bqe points to a CONSTR + end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), + bqe = end ? END_BQ_QUEUE : bqe->link) { + ASSERT(bqe != END_BQ_QUEUE); // sanity check + ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check + /* types of closures that may appear in a blocking queue */ + ASSERT(get_itbl(bqe)->type == TSO || + get_itbl(bqe)->type == BLOCKED_FETCH || + get_itbl(bqe)->type == CONSTR); + /* only BQs of an RBH end with an RBH_Save closure */ + //ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH); + + switch (get_itbl(bqe)->type) { + case TSO: + fprintf(stderr," TSO %u (%x),", + ((StgTSO *)bqe)->id, ((StgTSO *)bqe)); + break; + case BLOCKED_FETCH: + fprintf(stderr," BF (node=%p, ga=((%x, %d, %x)),", + ((StgBlockedFetch *)bqe)->node, + ((StgBlockedFetch *)bqe)->ga.payload.gc.gtid, + ((StgBlockedFetch *)bqe)->ga.payload.gc.slot, + ((StgBlockedFetch *)bqe)->ga.weight); + break; + case CONSTR: + fprintf(stderr," %s (IP %p),", + (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" : + get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" : + get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" : + "RBH_Save_?"), get_itbl(bqe)); + break; + default: + barf("Unexpected closure type %s in blocking queue", // of %p (%s)", + info_type((StgClosure *)bqe)); // , node, info_type(node)); + break; + } + } /* for */ + fputc('\n', stderr); +} +# elif defined(GRAN) +void +print_bq (StgClosure *node) +{ + StgBlockingQueueElement *bqe; + PEs node_loc, tso_loc; + rtsBool end; + + /* should cover all closures that may have a blocking queue */ + ASSERT(get_itbl(node)->type == BLACKHOLE_BQ || + get_itbl(node)->type == FETCH_ME_BQ || + get_itbl(node)->type == RBH); + + ASSERT(node!=(StgClosure*)NULL); // sanity check + node_loc = where_is(node); + + fprintf(stderr,"## BQ of closure %p (%s) on [PE %d]: ", + node, info_type(node), node_loc); + + /* + NB: In a parallel setup a BQ of an RBH must end with an RBH_Save closure; + */ + for (bqe = ((StgBlockingQueue*)node)->blocking_queue, end = (bqe==END_BQ_QUEUE); + !end; // iterate until bqe points to a CONSTR + end = (get_itbl(bqe)->type == CONSTR) || (bqe->link==END_BQ_QUEUE), bqe = end ? END_BQ_QUEUE : bqe->link) { + ASSERT(bqe != END_BQ_QUEUE); // sanity check + ASSERT(bqe != (StgBlockingQueueElement *)NULL); // sanity check + /* types of closures that may appear in a blocking queue */ + ASSERT(get_itbl(bqe)->type == TSO || + get_itbl(bqe)->type == CONSTR); + /* only BQs of an RBH end with an RBH_Save closure */ + ASSERT(get_itbl(bqe)->type != CONSTR || get_itbl(node)->type == RBH); + + tso_loc = where_is((StgClosure *)bqe); + switch (get_itbl(bqe)->type) { + case TSO: + fprintf(stderr," TSO %d (%p) on [PE %d],", + ((StgTSO *)bqe)->id, (StgTSO *)bqe, tso_loc); + break; + case CONSTR: + fprintf(stderr," %s (IP %p),", + (get_itbl(bqe) == &stg_RBH_Save_0_info ? "RBH_Save_0" : + get_itbl(bqe) == &stg_RBH_Save_1_info ? "RBH_Save_1" : + get_itbl(bqe) == &stg_RBH_Save_2_info ? "RBH_Save_2" : + "RBH_Save_?"), get_itbl(bqe)); + break; + default: + barf("Unexpected closure type %s in blocking queue of %p (%s)", + info_type((StgClosure *)bqe), node, info_type(node)); + break; + } + } /* for */ + fputc('\n', stderr); +} +#else +/* + Nice and easy: only TSOs on the blocking queue +*/ +void +print_bq (StgClosure *node) +{ + StgTSO *tso; + + ASSERT(node!=(StgClosure*)NULL); // sanity check + for (tso = ((StgBlockingQueue*)node)->blocking_queue; + tso != END_TSO_QUEUE; + tso=tso->link) { + ASSERT(tso!=NULL && tso!=END_TSO_QUEUE); // sanity check + ASSERT(get_itbl(tso)->type == TSO); // guess what, sanity check + fprintf(stderr," TSO %d (%p),", tso->id, tso); } + fputc('\n', stderr); } +# endif -/* ----------------------------------------------------------------------------- - Interrupt execution - - usually called inside a signal handler so it mustn't do anything fancy. - -------------------------------------------------------------------------- */ +#if defined(PAR) +static nat +run_queue_len(void) +{ + nat i; + StgTSO *tso; + + for (i=0, tso=run_queue_hd; + tso != END_TSO_QUEUE; + i++, tso=tso->link) + /* nothing */ -void interruptStgRts(void) + return i; +} +#endif + +static void +sched_belch(char *s, ...) { - interrupted = 1; - context_switch = 1; + va_list ap; + va_start(ap,s); +#ifdef SMP + fprintf(stderr, "scheduler (task %ld): ", pthread_self()); +#elif defined(PAR) + fprintf(stderr, "== "); +#else + fprintf(stderr, "scheduler: "); +#endif + vfprintf(stderr, s, ap); + fprintf(stderr, "\n"); } +#endif /* DEBUG */ + + +//@node Index, , Debugging Routines, Main scheduling code +//@subsection Index + +//@index +//* MainRegTable:: @cindex\s-+MainRegTable +//* StgMainThread:: @cindex\s-+StgMainThread +//* awaken_blocked_queue:: @cindex\s-+awaken_blocked_queue +//* blocked_queue_hd:: @cindex\s-+blocked_queue_hd +//* blocked_queue_tl:: @cindex\s-+blocked_queue_tl +//* context_switch:: @cindex\s-+context_switch +//* createThread:: @cindex\s-+createThread +//* free_capabilities:: @cindex\s-+free_capabilities +//* gc_pending_cond:: @cindex\s-+gc_pending_cond +//* initScheduler:: @cindex\s-+initScheduler +//* interrupted:: @cindex\s-+interrupted +//* n_free_capabilities:: @cindex\s-+n_free_capabilities +//* next_thread_id:: @cindex\s-+next_thread_id +//* print_bq:: @cindex\s-+print_bq +//* run_queue_hd:: @cindex\s-+run_queue_hd +//* run_queue_tl:: @cindex\s-+run_queue_tl +//* sched_mutex:: @cindex\s-+sched_mutex +//* schedule:: @cindex\s-+schedule +//* take_off_run_queue:: @cindex\s-+take_off_run_queue +//* task_ids:: @cindex\s-+task_ids +//* term_mutex:: @cindex\s-+term_mutex +//* thread_ready_cond:: @cindex\s-+thread_ready_cond +//@end index