#include "PosixSource.h"
#include "Rts.h"
-#include "RtsUtils.h"
-#include "RtsFlags.h"
-#include "STM.h"
-#include "OSThreads.h"
+
#include "Capability.h"
#include "Schedule.h"
#include "Sparks.h"
#include "Trace.h"
+#include "sm/GC.h" // for gcWorkerThread()
+#include "STM.h"
+#include "RtsUtils.h"
// one global capability, this is the Capability for non-threaded
// builds, and for +RTS -N1
Capability MainCapability;
-nat n_capabilities;
+nat n_capabilities = 0;
Capability *capabilities = NULL;
// Holds the Capability which last became free. This is used so that
// an in-call has a chance of quickly finding a free Capability.
// Maintaining a global free list of Capabilities would require global
// locking, so we don't do that.
-Capability *last_free_capability;
+Capability *last_free_capability = NULL;
/* GC indicator, in scope for the scheduler, init'ed to false */
volatile StgWord waiting_for_gc = 0;
+/* Let foreign code get the current Capability -- assuming there is one!
+ * This is useful for unsafe foreign calls because they are called with
+ * the current Capability held, but they are not passed it. For example,
+ * see see the integer-gmp package which calls allocateLocal() in its
+ * stgAllocForGMP() function (which gets called by gmp functions).
+ * */
+Capability * rts_unsafeGetMyCapability (void)
+{
+#if defined(THREADED_RTS)
+ return myTask()->cap;
+#else
+ return &MainCapability;
+#endif
+}
+
#if defined(THREADED_RTS)
STATIC_INLINE rtsBool
globalWorkToDo (void)
{
- return blackholes_need_checking
- || sched_state >= SCHED_INTERRUPTING
- ;
+ return sched_state >= SCHED_INTERRUPTING
+ || recent_activity == ACTIVITY_INACTIVE; // need to check for deadlock
}
#endif
StgClosure *
findSpark (Capability *cap)
{
- /* use the normal Sparks.h interface (internally modified to enable
- concurrent stealing)
- and immediately turn the spark into a thread when successful
- */
Capability *robbed;
StgClosurePtr spark;
rtsBool retry;
nat i = 0;
- // first try to get a spark from our own pool.
- // We should be using reclaimSpark(), because it works without
- // needing any atomic instructions:
- // spark = reclaimSpark(cap->sparks);
- // However, measurements show that this makes at least one benchmark
- // slower (prsa) and doesn't affect the others.
- spark = tryStealSpark(cap);
- if (spark != NULL) {
- cap->sparks_converted++;
- return spark;
+ if (!emptyRunQueue(cap) || cap->returning_tasks_hd != NULL) {
+ // If there are other threads, don't try to run any new
+ // sparks: sparks might be speculative, we don't want to take
+ // resources away from the main computation.
+ return 0;
}
- if (n_capabilities == 1) { return NULL; } // makes no sense...
-
- debugTrace(DEBUG_sched,
- "cap %d: Trying to steal work from other capabilities",
- cap->no);
-
do {
retry = rtsFalse;
+ // first try to get a spark from our own pool.
+ // We should be using reclaimSpark(), because it works without
+ // needing any atomic instructions:
+ // spark = reclaimSpark(cap->sparks);
+ // However, measurements show that this makes at least one benchmark
+ // slower (prsa) and doesn't affect the others.
+ spark = tryStealSpark(cap);
+ if (spark != NULL) {
+ cap->sparks_converted++;
+
+ // Post event for running a spark from capability's own pool.
+ traceEventRunSpark(cap, cap->r.rCurrentTSO);
+
+ return spark;
+ }
+ if (!emptySparkPoolCap(cap)) {
+ retry = rtsTrue;
+ }
+
+ if (n_capabilities == 1) { return NULL; } // makes no sense...
+
+ debugTrace(DEBUG_sched,
+ "cap %d: Trying to steal work from other capabilities",
+ cap->no);
+
/* visit cap.s 0..n-1 in sequence until a theft succeeds. We could
start at a random place instead of 0 as well. */
for ( i=0 ; i < n_capabilities ; i++ ) {
}
if (spark != NULL) {
- debugTrace(DEBUG_sched,
- "cap %d: Stole a spark from capability %d",
- cap->no, robbed->no);
cap->sparks_converted++;
+
+ traceEventStealSpark(cap, cap->r.rCurrentTSO, robbed->no);
+
return spark;
}
// otherwise: no success, try next one
newReturningTask (Capability *cap, Task *task)
{
ASSERT_LOCK_HELD(&cap->lock);
- ASSERT(task->return_link == NULL);
+ ASSERT(task->next == NULL);
if (cap->returning_tasks_hd) {
- ASSERT(cap->returning_tasks_tl->return_link == NULL);
- cap->returning_tasks_tl->return_link = task;
+ ASSERT(cap->returning_tasks_tl->next == NULL);
+ cap->returning_tasks_tl->next = task;
} else {
cap->returning_tasks_hd = task;
}
Task *task;
task = cap->returning_tasks_hd;
ASSERT(task);
- cap->returning_tasks_hd = task->return_link;
+ cap->returning_tasks_hd = task->next;
if (!cap->returning_tasks_hd) {
cap->returning_tasks_tl = NULL;
}
- task->return_link = NULL;
+ task->next = NULL;
return task;
}
#endif
initMutex(&cap->lock);
cap->running_task = NULL; // indicates cap is free
cap->spare_workers = NULL;
- cap->suspended_ccalling_tasks = NULL;
+ cap->n_spare_workers = 0;
+ cap->suspended_ccalls = NULL;
cap->returning_tasks_hd = NULL;
cap->returning_tasks_tl = NULL;
- cap->wakeup_queue_hd = END_TSO_QUEUE;
- cap->wakeup_queue_tl = END_TSO_QUEUE;
+ cap->inbox = (Message*)END_TSO_QUEUE;
cap->sparks_created = 0;
+ cap->sparks_dud = 0;
cap->sparks_converted = 0;
- cap->sparks_pruned = 0;
+ cap->sparks_gcd = 0;
+ cap->sparks_fizzled = 0;
#endif
cap->f.stgEagerBlackholeInfo = (W_)&__stg_EAGER_BLACKHOLE_info;
- cap->f.stgGCEnter1 = (F_)__stg_gc_enter_1;
- cap->f.stgGCFun = (F_)__stg_gc_fun;
+ cap->f.stgGCEnter1 = (StgFunPtr)__stg_gc_enter_1;
+ cap->f.stgGCFun = (StgFunPtr)__stg_gc_fun;
cap->mut_lists = stgMallocBytes(sizeof(bdescr *) *
RtsFlags.GcFlags.generations,
"initCapability");
+ cap->saved_mut_lists = stgMallocBytes(sizeof(bdescr *) *
+ RtsFlags.GcFlags.generations,
+ "initCapability");
for (g = 0; g < RtsFlags.GcFlags.generations; g++) {
cap->mut_lists[g] = NULL;
cap->free_trec_headers = NO_TREC;
cap->transaction_tokens = 0;
cap->context_switch = 0;
+ cap->pinned_object_block = NULL;
}
/* ---------------------------------------------------------------------------
void setContextSwitches(void)
{
- nat i;
- for (i=0; i < n_capabilities; i++) {
- capabilities[i].context_switch = 1;
- }
+ nat i;
+ for (i=0; i < n_capabilities; i++) {
+ contextSwitchCapability(&capabilities[i]);
+ }
}
/* ----------------------------------------------------------------------------
{
ASSERT_LOCK_HELD(&cap->lock);
ASSERT(task->cap == cap);
- trace(TRACE_sched | DEBUG_sched,
- "passing capability %d to %s %p",
- cap->no, task->tso ? "bound task" : "worker",
- (void *)task->id);
+ debugTrace(DEBUG_sched, "passing capability %d to %s %p",
+ cap->no, task->incall->tso ? "bound task" : "worker",
+ (void *)task->id);
ACQUIRE_LOCK(&task->lock);
task->wakeup = rtsTrue;
// the wakeup flag is needed because signalCondition() doesn't
return;
}
- /* if waiting_for_gc was the reason to release the cap: thread
- comes from yieldCap->releaseAndQueueWorker. Unconditionally set
- cap. free and return (see default after the if-protected other
- special cases). Thread will wait on cond.var and re-acquire the
- same cap after GC (GC-triggering cap. calls releaseCap and
- enters the spare_workers case)
- */
- if (waiting_for_gc) {
+ if (waiting_for_gc == PENDING_GC_SEQ) {
last_free_capability = cap; // needed?
- trace(TRACE_sched | DEBUG_sched,
- "GC pending, set capability %d free", cap->no);
+ debugTrace(DEBUG_sched, "GC pending, set capability %d free", cap->no);
return;
}
// give this Capability to the appropriate Task.
if (!emptyRunQueue(cap) && cap->run_queue_hd->bound) {
// Make sure we're not about to try to wake ourselves up
- ASSERT(task != cap->run_queue_hd->bound);
- task = cap->run_queue_hd->bound;
+ // ASSERT(task != cap->run_queue_hd->bound);
+ // assertion is false: in schedule() we force a yield after
+ // ThreadBlocked, but the thread may be back on the run queue
+ // by now.
+ task = cap->run_queue_hd->bound->task;
giveCapabilityToTask(cap,task);
return;
}
if (sched_state < SCHED_SHUTTING_DOWN || !emptyRunQueue(cap)) {
debugTrace(DEBUG_sched,
"starting new worker on capability %d", cap->no);
- startWorkerTask(cap, workerStart);
+ startWorkerTask(cap);
return;
}
}
// If we have an unbound thread on the run queue, or if there's
// anything else to do, give the Capability to a worker thread.
if (always_wakeup ||
- !emptyRunQueue(cap) || !emptyWakeupQueue(cap) ||
+ !emptyRunQueue(cap) || !emptyInbox(cap) ||
!emptySparkPoolCap(cap) || globalWorkToDo()) {
if (cap->spare_workers) {
giveCapabilityToTask(cap,cap->spare_workers);
}
last_free_capability = cap;
- trace(TRACE_sched | DEBUG_sched, "freeing capability %d", cap->no);
+ debugTrace(DEBUG_sched, "freeing capability %d", cap->no);
}
void
task = cap->running_task;
+ // If the Task is stopped, we shouldn't be yielding, we should
+ // be just exiting.
+ ASSERT(!task->stopped);
+
// If the current task is a worker, save it on the spare_workers
// list of this Capability. A worker can mark itself as stopped,
// in which case it is not replaced on the spare_worker queue.
// This happens when the system is shutting down (see
// Schedule.c:workerStart()).
- // Also, be careful to check that this task hasn't just exited
- // Haskell to do a foreign call (task->suspended_tso).
- if (!isBoundTask(task) && !task->stopped && !task->suspended_tso) {
- task->next = cap->spare_workers;
- cap->spare_workers = task;
+ if (!isBoundTask(task))
+ {
+ if (cap->n_spare_workers < MAX_SPARE_WORKERS)
+ {
+ task->next = cap->spare_workers;
+ cap->spare_workers = task;
+ cap->n_spare_workers++;
+ }
+ else
+ {
+ debugTrace(DEBUG_sched, "%d spare workers already, exiting",
+ cap->n_spare_workers);
+ releaseCapability_(cap,rtsFalse);
+ // hold the lock until after workerTaskStop; c.f. scheduleWorker()
+ workerTaskStop(task);
+ RELEASE_LOCK(&cap->lock);
+ shutdownThread();
+ }
}
// Bound tasks just float around attached to their TSOs.
if (cap == NULL) {
// Try last_free_capability first
cap = last_free_capability;
- if (!cap->running_task) {
+ if (cap->running_task) {
nat i;
// otherwise, search for a free capability
+ cap = NULL;
for (i = 0; i < n_capabilities; i++) {
- cap = &capabilities[i];
- if (!cap->running_task) {
+ if (!capabilities[i].running_task) {
+ cap = &capabilities[i];
break;
}
}
- // Can't find a free one, use last_free_capability.
- cap = last_free_capability;
+ if (cap == NULL) {
+ // Can't find a free one, use last_free_capability.
+ cap = last_free_capability;
+ }
}
// record the Capability as the one this Task is now assocated with.
ASSERT_FULL_CAPABILITY_INVARIANTS(cap,task);
- trace(TRACE_sched | DEBUG_sched, "resuming capability %d", cap->no);
+ debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
*pCap = cap;
#endif
{
Capability *cap = *pCap;
+ if (waiting_for_gc == PENDING_GC_PAR) {
+ traceEventGcStart(cap);
+ gcWorkerThread(cap);
+ traceEventGcEnd(cap);
+ return;
+ }
+
debugTrace(DEBUG_sched, "giving up capability %d", cap->no);
// We must now release the capability and wait to be woken up
continue;
}
- if (task->tso == NULL) {
+ if (task->incall->tso == NULL) {
ASSERT(cap->spare_workers != NULL);
// if we're not at the front of the queue, release it
// again. This is unlikely to happen.
}
cap->spare_workers = task->next;
task->next = NULL;
- }
+ cap->n_spare_workers--;
+ }
cap->running_task = task;
RELEASE_LOCK(&cap->lock);
break;
}
- trace(TRACE_sched | DEBUG_sched, "resuming capability %d", cap->no);
+ debugTrace(DEBUG_sched, "resuming capability %d", cap->no);
ASSERT(cap->running_task == task);
*pCap = cap;
}
/* ----------------------------------------------------------------------------
- * Wake up a thread on a Capability.
+ * prodCapability
*
- * This is used when the current Task is running on a Capability and
- * wishes to wake up a thread on a different Capability.
+ * If a Capability is currently idle, wake up a Task on it. Used to
+ * get every Capability into the GC.
* ------------------------------------------------------------------------- */
void
-wakeupThreadOnCapability (Capability *my_cap,
- Capability *other_cap,
- StgTSO *tso)
-{
- ACQUIRE_LOCK(&other_cap->lock);
-
- // ASSUMES: cap->lock is held (asserted in wakeupThreadOnCapability)
- if (tso->bound) {
- ASSERT(tso->bound->cap == tso->cap);
- tso->bound->cap = other_cap;
- }
- tso->cap = other_cap;
-
- ASSERT(tso->bound ? tso->bound->cap == other_cap : 1);
-
- if (other_cap->running_task == NULL) {
- // nobody is running this Capability, we can add our thread
- // directly onto the run queue and start up a Task to run it.
-
- other_cap->running_task = myTask();
- // precond for releaseCapability_() and appendToRunQueue()
-
- appendToRunQueue(other_cap,tso);
-
- trace(TRACE_sched, "resuming capability %d", other_cap->no);
- releaseCapability_(other_cap,rtsFalse);
- } else {
- appendToWakeupQueue(my_cap,other_cap,tso);
- other_cap->context_switch = 1;
- // someone is running on this Capability, so it cannot be
- // freed without first checking the wakeup queue (see
- // releaseCapability_).
- }
-
- RELEASE_LOCK(&other_cap->lock);
-}
-
-/* ----------------------------------------------------------------------------
- * prodCapabilities
- *
- * Used to indicate that the interrupted flag is now set, or some
- * other global condition that might require waking up a Task on each
- * Capability.
- * ------------------------------------------------------------------------- */
-
-static void
-prodCapabilities(rtsBool all)
+prodCapability (Capability *cap, Task *task)
{
- nat i;
- Capability *cap;
- Task *task;
-
- for (i=0; i < n_capabilities; i++) {
- cap = &capabilities[i];
- ACQUIRE_LOCK(&cap->lock);
- if (!cap->running_task) {
- if (cap->spare_workers) {
- trace(TRACE_sched, "resuming capability %d", cap->no);
- task = cap->spare_workers;
- ASSERT(!task->stopped);
- giveCapabilityToTask(cap,task);
- if (!all) {
- RELEASE_LOCK(&cap->lock);
- return;
- }
- }
- }
- RELEASE_LOCK(&cap->lock);
+ ACQUIRE_LOCK(&cap->lock);
+ if (!cap->running_task) {
+ cap->running_task = task;
+ releaseCapability_(cap,rtsTrue);
}
- return;
-}
-
-void
-prodAllCapabilities (void)
-{
- prodCapabilities(rtsTrue);
-}
-
-/* ----------------------------------------------------------------------------
- * prodOneCapability
- *
- * Like prodAllCapabilities, but we only require a single Task to wake
- * up in order to service some global event, such as checking for
- * deadlock after some idle time has passed.
- * ------------------------------------------------------------------------- */
-
-void
-prodOneCapability (void)
-{
- prodCapabilities(rtsFalse);
+ RELEASE_LOCK(&cap->lock);
}
/* ----------------------------------------------------------------------------
if (!osThreadIsAlive(t->id)) {
debugTrace(DEBUG_sched,
"worker thread %p has died unexpectedly", (void *)t->id);
- if (!prev) {
- cap->spare_workers = t->next;
- } else {
- prev->next = t->next;
- }
- prev = t;
+ cap->n_spare_workers--;
+ if (!prev) {
+ cap->spare_workers = t->next;
+ } else {
+ prev->next = t->next;
+ }
+ prev = t;
}
}
}
// that will try to return to code that has been unloaded.
// We can be a bit more relaxed when this is a standalone
// program that is about to terminate, and let safe=false.
- if (cap->suspended_ccalling_tasks && safe) {
+ if (cap->suspended_ccalls && safe) {
debugTrace(DEBUG_sched,
"thread(s) are involved in foreign calls, yielding");
cap->running_task = NULL;
RELEASE_LOCK(&cap->lock);
+ // The IO manager thread might have been slow to start up,
+ // so the first attempt to kill it might not have
+ // succeeded. Just in case, try again - the kill message
+ // will only be sent once.
+ //
+ // To reproduce this deadlock: run ffi002(threaded1)
+ // repeatedly on a loaded machine.
+ ioManagerDie();
yieldThread();
continue;
}
-
- debugTrace(DEBUG_sched, "capability %d is stopped.", cap->no);
+
+ traceEventShutdown(cap);
RELEASE_LOCK(&cap->lock);
break;
}
freeCapability (Capability *cap)
{
stgFree(cap->mut_lists);
-#if defined(THREADED_RTS) || defined(PARALLEL_HASKELL)
+ stgFree(cap->saved_mut_lists);
+#if defined(THREADED_RTS)
freeSparkPool(cap->sparks);
#endif
}
------------------------------------------------------------------------ */
void
-markSomeCapabilities (evac_fn evac, void *user, nat i0, nat delta,
- rtsBool prune_sparks USED_IF_THREADS)
+markCapability (evac_fn evac, void *user, Capability *cap,
+ rtsBool no_mark_sparks USED_IF_THREADS)
{
- nat i;
- Capability *cap;
- Task *task;
+ InCall *incall;
// Each GC thread is responsible for following roots from the
// Capability of the same number. There will usually be the same
// or fewer Capabilities as GC threads, but just in case there
// are more, we mark every Capability whose number is the GC
// thread's index plus a multiple of the number of GC threads.
- for (i = i0; i < n_capabilities; i += delta) {
- cap = &capabilities[i];
- evac(user, (StgClosure **)(void *)&cap->run_queue_hd);
- evac(user, (StgClosure **)(void *)&cap->run_queue_tl);
+ evac(user, (StgClosure **)(void *)&cap->run_queue_hd);
+ evac(user, (StgClosure **)(void *)&cap->run_queue_tl);
#if defined(THREADED_RTS)
- evac(user, (StgClosure **)(void *)&cap->wakeup_queue_hd);
- evac(user, (StgClosure **)(void *)&cap->wakeup_queue_tl);
+ evac(user, (StgClosure **)(void *)&cap->inbox);
#endif
- for (task = cap->suspended_ccalling_tasks; task != NULL;
- task=task->next) {
- debugTrace(DEBUG_sched,
- "evac'ing suspended TSO %lu", (unsigned long)task->suspended_tso->id);
- evac(user, (StgClosure **)(void *)&task->suspended_tso);
- }
+ for (incall = cap->suspended_ccalls; incall != NULL;
+ incall=incall->next) {
+ evac(user, (StgClosure **)(void *)&incall->suspended_tso);
+ }
#if defined(THREADED_RTS)
- if (prune_sparks) {
- pruneSparkQueue (evac, user, cap);
- } else {
- traverseSparkQueue (evac, user, cap);
- }
-#endif
+ if (!no_mark_sparks) {
+ traverseSparkQueue (evac, user, cap);
}
+#endif
-#if !defined(THREADED_RTS)
- evac(user, (StgClosure **)(void *)&blocked_queue_hd);
- evac(user, (StgClosure **)(void *)&blocked_queue_tl);
- evac(user, (StgClosure **)(void *)&sleeping_queue);
-#endif
+ // Free STM structures for this Capability
+ stmPreGCHook(cap);
}
void
markCapabilities (evac_fn evac, void *user)
{
- markSomeCapabilities(evac, user, 0, 1, rtsFalse);
+ nat n;
+ for (n = 0; n < n_capabilities; n++) {
+ markCapability(evac, user, &capabilities[n], rtsFalse);
+ }
}