+ debugTrace(DEBUG_gc, "%d GC threads still running", r);
+
+ while (gc_running_threads != 0) {
+ // usleep(1);
+ if (any_work()) {
+ inc_running();
+ goto loop;
+ }
+ // any_work() does not remove the work from the queue, it
+ // just checks for the presence of work. If we find any,
+ // then we increment gc_running_threads and go back to
+ // scavenge_loop() to perform any pending work.
+ }
+
+ traceEvent(&capabilities[gct->thread_index], EVENT_GC_DONE);
+}
+
+#if defined(THREADED_RTS)
+
+void
+gcWorkerThread (Capability *cap)
+{
+ gc_thread *saved_gct;
+
+ // necessary if we stole a callee-saves register for gct:
+ saved_gct = gct;
+
+ cap->in_gc = rtsTrue;
+
+ gct = gc_threads[cap->no];
+ gct->id = osThreadId();
+
+ // Wait until we're told to wake up
+ RELEASE_SPIN_LOCK(&gct->mut_spin);
+ gct->wakeup = GC_THREAD_STANDING_BY;
+ debugTrace(DEBUG_gc, "GC thread %d standing by...", gct->thread_index);
+ ACQUIRE_SPIN_LOCK(&gct->gc_spin);
+
+#ifdef USE_PAPI
+ // start performance counters in this thread...
+ if (gct->papi_events == -1) {
+ papi_init_eventset(&gct->papi_events);
+ }
+ papi_thread_start_gc1_count(gct->papi_events);
+#endif
+
+ // Every thread evacuates some roots.
+ gct->evac_step = 0;
+ markSomeCapabilities(mark_root, gct, gct->thread_index, n_gc_threads,
+ rtsTrue/*prune sparks*/);
+ scavenge_capability_mut_lists(&capabilities[gct->thread_index]);
+
+ scavenge_until_all_done();
+
+#ifdef USE_PAPI
+ // count events in this thread towards the GC totals
+ papi_thread_stop_gc1_count(gct->papi_events);
+#endif
+
+ // Wait until we're told to continue
+ RELEASE_SPIN_LOCK(&gct->gc_spin);
+ gct->wakeup = GC_THREAD_WAITING_TO_CONTINUE;
+ debugTrace(DEBUG_gc, "GC thread %d waiting to continue...",
+ gct->thread_index);
+ ACQUIRE_SPIN_LOCK(&gct->mut_spin);
+ debugTrace(DEBUG_gc, "GC thread %d on my way...", gct->thread_index);
+
+ SET_GCT(saved_gct);
+}
+
+#endif
+
+#if defined(THREADED_RTS)
+
+void
+waitForGcThreads (Capability *cap USED_IF_THREADS)
+{
+ nat n_threads = RtsFlags.ParFlags.nNodes;
+ nat me = cap->no;
+ nat i, j;
+ rtsBool retry = rtsTrue;
+
+ while(retry) {
+ for (i=0; i < n_threads; i++) {
+ if (i == me) continue;
+ if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
+ prodCapability(&capabilities[i], cap->running_task);
+ }
+ }
+ for (j=0; j < 10; j++) {
+ retry = rtsFalse;
+ for (i=0; i < n_threads; i++) {
+ if (i == me) continue;
+ write_barrier();
+ setContextSwitches();
+ if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) {
+ retry = rtsTrue;
+ }
+ }
+ if (!retry) break;
+ yieldThread();
+ }
+ }
+}
+
+#endif // THREADED_RTS
+
+static void
+start_gc_threads (void)
+{
+#if defined(THREADED_RTS)
+ gc_running_threads = 0;
+#endif
+}
+
+static void
+wakeup_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
+{
+#if defined(THREADED_RTS)
+ nat i;
+ for (i=0; i < n_threads; i++) {
+ if (i == me) continue;
+ inc_running();
+ debugTrace(DEBUG_gc, "waking up gc thread %d", i);
+ if (gc_threads[i]->wakeup != GC_THREAD_STANDING_BY) barf("wakeup_gc_threads");
+
+ gc_threads[i]->wakeup = GC_THREAD_RUNNING;
+ ACQUIRE_SPIN_LOCK(&gc_threads[i]->mut_spin);
+ RELEASE_SPIN_LOCK(&gc_threads[i]->gc_spin);
+ }
+#endif
+}
+
+// After GC is complete, we must wait for all GC threads to enter the
+// standby state, otherwise they may still be executing inside
+// any_work(), and may even remain awake until the next GC starts.
+static void
+shutdown_gc_threads (nat n_threads USED_IF_THREADS, nat me USED_IF_THREADS)
+{
+#if defined(THREADED_RTS)
+ nat i;
+ for (i=0; i < n_threads; i++) {
+ if (i == me) continue;
+ while (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE) { write_barrier(); }
+ }
+#endif
+}
+
+#if defined(THREADED_RTS)
+void
+releaseGCThreads (Capability *cap USED_IF_THREADS)
+{
+ nat n_threads = RtsFlags.ParFlags.nNodes;
+ nat me = cap->no;
+ nat i;
+ for (i=0; i < n_threads; i++) {
+ if (i == me) continue;
+ if (gc_threads[i]->wakeup != GC_THREAD_WAITING_TO_CONTINUE)
+ barf("releaseGCThreads");
+
+ gc_threads[i]->wakeup = GC_THREAD_INACTIVE;
+ ACQUIRE_SPIN_LOCK(&gc_threads[i]->gc_spin);
+ RELEASE_SPIN_LOCK(&gc_threads[i]->mut_spin);
+ }
+}
+#endif
+
+/* ----------------------------------------------------------------------------
+ Initialise a generation that is to be collected
+ ------------------------------------------------------------------------- */
+
+static void
+init_collected_gen (nat g, nat n_threads)
+{
+ nat s, t, i;
+ step_workspace *ws;
+ step *stp;
+ bdescr *bd;
+
+ // Throw away the current mutable list. Invariant: the mutable
+ // list always has at least one block; this means we can avoid a
+ // check for NULL in recordMutable().
+ if (g != 0) {
+ freeChain(generations[g].mut_list);
+ generations[g].mut_list = allocBlock();
+ for (i = 0; i < n_capabilities; i++) {
+ freeChain(capabilities[i].mut_lists[g]);
+ capabilities[i].mut_lists[g] = allocBlock();
+ }
+ }
+
+ for (s = 0; s < generations[g].n_steps; s++) {
+
+ stp = &generations[g].steps[s];
+ ASSERT(stp->gen_no == g);
+
+ // we'll construct a new list of threads in this step
+ // during GC, throw away the current list.
+ stp->old_threads = stp->threads;
+ stp->threads = END_TSO_QUEUE;
+
+ // generation 0, step 0 doesn't need to-space
+ if (g == 0 && s == 0 && RtsFlags.GcFlags.generations > 1) {
+ continue;
+ }
+
+ // deprecate the existing blocks
+ stp->old_blocks = stp->blocks;
+ stp->n_old_blocks = stp->n_blocks;
+ stp->blocks = NULL;
+ stp->n_blocks = 0;
+ stp->n_words = 0;
+ stp->live_estimate = 0;
+
+ // initialise the large object queues.
+ stp->scavenged_large_objects = NULL;
+ stp->n_scavenged_large_blocks = 0;
+
+ // mark the small objects as from-space
+ for (bd = stp->old_blocks; bd; bd = bd->link) {
+ bd->flags &= ~BF_EVACUATED;
+ }
+
+ // mark the large objects as from-space
+ for (bd = stp->large_objects; bd; bd = bd->link) {
+ bd->flags &= ~BF_EVACUATED;
+ }
+
+ // for a compacted step, we need to allocate the bitmap
+ if (stp->mark) {
+ nat bitmap_size; // in bytes
+ bdescr *bitmap_bdescr;
+ StgWord *bitmap;
+
+ bitmap_size = stp->n_old_blocks * BLOCK_SIZE / (sizeof(W_)*BITS_PER_BYTE);
+
+ if (bitmap_size > 0) {
+ bitmap_bdescr = allocGroup((lnat)BLOCK_ROUND_UP(bitmap_size)
+ / BLOCK_SIZE);
+ stp->bitmap = bitmap_bdescr;
+ bitmap = bitmap_bdescr->start;
+
+ debugTrace(DEBUG_gc, "bitmap_size: %d, bitmap: %p",
+ bitmap_size, bitmap);
+
+ // don't forget to fill it with zeros!
+ memset(bitmap, 0, bitmap_size);
+
+ // For each block in this step, point to its bitmap from the
+ // block descriptor.
+ for (bd=stp->old_blocks; bd != NULL; bd = bd->link) {
+ bd->u.bitmap = bitmap;
+ bitmap += BLOCK_SIZE_W / (sizeof(W_)*BITS_PER_BYTE);
+
+ // Also at this point we set the BF_MARKED flag
+ // for this block. The invariant is that
+ // BF_MARKED is always unset, except during GC
+ // when it is set on those blocks which will be
+ // compacted.
+ if (!(bd->flags & BF_FRAGMENTED)) {
+ bd->flags |= BF_MARKED;
+ }
+ }
+ }
+ }
+ }
+
+ // For each GC thread, for each step, allocate a "todo" block to
+ // store evacuated objects to be scavenged, and a block to store
+ // evacuated objects that do not need to be scavenged.
+ for (t = 0; t < n_threads; t++) {
+ for (s = 0; s < generations[g].n_steps; s++) {
+
+ // we don't copy objects into g0s0, unless -G0
+ if (g==0 && s==0 && RtsFlags.GcFlags.generations > 1) continue;
+
+ ws = &gc_threads[t]->steps[g * RtsFlags.GcFlags.steps + s];
+
+ ws->todo_large_objects = NULL;
+
+ ws->part_list = NULL;
+ ws->n_part_blocks = 0;
+
+ // allocate the first to-space block; extra blocks will be
+ // chained on as necessary.
+ ws->todo_bd = NULL;
+ ASSERT(looksEmptyWSDeque(ws->todo_q));
+ alloc_todo_block(ws,0);
+
+ ws->todo_overflow = NULL;
+ ws->n_todo_overflow = 0;
+
+ ws->scavd_list = NULL;
+ ws->n_scavd_blocks = 0;
+ }