rts: add max_n_todo_overflow internal counter

I've never observed this counter taking a non-zero value, however I do
think it's existence is justified by the comment in grab_local_todo_block.

I've not added it to RTSStats in GHC.Stats, as it doesn't seem worth the
api churn.

3 files changed, 18 insertions, 2 deletions

diff --git a/rts/sm/GC.c b/rts/sm/GC.c
index b78f993260..5e986f2296 100644
--- a/rts/sm/GC.c
+++ b/rts/sm/GC.c
@@ -270,7 +270,7 @@ GarbageCollect (uint32_t collect_gen,
   generation *gen;
   StgWord live_blocks, live_words, par_max_copied, par_balanced_copied,
       gc_spin_spin, gc_spin_yield, mut_spin_spin, mut_spin_yield,
-      any_work, scav_find_work;
+      any_work, scav_find_work, max_n_todo_overflow;
 #if defined(THREADED_RTS)
   gc_thread *saved_gct;
 #endif
@@ -594,6 +594,7 @@ GarbageCollect (uint32_t collect_gen,
   mut_spin_yield = 0;
   any_work = 0;
   scav_find_work = 0;
+  max_n_todo_overflow = 0;
   {
       uint32_t i;
       uint64_t par_balanced_copied_acc = 0;
@@ -625,6 +626,7 @@ GarbageCollect (uint32_t collect_gen,
 
               any_work += RELAXED_LOAD(&thread->any_work);
               scav_find_work += RELAXED_LOAD(&thread->scav_find_work);
+              max_n_todo_overflow = stg_max(RELAXED_LOAD(&thread->max_n_todo_overflow), max_n_todo_overflow);
 
               par_max_copied = stg_max(RELAXED_LOAD(&thread->copied), par_max_copied);
               par_balanced_copied_acc +=
@@ -1043,7 +1045,7 @@ GarbageCollect (uint32_t collect_gen,
              N, n_gc_threads, gc_threads,
              par_max_copied, par_balanced_copied,
              gc_spin_spin, gc_spin_yield, mut_spin_spin, mut_spin_yield,
-             any_work, scav_find_work);
+             any_work, scav_find_work, max_n_todo_overflow);
 
 #if defined(RTS_USER_SIGNALS)
   if (RtsFlags.MiscFlags.install_signal_handlers) {
@@ -1803,6 +1805,7 @@ init_gc_thread (gc_thread *t)
     t->scanned = 0;
     t->any_work = 0;
     t->scav_find_work = 0;
+    t->max_n_todo_overflow = 0;
 }
 
 /* -----------------------------------------------------------------------------
diff --git a/rts/sm/GCThread.h b/rts/sm/GCThread.h
index 90d15c69c5..31719ca020 100644
--- a/rts/sm/GCThread.h
+++ b/rts/sm/GCThread.h
@@ -183,6 +183,7 @@ typedef struct gc_thread_ {
     W_ scanned;
     W_ any_work;
     W_ scav_find_work;
+    W_ max_n_todo_overflow;
 
     Time gc_start_cpu;             // thread CPU time
     Time gc_end_cpu;               // thread CPU time
diff --git a/rts/sm/GCUtils.c b/rts/sm/GCUtils.c
index 89a92bc837..52ea27f263 100644
--- a/rts/sm/GCUtils.c
+++ b/rts/sm/GCUtils.c
@@ -183,6 +183,18 @@ push_todo_block(bdescr *bd, gen_workspace *ws)
         bd->link = ws->todo_overflow;
         ws->todo_overflow = bd;
         ws->n_todo_overflow++;
+
+        // In theory, if a gc thread pushes more blocks to it's todo_q than it
+        // pops, the todo_overflow list will continue to grow. Other gc threads
+        // can't steal from the todo_overflwo list, so they may be idle as the
+        // first gc thread works diligently on it's todo_overflow list. In
+        // practice this has not been observed to occur.
+        //
+        // The max_n_todo_overflow counter will allow us to observe large
+        // todo_overflow lists if they ever arise. As of now I've not observed
+        // any nonzero max_n_todo_overflow samples.
+        gct->max_n_todo_overflow =
+            stg_max(gct->max_n_todo_overflow, ws->n_todo_overflow);
     }
 
 #if defined(THREADED_RTS)