Avoid accumulating slop in the pinned_object_block.

The pinned_object_block is where we allocate small pinned ByteArray#
objects.  At a GC the pinned_object_block was being treated like other
large objects and promoted to the next step/generation, even if it was
only partly full.  Under some ByteString-heavy workloads this would
accumulate on average 2k of slop per GC, and this memory is never
released until the ByteArray# objects in the block are freed.

So now, we keep allocating into the pinned_object_block until it is
completely full, at which point it is handed over to the GC as before.
The pinned_object_block might therefore contain objects which a large
range of ages, but I don't think this is any worse than the situation
before.  We still have the fragmentation issue in general, but the new
scheme can improve the memory overhead for some workloads
dramatically.

3 files changed, 24 insertions, 10 deletions

diff --git a/rts/sm/GC.c b/rts/sm/GC.c
index 05bc8f22fb..30361401cc 100644
--- a/rts/sm/GC.c
+++ b/rts/sm/GC.c
@@ -597,11 +597,6 @@ GarbageCollect (rtsBool force_major_gc,
   // update the max size of older generations after a major GC
   resize_generations();
   
-  // Start a new pinned_object_block
-  for (n = 0; n < n_capabilities; n++) {
-      capabilities[n].pinned_object_block = NULL;
-  }
-
   // Free the mark stack.
   if (mark_stack_top_bd != NULL) {
       debugTrace(DEBUG_gc, "mark stack: %d blocks",
diff --git a/rts/sm/Sanity.c b/rts/sm/Sanity.c
index 8ebb9a2513..0ec552c047 100644
--- a/rts/sm/Sanity.c
+++ b/rts/sm/Sanity.c
@@ -789,6 +789,7 @@ findMemoryLeak (void)
 
     for (i = 0; i < n_capabilities; i++) {
         markBlocks(nurseries[i].blocks);
+        markBlocks(capabilities[i].pinned_object_block);
     }
 
 #ifdef PROFILING
@@ -880,6 +881,9 @@ memInventory (rtsBool show)
   for (i = 0; i < n_capabilities; i++) {
       ASSERT(countBlocks(nurseries[i].blocks) == nurseries[i].n_blocks);
       nursery_blocks += nurseries[i].n_blocks;
+      if (capabilities[i].pinned_object_block != NULL) {
+          nursery_blocks += capabilities[i].pinned_object_block->blocks;
+      }
   }
 
   retainer_blocks = 0;
diff --git a/rts/sm/Storage.c b/rts/sm/Storage.c
index ae3433a9b2..f8a9e559bf 100644
--- a/rts/sm/Storage.c
+++ b/rts/sm/Storage.c
@@ -657,17 +657,32 @@ allocatePinned (Capability *cap, lnat n)
     // If we don't have a block of pinned objects yet, or the current
     // one isn't large enough to hold the new object, allocate a new one.
     if (bd == NULL || (bd->free + n) > (bd->start + BLOCK_SIZE_W)) {
+        // The pinned_object_block remains attached to the capability
+        // until it is full, even if a GC occurs.  We want this
+        // behaviour because otherwise the unallocated portion of the
+        // block would be forever slop, and under certain workloads
+        // (allocating a few ByteStrings per GC) we accumulate a lot
+        // of slop.
+        //
+        // So, the pinned_object_block is initially marked
+        // BF_EVACUATED so the GC won't touch it.  When it is full,
+        // we place it on the large_objects list, and at the start of
+        // the next GC the BF_EVACUATED flag will be cleared, and the
+        // block will be promoted as usual (if anything in it is
+        // live).
         ACQUIRE_SM_LOCK;
-	cap->pinned_object_block = bd = allocBlock();
-	dbl_link_onto(bd, &g0->large_objects);
-	g0->n_large_blocks++;
+        if (bd != NULL) {
+            dbl_link_onto(bd, &g0->large_objects);
+            g0->n_large_blocks++;
+            g0->n_new_large_words += bd->free - bd->start;
+        }
+        cap->pinned_object_block = bd = allocBlock();
         RELEASE_SM_LOCK;
         initBdescr(bd, g0, g0);
-	bd->flags  = BF_PINNED | BF_LARGE;
+        bd->flags  = BF_PINNED | BF_LARGE | BF_EVACUATED;
 	bd->free   = bd->start;
     }
 
-    g0->n_new_large_words += n;
     p = bd->free;
     bd->free += n;
     return p;