8079555: REDO - Determining the desired PLAB size adjusts to the the number of threads at the wrong place

Calculate the desired PLAB value for a single thread and then return desired PLAB size according to the current number of threads when needed

Reviewed-by: jmasa, tschatzl

hotspot/src/share/vm/gc/cms/parNewGeneration.cpp

@@ -1021,7 +1021,7 @@ void ParNewGeneration::collect(bool   full,
   to()->set_concurrent_iteration_safe_limit(to()->top());
 
   if (ResizePLAB) {
-    plab_stats()->adjust_desired_plab_sz(active_workers);
+    plab_stats()->adjust_desired_plab_sz();
   }
 
   if (PrintGC && !PrintGCDetails) {
@@ -1059,6 +1059,10 @@ void ParNewGeneration::collect(bool   full,
   _gc_tracer.report_gc_end(_gc_timer->gc_end(), _gc_timer->time_partitions());
 }
 
+size_t ParNewGeneration::desired_plab_sz() {
+  return _plab_stats.desired_plab_sz(GenCollectedHeap::heap()->workers()->active_workers());
+}
+
 static int sum;
 void ParNewGeneration::waste_some_time() {
   for (int i = 0; i < 100; i++) {

hotspot/src/share/vm/gc/cms/parNewGeneration.hpp

@@ -411,9 +411,7 @@ class ParNewGeneration: public DefNewGeneration {
     return &_plab_stats;
   }
 
-  size_t desired_plab_sz() {
-    return _plab_stats.desired_plab_sz();
-  }
+  size_t desired_plab_sz();
 
   const ParNewTracer* gc_tracer() const {
     return &_gc_tracer;

hotspot/src/share/vm/gc/g1/g1Allocator.cpp

@@ -86,7 +86,7 @@ void G1DefaultAllocator::init_gc_alloc_regions(EvacuationInfo& evacuation_info)
                             &_retained_old_gc_alloc_region);
 }
 
-void G1DefaultAllocator::release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info) {
+void G1DefaultAllocator::release_gc_alloc_regions(EvacuationInfo& evacuation_info) {
   AllocationContext_t context = AllocationContext::current();
   evacuation_info.set_allocation_regions(survivor_gc_alloc_region(context)->count() +
                                          old_gc_alloc_region(context)->count());
@@ -102,8 +102,8 @@ void G1DefaultAllocator::release_gc_alloc_regions(uint no_of_gc_workers, Evacuat
   }
 
   if (ResizePLAB) {
-    _g1h->alloc_buffer_stats(InCSetState::Young)->adjust_desired_plab_sz(no_of_gc_workers);
-    _g1h->alloc_buffer_stats(InCSetState::Old)->adjust_desired_plab_sz(no_of_gc_workers);
+    _g1h->alloc_buffer_stats(InCSetState::Young)->adjust_desired_plab_sz();
+    _g1h->alloc_buffer_stats(InCSetState::Old)->adjust_desired_plab_sz();
   }
 }
 

hotspot/src/share/vm/gc/g1/g1Allocator.hpp

@@ -53,7 +53,7 @@ public:
    virtual void release_mutator_alloc_region() = 0;
 
    virtual void init_gc_alloc_regions(EvacuationInfo& evacuation_info) = 0;
-   virtual void release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info) = 0;
+   virtual void release_gc_alloc_regions(EvacuationInfo& evacuation_info) = 0;
    virtual void abandon_gc_alloc_regions() = 0;
 
    virtual MutatorAllocRegion*    mutator_alloc_region(AllocationContext_t context) = 0;
@@ -114,7 +114,7 @@ public:
   virtual void release_mutator_alloc_region();
 
   virtual void init_gc_alloc_regions(EvacuationInfo& evacuation_info);
-  virtual void release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info);
+  virtual void release_gc_alloc_regions(EvacuationInfo& evacuation_info);
   virtual void abandon_gc_alloc_regions();
 
   virtual bool is_retained_old_region(HeapRegion* hr) {

hotspot/src/share/vm/gc/g1/g1CollectedHeap.cpp

@@ -5644,7 +5644,7 @@ void G1CollectedHeap::evacuate_collection_set(EvacuationInfo& evacuation_info) {
     phase_times->record_string_dedup_fixup_time(fixup_time_ms);
   }
 
-  _allocator->release_gc_alloc_regions(n_workers, evacuation_info);
+  _allocator->release_gc_alloc_regions(evacuation_info);
   g1_rem_set()->cleanup_after_oops_into_collection_set_do();
 
   // Reset and re-enable the hot card cache.

hotspot/src/share/vm/gc/g1/g1CollectedHeap.hpp

@@ -281,7 +281,7 @@ private:
   void init_gc_alloc_regions(EvacuationInfo& evacuation_info);
 
   // It releases the GC alloc regions at the end of a GC.
-  void release_gc_alloc_regions(uint no_of_gc_workers, EvacuationInfo& evacuation_info);
+  void release_gc_alloc_regions(EvacuationInfo& evacuation_info);
 
   // It does any cleanup that needs to be done on the GC alloc regions
   // before a Full GC.

hotspot/src/share/vm/gc/g1/g1CollectedHeap.inline.hpp

@@ -49,7 +49,7 @@ PLABStats* G1CollectedHeap::alloc_buffer_stats(InCSetState dest) {
 }
 
 size_t G1CollectedHeap::desired_plab_sz(InCSetState dest) {
-  size_t gclab_word_size = alloc_buffer_stats(dest)->desired_plab_sz();
+  size_t gclab_word_size = alloc_buffer_stats(dest)->desired_plab_sz(G1CollectedHeap::heap()->workers()->active_workers());
   // Prevent humongous PLAB sizes for two reasons:
   // * PLABs are allocated using a similar paths as oops, but should
   //   never be in a humongous region

hotspot/src/share/vm/gc/shared/plab.cpp

@@ -109,10 +109,15 @@ void PLAB::undo_allocation(HeapWord* obj, size_t word_sz) {
   }
 }
 
-// Compute desired plab size and latch result for later
+// Calculates plab size for current number of gc worker threads.
+size_t PLABStats::desired_plab_sz(uint no_of_gc_workers) {
+  return MAX2(min_size(), (size_t)align_object_size(_desired_net_plab_sz / no_of_gc_workers));
+}
+
+// Compute desired plab size for one gc worker thread and latch result for later
 // use. This should be called once at the end of parallel
 // scavenge; it clears the sensor accumulators.
-void PLABStats::adjust_desired_plab_sz(uint no_of_gc_workers) {
+void PLABStats::adjust_desired_plab_sz() {
   assert(ResizePLAB, "Not set");
 
   assert(is_object_aligned(max_size()) && min_size() <= max_size(),
@@ -135,7 +140,8 @@ void PLABStats::adjust_desired_plab_sz(uint no_of_gc_workers) {
     target_refills = 1;
   }
   size_t used = _allocated - _wasted - _unused;
-  size_t recent_plab_sz = used / (target_refills * no_of_gc_workers);
+  // Assumed to have 1 gc worker thread
+  size_t recent_plab_sz = used / target_refills;
   // Take historical weighted average
   _filter.sample(recent_plab_sz);
   // Clip from above and below, and align to object boundary
@@ -144,9 +150,9 @@ void PLABStats::adjust_desired_plab_sz(uint no_of_gc_workers) {
   new_plab_sz = align_object_size(new_plab_sz);
   // Latch the result
   if (PrintPLAB) {
-    gclog_or_tty->print(" (plab_sz = " SIZE_FORMAT " desired_plab_sz = " SIZE_FORMAT ") ", recent_plab_sz, new_plab_sz);
+    gclog_or_tty->print(" (plab_sz = " SIZE_FORMAT " desired_net_plab_sz = " SIZE_FORMAT ") ", recent_plab_sz, new_plab_sz);
   }
-  _desired_plab_sz = new_plab_sz;
+  _desired_net_plab_sz = new_plab_sz;
 
   reset();
 }

hotspot/src/share/vm/gc/shared/plab.hpp

@@ -154,7 +154,7 @@ class PLABStats VALUE_OBJ_CLASS_SPEC {
   size_t _wasted;             // of which wasted (internal fragmentation)
   size_t _undo_wasted;        // of which wasted on undo (is not used for calculation of PLAB size)
   size_t _unused;             // Unused in last buffer
-  size_t _desired_plab_sz;// Output of filter (below), suitably trimmed and quantized
+  size_t _desired_net_plab_sz;// Output of filter (below), suitably trimmed and quantized
   AdaptiveWeightedAverage
          _filter;             // Integrator with decay
 
@@ -165,12 +165,12 @@ class PLABStats VALUE_OBJ_CLASS_SPEC {
     _unused      = 0;
   }
  public:
-  PLABStats(size_t desired_plab_sz_, unsigned wt) :
+  PLABStats(size_t desired_net_plab_sz_, unsigned wt) :
     _allocated(0),
     _wasted(0),
     _undo_wasted(0),
     _unused(0),
-    _desired_plab_sz(desired_plab_sz_),
+    _desired_net_plab_sz(desired_net_plab_sz_),
     _filter(wt)
   { }
 
@@ -182,13 +182,12 @@ class PLABStats VALUE_OBJ_CLASS_SPEC {
     return PLAB::max_size();
   }
 
-  size_t desired_plab_sz() {
-    return _desired_plab_sz;
-  }
+  // Calculates plab size for current number of gc worker threads.
+  size_t desired_plab_sz(uint no_of_gc_workers);
 
-  // Updates the current desired PLAB size. Computes the new desired PLAB size,
+  // Updates the current desired PLAB size. Computes the new desired PLAB size with one gc worker thread,
   // updates _desired_plab_sz and clears sensor accumulators.
-  void adjust_desired_plab_sz(uint no_of_gc_workers);
+  void adjust_desired_plab_sz();
 
   void add_allocated(size_t v) {
     Atomic::add_ptr(v, &_allocated);