RFR: Elastic TLABs may avoid racily peeking into the free set

[Aleksey Shipilev]

CollectedHeap::unsafe_max_tlab_alloc is used to determine the momentary max tlab size, peeking
racily into GC state (thus "unsafe_"). This is needed to avoid retiring regions that would not fit
the max tlab if there is not enough space available in current allocation region. Because it is
racy, it can grossly miscalculate under heavy allocations, when other threads are actively mutating
the current allocation region. Fortunately, with Elastic TLABs, this peeking is not needed, as
allocation path would figure it out itself.

diff -r e2c2bf800963 src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp
--- a/src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp        Tue Jul 17 16:54:17 2018 +0200
+++ b/src/hotspot/share/gc/shenandoah/shenandoahHeap.cpp        Tue Jul 17 17:29:56 2018 +0200
@@ -1370,13 +1370,19 @@

 bool ShenandoahHeap::supports_tlab_allocation() const {
   return true;
 }

-size_t  ShenandoahHeap::unsafe_max_tlab_alloc(Thread *thread) const {
-  // Returns size in bytes
-  return MIN2(_free_set->unsafe_peek_free(), ShenandoahHeapRegion::max_tlab_size_bytes());
+// Returns size in bytes
+size_t ShenandoahHeap::unsafe_max_tlab_alloc(Thread *thread) const {
+  if (ShenandoahElasticTLAB) {
+    // With Elastic TLABs, return the max allowed size, and let the allocation path
+    // figure out the safe size for current allocation.
+    return ShenandoahHeapRegion::max_tlab_size_bytes();
+  } else {
+    return MIN2(_free_set->unsafe_peek_free(), ShenandoahHeapRegion::max_tlab_size_bytes());
+  }
 }

 size_t ShenandoahHeap::max_tlab_size() const {
   // Returns size in words
   return ShenandoahHeapRegion::max_tlab_size_words();

Testing: ad-hoc benchmarks, tier1_gc_shenandoah, tier3_gc_shenandoah (running)

Thanks,
-Aleksey