8267579: Thread::cooked_allocated_bytes() hits assert(left >= right) failed: avoid underflow

Backport-of: c420735

Author: Markus Grönlund

src/hotspot/share/gc/shared/threadLocalAllocBuffer.cpp

@@ -31,6 +31,7 @@
 #include "memory/resourceArea.hpp"
 #include "memory/universe.hpp"
 #include "oops/oop.inline.hpp"
+#include "runtime/atomic.hpp"
 #include "runtime/perfData.hpp"
 #include "runtime/thread.inline.hpp"
 #include "runtime/threadSMR.hpp"
@@ -473,3 +474,11 @@ size_t ThreadLocalAllocBuffer::end_reserve() {
   size_t reserve_size = Universe::heap()->tlab_alloc_reserve();
   return MAX2(reserve_size, (size_t)_reserve_for_allocation_prefetch);
 }
+
+const HeapWord* ThreadLocalAllocBuffer::start_relaxed() const {
+  return Atomic::load(&_start);
+}
+
+const HeapWord* ThreadLocalAllocBuffer::top_relaxed() const {
+  return Atomic::load(&_top);
+}

src/hotspot/share/gc/shared/threadLocalAllocBuffer.hpp

@@ -129,6 +129,10 @@ class ThreadLocalAllocBuffer: public CHeapObj<mtThread> {
   size_t refill_waste_limit() const              { return _refill_waste_limit; }
   size_t bytes_since_last_sample_point() const   { return _bytes_since_last_sample_point; }
 
+  // For external inspection.
+  const HeapWord* start_relaxed() const;
+  const HeapWord* top_relaxed() const;
+
   // Allocate size HeapWords. The memory is NOT initialized to zero.
   inline HeapWord* allocate(size_t size);
 

src/hotspot/share/runtime/thread.inline.hpp

@@ -41,7 +41,14 @@
 inline jlong Thread::cooked_allocated_bytes() {
   jlong allocated_bytes = Atomic::load_acquire(&_allocated_bytes);
   if (UseTLAB) {
-    size_t used_bytes = tlab().used_bytes();
+    // These reads are unsynchronized and unordered with the thread updating its tlab pointers.
+    // Use only if top > start && used_bytes <= max_tlab_size_bytes.
+    const HeapWord* const top = tlab().top_relaxed();
+    const HeapWord* const start = tlab().start_relaxed();
+    if (top <= start) {
+      return allocated_bytes;
+    }
+    const size_t used_bytes = pointer_delta(top, start, 1);
     if (used_bytes <= ThreadLocalAllocBuffer::max_size_in_bytes()) {
       // Comparing used_bytes with the maximum allowed size will ensure
       // that we don't add the used bytes from a semi-initialized TLAB