Split v8_enable_concurrent_marking into two flags

The new flags are
- v8_enable_atomic_object_field_writes that makes field write operations
  relaxed atomic.
- v8_enable_atomic_marking_state that makes the marking state and the
  write-barrier thread-safe.

The motivation is that we want to disable atomic object fields while
keeping the marking states thread-safe. This allows us to increase
TSAN coverage for background compilation and streaming tasks while
keeping the write-barrier used by the tasks thread-safe.

Bug: v8:10988
Change-Id: I11d66954dda4bf36d24c5e6f14ee5bc7a0f86094
Reviewed-on: https://chromium-review.googlesource.com/c/v8/v8/+/2448467
Reviewed-by: Georg Neis <neis@chromium.org>
Reviewed-by: Michael Achenbach <machenbach@chromium.org>
Commit-Queue: Ulan Degenbaev <ulan@chromium.org>
Cr-Commit-Position: refs/heads/master@{#70329}

BUILD.gn

@@ -131,7 +131,15 @@ declare_args() {
   # Sets -dV8_TRACE_FEEDBACK_UPDATES.
   v8_enable_trace_feedback_updates = false
 
-  # Sets -dV8_CONCURRENT_MARKING
+  # Sets -dV8_ATOMIC_OBJECT_FIELD_WRITES and turns all field write operations
+  # into relaxed atomic operations.
+  v8_enable_atomic_object_field_writes = ""
+
+  # Sets -dV8_ATOMIC_MARKING_STATE
+  v8_enable_atomic_marking_state = ""
+
+  # Controls the default values of v8_enable_atomic_object_field_writes and
+  # v8_enable_concurrent_marking_state. See the default setting code below.
   v8_enable_concurrent_marking = true
 
   # Runs mksnapshot with --turbo-profiling. After building in this
@@ -318,6 +326,16 @@ if (v8_enable_heap_sandbox == "") {
 if (v8_enable_single_generation == "") {
   v8_enable_single_generation = v8_disable_write_barriers
 }
+if (v8_enable_atomic_object_field_writes == "") {
+  v8_enable_atomic_object_field_writes = v8_enable_concurrent_marking
+}
+if (v8_enable_atomic_marking_state == "") {
+  v8_enable_atomic_marking_state = v8_enable_concurrent_marking
+}
+assert(!v8_enable_concurrent_marking || v8_enable_atomic_object_field_writes,
+       "Concurrent marking requires atomic object field writes.")
+assert(!v8_enable_concurrent_marking || v8_enable_atomic_marking_state,
+       "Concurrent marking requires atomic marking state.")
 
 # Toggle pointer compression for correctness fuzzing when building the
 # clang_x64_pointer_compression toolchain. We'll correctness-compare the
@@ -611,8 +629,11 @@ config("features") {
   if (v8_use_external_startup_data) {
     defines += [ "V8_USE_EXTERNAL_STARTUP_DATA" ]
   }
-  if (v8_enable_concurrent_marking) {
+  if (v8_enable_atomic_object_field_writes) {
-    defines += [ "V8_CONCURRENT_MARKING" ]
+    defines += [ "V8_ATOMIC_OBJECT_FIELD_WRITES" ]
+  }
+  if (v8_enable_atomic_marking_state) {
+    defines += [ "V8_ATOMIC_MARKING_STATE" ]
   }
   if (v8_enable_lazy_source_positions) {
     defines += [ "V8_ENABLE_LAZY_SOURCE_POSITIONS" ]
@@ -1671,6 +1692,9 @@ action("v8_dump_build_config") {
     "is_ubsan_vptr=$is_ubsan_vptr",
     "target_cpu=\"$target_cpu\"",
     "v8_current_cpu=\"$v8_current_cpu\"",
+    "v8_enable_atomic_marking_state=$v8_enable_atomic_marking_state",
+    "v8_enable_atomic_object_field_writes=" +
+        "$v8_enable_atomic_object_field_writes",
     "v8_enable_concurrent_marking=$v8_enable_concurrent_marking",
     "v8_enable_i18n_support=$v8_enable_i18n_support",
     "v8_enable_verify_predictable=$v8_enable_verify_predictable",

infra/mb/mb_config.pyl

@@ -661,7 +661,13 @@
     },
 
     'disable_concurrent_marking': {
-      'gn_args': 'v8_enable_concurrent_marking=false',
+      # Disable concurrent marking and atomic object field writes in order to
+      # increase the TSAN coverage for background tasks. We need to keep the 
+      # atomic marking state enabled because that is needed for the concurrent
+      # write-barrier used by background compilation.
+      'gn_args': 'v8_enable_concurrent_marking=false '
+                 'v8_enable_atomic_object_field_writes=false '
+                 'v8_enable_atomic_marking_state=true ',
     },
 
     'disable_pgo': {

src/flags/flag-definitions.h

@@ -999,7 +999,7 @@ DEFINE_BOOL(scavenge_separate_stack_scanning, false,
             "use a separate phase for stack scanning in scavenge")
 DEFINE_BOOL(trace_parallel_scavenge, false, "trace parallel scavenge")
 DEFINE_BOOL(write_protect_code_memory, true, "write protect code memory")
-#ifdef V8_CONCURRENT_MARKING
+#if defined(V8_ATOMIC_MARKING_STATE) && defined(V8_ATOMIC_OBJECT_FIELD_WRITES)
 #define V8_CONCURRENT_MARKING_BOOL true
 #else
 #define V8_CONCURRENT_MARKING_BOOL false

src/heap/concurrent-marking.cc

@@ -375,10 +375,14 @@ ConcurrentMarking::ConcurrentMarking(Heap* heap,
     : heap_(heap),
       marking_worklists_(marking_worklists),
       weak_objects_(weak_objects) {
-// The runtime flag should be set only if the compile time flag was set.
+#ifndef V8_ATOMIC_MARKING_STATE
-#ifndef V8_CONCURRENT_MARKING
+  // Concurrent and parallel marking require atomic marking state.
   CHECK(!FLAG_concurrent_marking && !FLAG_parallel_marking);
 #endif
+#ifndef V8_ATOMIC_OBJECT_FIELD_WRITES
+  // Concurrent marking requires atomic object field writes.
+  CHECK(!FLAG_concurrent_marking);
+#endif
 }
 
 void ConcurrentMarking::Run(int task_id, TaskState* task_state) {

src/heap/incremental-marking.h

@@ -81,7 +81,7 @@ class V8_EXPORT_PRIVATE IncrementalMarking final {
   static constexpr size_t kGlobalActivationThreshold = 0;
 #endif
 
-#ifdef V8_CONCURRENT_MARKING
+#ifdef V8_ATOMIC_MARKING_STATE
   static const AccessMode kAtomicity = AccessMode::ATOMIC;
 #else
   static const AccessMode kAtomicity = AccessMode::NON_ATOMIC;

src/heap/mark-compact.h

@@ -434,11 +434,11 @@ class MainMarkingVisitor final
 // Collector for young and old generation.
 class MarkCompactCollector final : public MarkCompactCollectorBase {
  public:
-#ifdef V8_CONCURRENT_MARKING
+#ifdef V8_ATOMIC_MARKING_STATE
   using MarkingState = MajorMarkingState;
 #else
   using MarkingState = MajorNonAtomicMarkingState;
-#endif  // V8_CONCURRENT_MARKING
+#endif  // V8_ATOMIC_MARKING_STATE
   using AtomicMarkingState = MajorAtomicMarkingState;
   using NonAtomicMarkingState = MajorNonAtomicMarkingState;
 

src/objects/tagged-field-inl.h

@@ -70,7 +70,7 @@ T TaggedField<T, kFieldOffset>::load(const Isolate* isolate, HeapObject host,
 // static
 template <typename T, int kFieldOffset>
 void TaggedField<T, kFieldOffset>::store(HeapObject host, T value) {
-#ifdef V8_CONCURRENT_MARKING
+#ifdef V8_ATOMIC_OBJECT_FIELD_WRITES
   Relaxed_Store(host, value);
 #else
   *location(host) = full_to_tagged(value.ptr());
@@ -80,7 +80,7 @@ void TaggedField<T, kFieldOffset>::store(HeapObject host, T value) {
 // static
 template <typename T, int kFieldOffset>
 void TaggedField<T, kFieldOffset>::store(HeapObject host, int offset, T value) {
-#ifdef V8_CONCURRENT_MARKING
+#ifdef V8_ATOMIC_OBJECT_FIELD_WRITES
   Relaxed_Store(host, offset, value);
 #else
   *location(host, offset) = full_to_tagged(value.ptr());