Background house-keeping & pre-existing thread

Break apart ProcessBackgroundActions into an "Init" and "Tick" stage and
leave periodic invocation to the caller. This lets users reduce
the number of threads spun up by 1 if they already have a background
house keeping thread that knows when (& perhaps even how much) memory to
free.

Author: vlovich

tcmalloc/background.cc

@@ -99,43 +99,54 @@ void ShuffleCpuCaches() {
 GOOGLE_MALLOC_SECTION_END
 
 // Release memory to the system at a constant rate.
-void MallocExtension_Internal_ProcessBackgroundActions() {
-  tcmalloc::MallocExtension::MarkThreadIdle();
+static absl::Time prev_time;
+static absl::Time last_shuffle;
 
+void MallocExtension_Internal_ProcessBackgroundActionsInit() {
   // Initialize storage for ReleasePerCpuMemoryToOS().
   CPU_ZERO(&tcmalloc::tcmalloc_internal::prev_allowed_cpus);
 
-  absl::Time prev_time = absl::Now();
+  prev_time = absl::Now();
+  last_shuffle = absl::InfinitePast();
+}
+
+void MallocExtension_Internal_ProcessBackgroundActions() {
   constexpr absl::Duration kSleepTime = absl::Seconds(1);
 
+  tcmalloc::MallocExtension::MarkThreadIdle();
+  MallocExtension_Internal_ProcessBackgroundActionsInit();
+
+  while (true) {
+    MallocExtension_Internal_ProcessBackgroundActionsTick();
+    absl::SleepFor(kSleepTime);
+  }
+}
+
+void MallocExtension_Internal_ProcessBackgroundActionsTick() {
   // Shuffle per-cpu caches once per kCpuCacheShufflePeriod secs.
   constexpr absl::Duration kCpuCacheShufflePeriod = absl::Seconds(5);
-  absl::Time last_shuffle = absl::InfinitePast();
 
-  while (true) {
-    absl::Time now = absl::Now();
-    const ssize_t bytes_to_release =
-        static_cast<size_t>(tcmalloc::tcmalloc_internal::Parameters::
-                                background_release_rate()) *
-        absl::ToDoubleSeconds(now - prev_time);
-    if (bytes_to_release > 0) {  // may be negative if time goes backwards
-      tcmalloc::MallocExtension::ReleaseMemoryToSystem(bytes_to_release);
-    }
+  absl::Time now = absl::Now();
+  const ssize_t bytes_to_release =
+      static_cast<size_t>(tcmalloc::tcmalloc_internal::Parameters::
+                              background_release_rate()) *
+      absl::ToDoubleSeconds(now - prev_time);
+  if (bytes_to_release > 0) {  // may be negative if time goes backwards
+    tcmalloc::MallocExtension::ReleaseMemoryToSystem(bytes_to_release);
+  }
 
-    tcmalloc::tcmalloc_internal::ReleasePerCpuMemoryToOS();
+  tcmalloc::tcmalloc_internal::ReleasePerCpuMemoryToOS();
 
-    const bool shuffle_per_cpu_caches =
-        tcmalloc::tcmalloc_internal::Parameters::shuffle_per_cpu_caches();
+  const bool shuffle_per_cpu_caches =
+      tcmalloc::tcmalloc_internal::Parameters::shuffle_per_cpu_caches();
 
-    if (shuffle_per_cpu_caches) {
-      if (now - last_shuffle >= kCpuCacheShufflePeriod) {
-        tcmalloc::tcmalloc_internal::ShuffleCpuCaches();
-        last_shuffle = now;
-      }
+  if (shuffle_per_cpu_caches) {
+    if (now - last_shuffle >= kCpuCacheShufflePeriod) {
+      tcmalloc::tcmalloc_internal::ShuffleCpuCaches();
+      last_shuffle = now;
     }
-
-    tcmalloc::tcmalloc_internal::Static().sharded_transfer_cache().Plunder();
-    prev_time = now;
-    absl::SleepFor(kSleepTime);
   }
+
+  tcmalloc::tcmalloc_internal::Static().sharded_transfer_cache().Plunder();
+  prev_time = now;
 }

tcmalloc/internal_malloc_extension.h

@@ -101,6 +101,8 @@ ABSL_ATTRIBUTE_WEAK void MallocExtension_Internal_SetProfileSamplingRate(
     int64_t);
 
 ABSL_ATTRIBUTE_WEAK void MallocExtension_Internal_ProcessBackgroundActions();
+ABSL_ATTRIBUTE_WEAK void MallocExtension_Internal_ProcessBackgroundActionsInit();
+ABSL_ATTRIBUTE_WEAK void MallocExtension_Internal_ProcessBackgroundActionsTick();
 
 ABSL_ATTRIBUTE_WEAK tcmalloc::MallocExtension::BytesPerSecond
 MallocExtension_Internal_GetBackgroundReleaseRate();

tcmalloc/malloc_extension.cc

@@ -425,9 +425,26 @@ void MallocExtension::ProcessBackgroundActions() {
 #endif
 }
 
+void MallocExtension::ProcessBackgroundActionsInit() {
+#if ABSL_INTERNAL_HAVE_WEAK_MALLOCEXTENSION_STUBS
+  if (NeedsProcessBackgroundActions()) {
+    MallocExtension_Internal_ProcessBackgroundActionsInit();
+  }
+#endif
+}
+
+void MallocExtension::ProcessBackgroundActionsTick() {
+#if ABSL_INTERNAL_HAVE_WEAK_MALLOCEXTENSION_STUBS
+  if (NeedsProcessBackgroundActions()) {
+    MallocExtension_Internal_ProcessBackgroundActionsTick();
+  }
+#endif
+}
+
 bool MallocExtension::NeedsProcessBackgroundActions() {
 #if ABSL_INTERNAL_HAVE_WEAK_MALLOCEXTENSION_STUBS
-  return &MallocExtension_Internal_ProcessBackgroundActions != nullptr;
+  static bool needed = &MallocExtension_Internal_ProcessBackgroundActions != nullptr;
+  return needed;
 #else
   return false;
 #endif

tcmalloc/malloc_extension.h

@@ -433,6 +433,19 @@ class MallocExtension final {
   // null if the implementation does not support profiling.
   static AllocationProfilingToken StartAllocationProfiling();
 
+  // Initializes the state needed for this thread to call
+  // ProcessBackgroundActionsTick. This is the counterpart to
+  // ProcessBackgroundActions that allows co-operative task keeping on a thread
+  // that is performing other task keeping.
+  static void ProcessBackgroundActionsInit();
+
+  // This is to be called periodically to run housekeeping actions for the
+  // allocator off of the main allocation paths of new/delete. Note that unlike
+  // ProcessBackgroundActions, the thread is not automatically marked as
+  // idle/busy and the caller is responsible for doing that correctly.
+  // See ProcessBackgroundActions for details of the actions performed.
+  static void ProcessBackgroundActionsTick();
+
   // Runs housekeeping actions for the allocator off of the main allocation path
   // of new/delete.  As of 2020, this includes:
   // * Inspecting the current CPU mask and releasing memory from inaccessible