Initial Commit

26 files changed, 9346 insertions, 0 deletions

diff --git a/absl/synchronization/BUILD.bazel b/absl/synchronization/BUILD.bazel
new file mode 100644
index 000000000000..bddd2eca8083
--- /dev/null
+++ b/absl/synchronization/BUILD.bazel
@@ -0,0 +1,178 @@
+#
+# Copyright 2017 The Abseil Authors.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#      http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+#
+
+load(
+    "//absl:copts.bzl",
+    "ABSL_DEFAULT_COPTS",
+    "ABSL_TEST_COPTS",
+)
+
+package(default_visibility = ["//visibility:public"])
+
+licenses(["notice"])  # Apache 2.0
+
+# Internal data structure for efficiently detecting mutex dependency cycles
+cc_library(
+    name = "graphcycles_internal",
+    srcs = [
+        "internal/graphcycles.cc",
+    ],
+    hdrs = [
+        "internal/graphcycles.h",
+    ],
+    copts = ABSL_DEFAULT_COPTS,
+    deps = [
+        "//absl/base",
+        "//absl/base:core_headers",
+        "//absl/base:malloc_internal",
+    ],
+)
+
+cc_library(
+    name = "synchronization",
+    srcs = [
+        "barrier.cc",
+        "blocking_counter.cc",
+        "internal/create_thread_identity.cc",
+        "internal/per_thread_sem.cc",
+        "internal/waiter.cc",
+        "notification.cc",
+    ] + select({
+        "//conditions:default": ["mutex.cc"],
+    }),
+    hdrs = [
+        "barrier.h",
+        "blocking_counter.h",
+        "internal/create_thread_identity.h",
+        "internal/kernel_timeout.h",
+        "internal/mutex_nonprod.inc",
+        "internal/per_thread_sem.h",
+        "internal/waiter.h",
+        "mutex.h",
+        "notification.h",
+    ],
+    copts = ABSL_DEFAULT_COPTS,
+    deps = [
+        ":graphcycles_internal",
+        "//absl/base",
+        "//absl/base:base_internal",
+        "//absl/base:config",
+        "//absl/base:core_headers",
+        "//absl/base:dynamic_annotations",
+        "//absl/base:malloc_extension",
+        "//absl/base:malloc_internal",
+        "//absl/debugging:stacktrace",
+        "//absl/time",
+    ],
+)
+
+cc_test(
+    name = "blocking_counter_test",
+    size = "small",
+    srcs = ["blocking_counter_test.cc"],
+    copts = ABSL_TEST_COPTS,
+    deps = [
+        ":synchronization",
+        "//absl/time",
+        "@com_google_googletest//:gtest_main",
+    ],
+)
+
+cc_test(
+    name = "graphcycles_test",
+    size = "medium",
+    srcs = ["internal/graphcycles_test.cc"],
+    copts = ABSL_TEST_COPTS,
+    deps = [
+        ":graphcycles_internal",
+        "//absl/base",
+        "//absl/base:core_headers",
+        "@com_google_googletest//:gtest_main",
+    ],
+)
+
+cc_library(
+    name = "thread_pool",
+    testonly = 1,
+    hdrs = ["internal/thread_pool.h"],
+    deps = [
+        ":synchronization",
+        "//absl/base:core_headers",
+    ],
+)
+
+cc_test(
+    name = "mutex_test",
+    size = "large",
+    timeout = "moderate",
+    srcs = ["mutex_test.cc"],
+    copts = ABSL_TEST_COPTS,
+    tags = [
+        "no_test_loonix",  # Too slow.
+    ],
+    deps = [
+        ":synchronization",
+        ":thread_pool",
+        "//absl/base",
+        "//absl/base:core_headers",
+        "//absl/memory",
+        "//absl/time",
+        "@com_google_googletest//:gtest_main",
+    ],
+)
+
+cc_test(
+    name = "notification_test",
+    size = "small",
+    srcs = ["notification_test.cc"],
+    copts = ABSL_TEST_COPTS,
+    deps = [
+        ":synchronization",
+        "//absl/time",
+        "@com_google_googletest//:gtest_main",
+    ],
+)
+
+cc_library(
+    name = "per_thread_sem_test_common",
+    testonly = 1,
+    srcs = ["internal/per_thread_sem_test.cc"],
+    copts = ABSL_TEST_COPTS,
+    deps = [
+        ":synchronization",
+        "//absl/base",
+        "//absl/base:malloc_extension",
+        "//absl/strings",
+        "//absl/time",
+        "@com_google_googletest//:gtest",
+    ],
+    alwayslink = 1,
+)
+
+cc_test(
+    name = "per_thread_sem_test",
+    size = "medium",
+    copts = ABSL_TEST_COPTS,
+    deps = [
+        ":per_thread_sem_test_common",
+        ":synchronization",
+        "//absl/base",
+        "//absl/base:malloc_extension",
+        "//absl/strings",
+        "//absl/time",
+        "@com_google_googletest//:gtest_main",
+    ],
+)
diff --git a/absl/synchronization/barrier.cc b/absl/synchronization/barrier.cc
new file mode 100644
index 000000000000..a1b3ad5c2174
--- /dev/null
+++ b/absl/synchronization/barrier.cc
@@ -0,0 +1,50 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/synchronization/barrier.h"
+
+#include "absl/base/internal/raw_logging.h"
+#include "absl/synchronization/mutex.h"
+
+namespace absl {
+
+// Return whether int *arg is zero.
+static bool IsZero(void *arg) {
+  return 0 == *reinterpret_cast<int *>(arg);
+}
+
+bool Barrier::Block() {
+  MutexLock l(&this->lock_);
+
+  this->num_to_block_--;
+  if (this->num_to_block_ < 0) {
+    ABSL_RAW_LOG(
+        FATAL,
+        "Block() called too many times.  num_to_block_=%d out of total=%d",
+        this->num_to_block_, this->num_to_exit_);
+  }
+
+  this->lock_.Await(Condition(IsZero, &this->num_to_block_));
+
+  // Determine which thread can safely delete this Barrier object
+  this->num_to_exit_--;
+  ABSL_RAW_CHECK(this->num_to_exit_ >= 0, "barrier underflow");
+
+  // If num_to_exit_ == 0 then all other threads in the barrier have
+  // exited the Wait() and have released the Mutex so this thread is
+  // free to delete the barrier.
+  return this->num_to_exit_ == 0;
+}
+
+}  // namespace absl
diff --git a/absl/synchronization/barrier.h b/absl/synchronization/barrier.h
new file mode 100644
index 000000000000..f834feec11dc
--- /dev/null
+++ b/absl/synchronization/barrier.h
@@ -0,0 +1,77 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// -----------------------------------------------------------------------------
+// barrier.h
+// -----------------------------------------------------------------------------
+
+#ifndef ABSL_SYNCHRONIZATION_BARRIER_H_
+#define ABSL_SYNCHRONIZATION_BARRIER_H_
+
+#include "absl/base/thread_annotations.h"
+#include "absl/synchronization/mutex.h"
+
+namespace absl {
+
+// Barrier
+//
+// This class creates a barrier which blocks threads until a prespecified
+// threshold of threads (`num_threads`) utilizes the barrier. A thread utilizes
+// the `Barrier` by calling `Block()` on the barrier, which will block that
+// thread; no call to `Block()` will return until `num_threads` threads have
+// called it.
+//
+// Exactly one call to `Block()` will return `true`, which is then responsible
+// for destroying the barrier; because stack allocation will cause the barrier
+// to be deleted when it is out of scope, barriers should not be stack
+// allocated.
+//
+// Example:
+//
+//   // Main thread creates a `Barrier`:
+//   barrier = new Barrier(num_threads);
+//
+//   // Each participating thread could then call:
+//   if (barrier->Block()) delete barrier;  // Exactly one call to `Block()`
+//                                          // returns `true`; that call
+//                                          // deletes the barrier.
+class Barrier {
+ public:
+  // `num_threads` is the number of threads that will participate in the barrier
+  explicit Barrier(int num_threads)
+      : num_to_block_(num_threads), num_to_exit_(num_threads) {}
+
+  Barrier(const Barrier&) = delete;
+  Barrier& operator=(const Barrier&) = delete;
+
+  // Barrier::Block()
+  //
+  // Blocks the current thread, and returns only when the `num_threads`
+  // threshold of threads utilizing this barrier has been reached. `Block()`
+  // returns `true` for precisely one caller, which may then destroy the
+  // barrier.
+  //
+  // Memory ordering: For any threads X and Y, any action taken by X
+  // before X calls `Block()` will be visible to Y after Y returns from
+  // `Block()`.
+  bool Block();
+
+ private:
+  Mutex lock_;
+  int num_to_block_ GUARDED_BY(lock_);
+  int num_to_exit_ GUARDED_BY(lock_);
+};
+
+}  // namespace absl
+#endif  // ABSL_SYNCHRONIZATION_BARRIER_H_
diff --git a/absl/synchronization/blocking_counter.cc b/absl/synchronization/blocking_counter.cc
new file mode 100644
index 000000000000..48e3650da246
--- /dev/null
+++ b/absl/synchronization/blocking_counter.cc
@@ -0,0 +1,53 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/synchronization/blocking_counter.h"
+
+namespace absl {
+
+// Return whether int *arg is zero.
+static bool IsZero(void *arg) {
+  return 0 == *reinterpret_cast<int *>(arg);
+}
+
+bool BlockingCounter::DecrementCount() {
+  MutexLock l(&lock_);
+  count_--;
+  if (count_ < 0) {
+    ABSL_RAW_LOG(
+        FATAL,
+        "BlockingCounter::DecrementCount() called too many times.  count=%d",
+        count_);
+  }
+  return count_ == 0;
+}
+
+void BlockingCounter::Wait() {
+  MutexLock l(&this->lock_);
+  ABSL_RAW_CHECK(count_ >= 0, "BlockingCounter underflow");
+
+  // only one thread may call Wait(). To support more than one thread,
+  // implement a counter num_to_exit, like in the Barrier class.
+  ABSL_RAW_CHECK(num_waiting_ == 0, "multiple threads called Wait()");
+  num_waiting_++;
+
+  this->lock_.Await(Condition(IsZero, &this->count_));
+
+  // At this point, We know that all threads executing DecrementCount have
+  // released the lock, and so will not touch this object again.
+  // Therefore, the thread calling this method is free to delete the object
+  // after we return from this method.
+}
+
+}  // namespace absl
diff --git a/absl/synchronization/blocking_counter.h b/absl/synchronization/blocking_counter.h
new file mode 100644
index 000000000000..476d5f8f9885
--- /dev/null
+++ b/absl/synchronization/blocking_counter.h
@@ -0,0 +1,96 @@
+//
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// -----------------------------------------------------------------------------
+// blocking_counter.h
+// -----------------------------------------------------------------------------
+
+#ifndef ABSL_SYNCHRONIZATION_BLOCKING_COUNTER_H_
+#define ABSL_SYNCHRONIZATION_BLOCKING_COUNTER_H_
+
+#include "absl/base/thread_annotations.h"
+#include "absl/synchronization/mutex.h"
+
+namespace absl {
+
+// BlockingCounter
+//
+// This class allows a thread to block for a pre-specified number of actions.
+// `BlockingCounter` maintains a single non-negative abstract integer "count"
+// with an initial value `initial_count`. A thread can then call `Wait()` on
+// this blocking counter to block until the specified number of events occur;
+// worker threads then call 'DecrementCount()` on the counter upon completion of
+// their work. Once the counter's internal "count" reaches zero, the blocked
+// thread unblocks.
+//
+// A `BlockingCounter` requires the following:
+//     - its `initial_count` is non-negative.
+//     - the number of calls to `DecrementCount()` on it is at most
+//       `initial_count`.
+//     - `Wait()` is called at most once on it.
+//
+// Given the above requirements, a `BlockingCounter` provides the following
+// guarantees:
+//     - Once its internal "count" reaches zero, no legal action on the object
+//       can further change the value of "count".
+//     - When `Wait()` returns, it is legal to destroy the `BlockingCounter`.
+//     - When `Wait()` returns, the number of calls to `DecrementCount()` on
+//       this blocking counter exactly equals `initial_count`.
+//
+// Example:
+//     BlockingCounter bcount(N);         // there are N items of work
+//     ... Allow worker threads to start.
+//     ... On completing each work item, workers do:
+//     ... bcount.DecrementCount();      // an item of work has been completed
+//
+//     bcount.Wait();                    // wait for all work to be complete
+//
+class BlockingCounter {
+ public:
+  explicit BlockingCounter(int initial_count)
+      : count_(initial_count), num_waiting_(0) {}
+
+  BlockingCounter(const BlockingCounter&) = delete;
+  BlockingCounter& operator=(const BlockingCounter&) = delete;
+
+  // BlockingCounter::DecrementCount()
+  //
+  // Decrements the counter's "count" by one, and return "count == 0". This
+  // function requires that "count != 0" when it is called.
+  //
+  // Memory ordering: For any threads X and Y, any action taken by X
+  // before it calls `DecrementCount()` is visible to thread Y after
+  // Y's call to `DecrementCount()`, provided Y's call returns `true`.
+  bool DecrementCount();
+
+  // BlockingCounter::Wait()
+  //
+  // Blocks until the counter reaches zero. This function may be called at most
+  // once. On return, `DecrementCount()` will have been called "initial_count"
+  // times and the blocking counter may be destroyed.
+  //
+  // Memory ordering: For any threads X and Y, any action taken by X
+  // before X calls `DecrementCount()` is visible to Y after Y returns
+  // from `Wait()`.
+  void Wait();
+
+ private:
+  Mutex lock_;
+  int count_ GUARDED_BY(lock_);
+  int num_waiting_ GUARDED_BY(lock_);
+};
+
+}  // namespace absl
+#endif  // ABSL_SYNCHRONIZATION_BLOCKING_COUNTER_H_
diff --git a/absl/synchronization/blocking_counter_test.cc b/absl/synchronization/blocking_counter_test.cc
new file mode 100644
index 000000000000..b4b6677234e9
--- /dev/null
+++ b/absl/synchronization/blocking_counter_test.cc
@@ -0,0 +1,67 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/synchronization/blocking_counter.h"
+
+#include <functional>
+#include <memory>
+#include <thread>  // NOLINT(build/c++11)
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "absl/time/clock.h"
+
+namespace absl {
+namespace {
+
+void PauseAndDecreaseCounter(BlockingCounter* counter, int* done) {
+  absl::SleepFor(absl::Seconds(1));
+  *done = 1;
+  counter->DecrementCount();
+}
+
+TEST(BlockingCounterTest, BasicFunctionality) {
+  // This test verifies that BlockingCounter functions correctly. Starts a
+  // number of threads that just sleep for a second and decrement a counter.
+
+  // Initialize the counter.
+  const int num_workers = 10;
+  BlockingCounter counter(num_workers);
+
+  std::vector<std::thread> workers;
+  std::vector<int> done(num_workers, 0);
+
+  // Start a number of parallel tasks that will just wait for a seconds and
+  // then decrement the count.
+  workers.reserve(num_workers);
+  for (int k = 0; k < num_workers; k++) {
+    workers.emplace_back(
+        [&counter, &done, k] { PauseAndDecreaseCounter(&counter, &done[k]); });
+  }
+
+  // Wait for the threads to have all finished.
+  counter.Wait();
+
+  // Check that all the workers have completed.
+  for (int k = 0; k < num_workers; k++) {
+    EXPECT_EQ(1, done[k]);
+  }
+
+  for (std::thread& w : workers) {
+    w.join();
+  }
+}
+
+}  // namespace
+}  // namespace absl
diff --git a/absl/synchronization/internal/create_thread_identity.cc b/absl/synchronization/internal/create_thread_identity.cc
new file mode 100644
index 000000000000..14976347b671
--- /dev/null
+++ b/absl/synchronization/internal/create_thread_identity.cc
@@ -0,0 +1,110 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// This file is a no-op if the required LowLevelAlloc support is missing.
+#include "absl/base/internal/low_level_alloc.h"
+#ifndef ABSL_LOW_LEVEL_ALLOC_MISSING
+
+#include <string.h>
+#include <atomic>
+#include <memory>
+
+#include "absl/base/internal/spinlock.h"
+#include "absl/base/internal/thread_identity.h"
+#include "absl/synchronization/internal/per_thread_sem.h"
+
+namespace absl {
+namespace synchronization_internal {
+
+// ThreadIdentity storage is persistent, we maintain a free-list of previously
+// released ThreadIdentity objects.
+static base_internal::SpinLock freelist_lock(base_internal::kLinkerInitialized);
+static base_internal::ThreadIdentity* thread_identity_freelist;
+
+// A per-thread destructor for reclaiming associated ThreadIdentity objects.
+// Since we must preserve their storage we cache them for re-use.
+static void ReclaimThreadIdentity(void* v) {
+  base_internal::ThreadIdentity* identity =
+      static_cast<base_internal::ThreadIdentity*>(v);
+
+  // all_locks might have been allocated by the Mutex implementation.
+  // We free it here when we are notified that our thread is dying.
+  if (identity->per_thread_synch.all_locks != nullptr) {
+    base_internal::LowLevelAlloc::Free(identity->per_thread_synch.all_locks);
+  }
+
+  // We must explicitly clear the current thread's identity:
+  // (a) Subsequent (unrelated) per-thread destructors may require an identity.
+  //     We must guarantee a new identity is used in this case (this instructor
+  //     will be reinvoked up to PTHREAD_DESTRUCTOR_ITERATIONS in this case).
+  // (b) ThreadIdentity implementations may depend on memory that is not
+  //     reinitialized before reuse.  We must allow explicit clearing of the
+  //     association state in this case.
+  base_internal::ClearCurrentThreadIdentity();
+  {
+    base_internal::SpinLockHolder l(&freelist_lock);
+    identity->next = thread_identity_freelist;
+    thread_identity_freelist = identity;
+  }
+}
+
+// Return value rounded up to next multiple of align.
+// Align must be a power of two.
+static intptr_t RoundUp(intptr_t addr, intptr_t align) {
+  return (addr + align - 1) & ~(align - 1);
+}
+
+static base_internal::ThreadIdentity* NewThreadIdentity() {
+  base_internal::ThreadIdentity* identity = nullptr;
+
+  {
+    // Re-use a previously released object if possible.
+    base_internal::SpinLockHolder l(&freelist_lock);
+    if (thread_identity_freelist) {
+      identity = thread_identity_freelist;  // Take list-head.
+      thread_identity_freelist = thread_identity_freelist->next;
+    }
+  }
+
+  if (identity == nullptr) {
+    // Allocate enough space to align ThreadIdentity to a multiple of
+    // PerThreadSynch::kAlignment. This space is never released (it is
+    // added to a freelist by ReclaimThreadIdentity instead).
+    void* allocation = base_internal::LowLevelAlloc::Alloc(
+        sizeof(*identity) + base_internal::PerThreadSynch::kAlignment - 1);
+    // Round up the address to the required alignment.
+    identity = reinterpret_cast<base_internal::ThreadIdentity*>(
+        RoundUp(reinterpret_cast<intptr_t>(allocation),
+                base_internal::PerThreadSynch::kAlignment));
+  }
+  memset(identity, 0, sizeof(*identity));
+
+  return identity;
+}
+
+// Allocates and attaches ThreadIdentity object for the calling thread.  Returns
+// the new identity.
+// REQUIRES: CurrentThreadIdentity(false) == nullptr
+base_internal::ThreadIdentity* CreateThreadIdentity() {
+  base_internal::ThreadIdentity* identity = NewThreadIdentity();
+  PerThreadSem::Init(identity);
+  // Associate the value with the current thread, and attach our destructor.
+  base_internal::SetCurrentThreadIdentity(identity, ReclaimThreadIdentity);
+  return identity;
+}
+
+}  // namespace synchronization_internal
+}  // namespace absl
+
+#endif  // ABSL_LOW_LEVEL_ALLOC_MISSING
diff --git a/absl/synchronization/internal/create_thread_identity.h b/absl/synchronization/internal/create_thread_identity.h
new file mode 100644
index 000000000000..1bb87dee6307
--- /dev/null
+++ b/absl/synchronization/internal/create_thread_identity.h
@@ -0,0 +1,53 @@
+/*
+ * Copyright 2017 The Abseil Authors.
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+// Interface for getting the current ThreadIdentity, creating one if necessary.
+// See thread_identity.h.
+//
+// This file is separate from thread_identity.h because creating a new
+// ThreadIdentity requires slightly higher level libraries (per_thread_sem
+// and low_level_alloc) than accessing an existing one.  This separation allows
+// us to have a smaller //absl/base:base.
+
+#ifndef ABSL_SYNCHRONIZATION_INTERNAL_CREATE_THREAD_IDENTITY_H_
+#define ABSL_SYNCHRONIZATION_INTERNAL_CREATE_THREAD_IDENTITY_H_
+
+#include "absl/base/internal/thread_identity.h"
+#include "absl/base/port.h"
+
+namespace absl {
+namespace synchronization_internal {
+
+// Allocates and attaches a ThreadIdentity object for the calling thread.
+// For private use only.
+base_internal::ThreadIdentity* CreateThreadIdentity();
+
+// Returns the ThreadIdentity object representing the calling thread; guaranteed
+// to be unique for its lifetime.  The returned object will remain valid for the
+// program's lifetime; although it may be re-assigned to a subsequent thread.
+// If one does not exist for the calling thread, allocate it now.
+inline base_internal::ThreadIdentity* GetOrCreateCurrentThreadIdentity() {
+  base_internal::ThreadIdentity* identity =
+      base_internal::CurrentThreadIdentityIfPresent();
+  if (ABSL_PREDICT_FALSE(identity == nullptr)) {
+    return CreateThreadIdentity();
+  }
+  return identity;
+}
+
+}  // namespace synchronization_internal
+}  // namespace absl
+#endif  // ABSL_SYNCHRONIZATION_INTERNAL_CREATE_THREAD_IDENTITY_H_
diff --git a/absl/synchronization/internal/graphcycles.cc b/absl/synchronization/internal/graphcycles.cc
new file mode 100644
index 000000000000..d7ae0cf320d7
--- /dev/null
+++ b/absl/synchronization/internal/graphcycles.cc
@@ -0,0 +1,709 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// GraphCycles provides incremental cycle detection on a dynamic
+// graph using the following algorithm:
+//
+// A dynamic topological sort algorithm for directed acyclic graphs
+// David J. Pearce, Paul H. J. Kelly
+// Journal of Experimental Algorithmics (JEA) JEA Homepage archive
+// Volume 11, 2006, Article No. 1.7
+//
+// Brief summary of the algorithm:
+//
+// (1) Maintain a rank for each node that is consistent
+//     with the topological sort of the graph. I.e., path from x to y
+//     implies rank[x] < rank[y].
+// (2) When a new edge (x->y) is inserted, do nothing if rank[x] < rank[y].
+// (3) Otherwise: adjust ranks in the neighborhood of x and y.
+
+// This file is a no-op if the required LowLevelAlloc support is missing.
+#include "absl/base/internal/low_level_alloc.h"
+#ifndef ABSL_LOW_LEVEL_ALLOC_MISSING
+
+#include "absl/synchronization/internal/graphcycles.h"
+
+#include <algorithm>
+#include <array>
+#include "absl/base/internal/raw_logging.h"
+#include "absl/base/internal/spinlock.h"
+
+// Do not use STL.   This module does not use standard memory allocation.
+
+namespace absl {
+namespace synchronization_internal {
+
+namespace {
+
+// Avoid LowLevelAlloc's default arena since it calls malloc hooks in
+// which people are doing things like acquiring Mutexes.
+static absl::base_internal::SpinLock arena_mu(
+    absl::base_internal::kLinkerInitialized);
+static base_internal::LowLevelAlloc::Arena* arena;
+
+static void InitArenaIfNecessary() {
+  arena_mu.Lock();
+  if (arena == nullptr) {
+    arena = base_internal::LowLevelAlloc::NewArena(
+        0, base_internal::LowLevelAlloc::DefaultArena());
+  }
+  arena_mu.Unlock();
+}
+
+// Number of inlined elements in Vec.  Hash table implementation
+// relies on this being a power of two.
+static const uint32_t kInline = 8;
+
+// A simple LowLevelAlloc based resizable vector with inlined storage
+// for a few elements.  T must be a plain type since constructor
+// and destructor are not run on elements of type T managed by Vec.
+template <typename T>
+class Vec {
+ public:
+  Vec() { Init(); }
+  ~Vec() { Discard(); }
+
+  void clear() {
+    Discard();
+    Init();
+  }
+
+  bool empty() const { return size_ == 0; }
+  uint32_t size() const { return size_; }
+  T* begin() { return ptr_; }
+  T* end() { return ptr_ + size_; }
+  const T& operator[](uint32_t i) const { return ptr_[i]; }
+  T& operator[](uint32_t i) { return ptr_[i]; }
+  const T& back() const { return ptr_[size_-1]; }
+  void pop_back() { size_--; }
+
+  void push_back(const T& v) {
+    if (size_ == capacity_) Grow(size_ + 1);
+    ptr_[size_] = v;
+    size_++;
+  }
+
+  void resize(uint32_t n) {
+    if (n > capacity_) Grow(n);
+    size_ = n;
+  }
+
+  void fill(const T& val) {
+    for (uint32_t i = 0; i < size(); i++) {
+      ptr_[i] = val;
+    }
+  }
+
+  // Guarantees src is empty at end.
+  // Provided for the hash table resizing code below.
+  void MoveFrom(Vec<T>* src) {
+    if (src->ptr_ == src->space_) {
+      // Need to actually copy
+      resize(src->size_);
+      std::copy(src->ptr_, src->ptr_ + src->size_, ptr_);
+      src->size_ = 0;
+    } else {
+      Discard();
+      ptr_ = src->ptr_;
+      size_ = src->size_;
+      capacity_ = src->capacity_;
+      src->Init();
+    }
+  }
+
+ private:
+  T* ptr_;
+  T space_[kInline];
+  uint32_t size_;
+  uint32_t capacity_;
+
+  void Init() {
+    ptr_ = space_;
+    size_ = 0;
+    capacity_ = kInline;
+  }
+
+  void Discard() {
+    if (ptr_ != space_) base_internal::LowLevelAlloc::Free(ptr_);
+  }
+
+  void Grow(uint32_t n) {
+    while (capacity_ < n) {
+      capacity_ *= 2;
+    }
+    size_t request = static_cast<size_t>(capacity_) * sizeof(T);
+    T* copy = static_cast<T*>(
+        base_internal::LowLevelAlloc::AllocWithArena(request, arena));
+    std::copy(ptr_, ptr_ + size_, copy);
+    Discard();
+    ptr_ = copy;
+  }
+
+  Vec(const Vec&) = delete;
+  Vec& operator=(const Vec&) = delete;
+};
+
+// A hash set of non-negative int32_t that uses Vec for its underlying storage.
+class NodeSet {
+ public:
+  NodeSet() { Init(); }
+
+  void clear() { Init(); }
+  bool contains(int32_t v) const { return table_[FindIndex(v)] == v; }
+
+  bool insert(int32_t v) {
+    uint32_t i = FindIndex(v);
+    if (table_[i] == v) {
+      return false;
+    }
+    if (table_[i] == kEmpty) {
+      // Only inserting over an empty cell increases the number of occupied
+      // slots.
+      occupied_++;
+    }
+    table_[i] = v;
+    // Double when 75% full.
+    if (occupied_ >= table_.size() - table_.size()/4) Grow();
+    return true;
+  }
+
+  void erase(uint32_t v) {
+    uint32_t i = FindIndex(v);
+    if (static_cast<uint32_t>(table_[i]) == v) {
+      table_[i] = kDel;
+    }
+  }
+
+  // Iteration: is done via HASH_FOR_EACH
+  // Example:
+  //    HASH_FOR_EACH(elem, node->out) { ... }
+#define HASH_FOR_EACH(elem, eset) \
+  for (int32_t elem, _cursor = 0; (eset).Next(&_cursor, &elem); )
+  bool Next(int32_t* cursor, int32_t* elem) {
+    while (static_cast<uint32_t>(*cursor) < table_.size()) {
+      int32_t v = table_[*cursor];
+      (*cursor)++;
+      if (v >= 0) {
+        *elem = v;
+        return true;
+      }
+    }
+    return false;
+  }
+
+ private:
+  static const int32_t kEmpty;
+  static const int32_t kDel;
+  Vec<int32_t> table_;
+  uint32_t occupied_;     // Count of non-empty slots (includes deleted slots)
+
+  static uint32_t Hash(uint32_t a) { return a * 41; }
+
+  // Return index for storing v.  May return an empty index or deleted index
+  int FindIndex(int32_t v) const {
+    // Search starting at hash index.
+    const uint32_t mask = table_.size() - 1;
+    uint32_t i = Hash(v) & mask;
+    int deleted_index = -1;  // If >= 0, index of first deleted element we see
+    while (true) {
+      int32_t e = table_[i];
+      if (v == e) {
+        return i;
+      } else if (e == kEmpty) {
+        // Return any previously encountered deleted slot.
+        return (deleted_index >= 0) ? deleted_index : i;
+      } else if (e == kDel && deleted_index < 0) {
+        // Keep searching since v might be present later.
+        deleted_index = i;
+      }
+      i = (i + 1) & mask;  // Linear probing; quadratic is slightly slower.
+    }
+  }
+
+  void Init() {
+    table_.clear();
+    table_.resize(kInline);
+    table_.fill(kEmpty);
+    occupied_ = 0;
+  }
+
+  void Grow() {
+    Vec<int32_t> copy;
+    copy.MoveFrom(&table_);
+    occupied_ = 0;
+    table_.resize(copy.size() * 2);
+    table_.fill(kEmpty);
+
+    for (const auto& e : copy) {
+      if (e >= 0) insert(e);
+    }
+  }
+
+  NodeSet(const NodeSet&) = delete;
+  NodeSet& operator=(const NodeSet&) = delete;
+};
+
+const int32_t NodeSet::kEmpty = -1;
+const int32_t NodeSet::kDel = -2;
+
+// We encode a node index and a node version in GraphId.  The version
+// number is incremented when the GraphId is freed which automatically
+// invalidates all copies of the GraphId.
+
+inline GraphId MakeId(int32_t index, uint32_t version) {
+  GraphId g;
+  g.handle =
+      (static_cast<uint64_t>(version) << 32) | static_cast<uint32_t>(index);
+  return g;
+}
+
+inline int32_t NodeIndex(GraphId id) {
+  return static_cast<uint32_t>(id.handle & 0xfffffffful);
+}
+
+inline uint32_t NodeVersion(GraphId id) {
+  return static_cast<uint32_t>(id.handle >> 32);
+}
+
+// We need to hide Mutexes (or other deadlock detection's pointers)
+// from the leak detector.  Xor with an arbitrary number with high bits set.
+static const uintptr_t kHideMask = static_cast<uintptr_t>(0xF03A5F7BF03A5F7Bll);
+
+static inline uintptr_t MaskPtr(void *ptr) {
+  return reinterpret_cast<uintptr_t>(ptr) ^ kHideMask;
+}
+
+static inline void* UnmaskPtr(uintptr_t word) {
+  return reinterpret_cast<void*>(word ^ kHideMask);
+}
+
+struct Node {
+  int32_t rank;               // rank number assigned by Pearce-Kelly algorithm
+  uint32_t version;           // Current version number
+  int32_t next_hash;          // Next entry in hash table
+  bool visited;               // Temporary marker used by depth-first-search
+  uintptr_t masked_ptr;       // User-supplied pointer
+  NodeSet in;                 // List of immediate predecessor nodes in graph
+  NodeSet out;                // List of immediate successor nodes in graph
+  int priority;               // Priority of recorded stack trace.
+  int nstack;                 // Depth of recorded stack trace.
+  void* stack[40];            // stack[0,nstack-1] holds stack trace for node.
+};
+
+// Hash table for pointer to node index lookups.
+class PointerMap {
+ public:
+  explicit PointerMap(const Vec<Node*>* nodes) : nodes_(nodes) {
+    table_.fill(-1);
+  }
+
+  int32_t Find(void* ptr) {
+    auto masked = MaskPtr(ptr);
+    for (int32_t i = table_[Hash(ptr)]; i != -1;) {
+      Node* n = (*nodes_)[i];
+      if (n->masked_ptr == masked) return i;
+      i = n->next_hash;
+    }
+    return -1;
+  }
+
+  void Add(void* ptr, int32_t i) {
+    int32_t* head = &table_[Hash(ptr)];
+    (*nodes_)[i]->next_hash = *head;
+    *head = i;
+  }
+
+  int32_t Remove(void* ptr) {
+    // Advance through linked list while keeping track of the
+    // predecessor slot that points to the current entry.
+    auto masked = MaskPtr(ptr);
+    for (int32_t* slot = &table_[Hash(ptr)]; *slot != -1; ) {
+      int32_t index = *slot;
+      Node* n = (*nodes_)[index];
+      if (n->masked_ptr == masked) {
+        *slot = n->next_hash;  // Remove n from linked list
+        n->next_hash = -1;
+        return index;
+      }
+      slot = &n->next_hash;
+    }
+    return -1;
+  }
+
+ private:
+  // Number of buckets in hash table for pointer lookups.
+  static constexpr uint32_t kHashTableSize = 8171;  // should be prime
+
+  const Vec<Node*>* nodes_;
+  std::array<int32_t, kHashTableSize> table_;
+
+  static uint32_t Hash(void* ptr) {
+    return reinterpret_cast<uintptr_t>(ptr) % kHashTableSize;
+  }
+};
+
+}  // namespace
+
+struct GraphCycles::Rep {
+  Vec<Node*> nodes_;
+  Vec<int32_t> free_nodes_;  // Indices for unused entries in nodes_
+  PointerMap ptrmap_;
+
+  // Temporary state.
+  Vec<int32_t> deltaf_;  // Results of forward DFS
+  Vec<int32_t> deltab_;  // Results of backward DFS
+  Vec<int32_t> list_;    // All nodes to reprocess
+  Vec<int32_t> merged_;  // Rank values to assign to list_ entries
+  Vec<int32_t> stack_;   // Emulates recursion stack for depth-first searches
+
+  Rep() : ptrmap_(&nodes_) {}
+};
+
+static Node* FindNode(GraphCycles::Rep* rep, GraphId id) {
+  Node* n = rep->nodes_[NodeIndex(id)];
+  return (n->version == NodeVersion(id)) ? n : nullptr;
+}
+
+GraphCycles::GraphCycles() {
+  InitArenaIfNecessary();
+  rep_ = new (base_internal::LowLevelAlloc::AllocWithArena(sizeof(Rep), arena))
+      Rep;
+}
+
+GraphCycles::~GraphCycles() {
+  for (auto* node : rep_->nodes_) {
+    node->Node::~Node();
+    base_internal::LowLevelAlloc::Free(node);
+  }
+  rep_->Rep::~Rep();
+  base_internal::LowLevelAlloc::Free(rep_);
+}
+
+bool GraphCycles::CheckInvariants() const {
+  Rep* r = rep_;
+  NodeSet ranks;  // Set of ranks seen so far.
+  for (uint32_t x = 0; x < r->nodes_.size(); x++) {
+    Node* nx = r->nodes_[x];
+    void* ptr = UnmaskPtr(nx->masked_ptr);
+    if (ptr != nullptr && static_cast<uint32_t>(r->ptrmap_.Find(ptr)) != x) {
+      ABSL_RAW_LOG(FATAL, "Did not find live node in hash table %u %p", x, ptr);
+    }
+    if (nx->visited) {
+      ABSL_RAW_LOG(FATAL, "Did not clear visited marker on node %u", x);
+    }
+    if (!ranks.insert(nx->rank)) {
+      ABSL_RAW_LOG(FATAL, "Duplicate occurrence of rank %d", nx->rank);
+    }
+    HASH_FOR_EACH(y, nx->out) {
+      Node* ny = r->nodes_[y];
+      if (nx->rank >= ny->rank) {
+        ABSL_RAW_LOG(FATAL, "Edge %u->%d has bad rank assignment %d->%d", x, y,
+                     nx->rank, ny->rank);
+      }
+    }
+  }
+  return true;
+}
+
+GraphId GraphCycles::GetId(void* ptr) {
+  int32_t i = rep_->ptrmap_.Find(ptr);
+  if (i != -1) {
+    return MakeId(i, rep_->nodes_[i]->version);
+  } else if (rep_->free_nodes_.empty()) {
+    Node* n =
+        new (base_internal::LowLevelAlloc::AllocWithArena(sizeof(Node), arena))
+            Node;
+    n->version = 1;  // Avoid 0 since it is used by InvalidGraphId()
+    n->visited = false;
+    n->rank = rep_->nodes_.size();
+    n->masked_ptr = MaskPtr(ptr);
+    n->nstack = 0;
+    n->priority = 0;
+    rep_->nodes_.push_back(n);
+    rep_->ptrmap_.Add(ptr, n->rank);
+    return MakeId(n->rank, n->version);
+  } else {
+    // Preserve preceding rank since the set of ranks in use must be
+    // a permutation of [0,rep_->nodes_.size()-1].
+    int32_t r = rep_->free_nodes_.back();
+    rep_->free_nodes_.pop_back();
+    Node* n = rep_->nodes_[r];
+    n->masked_ptr = MaskPtr(ptr);
+    n->nstack = 0;
+    n->priority = 0;
+    rep_->ptrmap_.Add(ptr, r);
+    return MakeId(r, n->version);
+  }
+}
+
+void GraphCycles::RemoveNode(void* ptr) {
+  int32_t i = rep_->ptrmap_.Remove(ptr);
+  if (i == -1) {
+    return;
+  }
+  Node* x = rep_->nodes_[i];
+  HASH_FOR_EACH(y, x->out) {
+    rep_->nodes_[y]->in.erase(i);
+  }
+  HASH_FOR_EACH(y, x->in) {
+    rep_->nodes_[y]->out.erase(i);
+  }
+  x->in.clear();
+  x->out.clear();
+  x->masked_ptr = MaskPtr(nullptr);
+  if (x->version == std::numeric_limits<uint32_t>::max()) {
+    // Cannot use x any more
+  } else {
+    x->version++;  // Invalidates all copies of node.
+    rep_->free_nodes_.push_back(i);
+  }
+}
+
+void* GraphCycles::Ptr(GraphId id) {
+  Node* n = FindNode(rep_, id);
+  return n == nullptr ? nullptr : UnmaskPtr(n->masked_ptr);
+}
+
+bool GraphCycles::HasNode(GraphId node) {
+  return FindNode(rep_, node) != nullptr;
+}
+
+bool GraphCycles::HasEdge(GraphId x, GraphId y) const {
+  Node* xn = FindNode(rep_, x);
+  return xn && FindNode(rep_, y) && xn->out.contains(NodeIndex(y));
+}
+
+void GraphCycles::RemoveEdge(GraphId x, GraphId y) {
+  Node* xn = FindNode(rep_, x);
+  Node* yn = FindNode(rep_, y);
+  if (xn && yn) {
+    xn->out.erase(NodeIndex(y));
+    yn->in.erase(NodeIndex(x));
+    // No need to update the rank assignment since a previous valid
+    // rank assignment remains valid after an edge deletion.
+  }
+}
+
+static bool ForwardDFS(GraphCycles::Rep* r, int32_t n, int32_t upper_bound);
+static void BackwardDFS(GraphCycles::Rep* r, int32_t n, int32_t lower_bound);
+static void Reorder(GraphCycles::Rep* r);
+static void Sort(const Vec<Node*>&, Vec<int32_t>* delta);
+static void MoveToList(
+    GraphCycles::Rep* r, Vec<int32_t>* src, Vec<int32_t>* dst);
+
+bool GraphCycles::InsertEdge(GraphId idx, GraphId idy) {
+  Rep* r = rep_;
+  const int32_t x = NodeIndex(idx);
+  const int32_t y = NodeIndex(idy);
+  Node* nx = FindNode(r, idx);
+  Node* ny = FindNode(r, idy);
+  if (nx == nullptr || ny == nullptr) return true;  // Expired ids
+
+  if (nx == ny) return false;  // Self edge
+  if (!nx->out.insert(y)) {
+    // Edge already exists.
+    return true;
+  }
+
+  ny->in.insert(x);
+
+  if (nx->rank <= ny->rank) {
+    // New edge is consistent with existing rank assignment.
+    return true;
+  }
+
+  // Current rank assignments are incompatible with the new edge.  Recompute.
+  // We only need to consider nodes that fall in the range [ny->rank,nx->rank].
+  if (!ForwardDFS(r, y, nx->rank)) {
+    // Found a cycle.  Undo the insertion and tell caller.
+    nx->out.erase(y);
+    ny->in.erase(x);
+    // Since we do not call Reorder() on this path, clear any visited
+    // markers left by ForwardDFS.
+    for (const auto& d : r->deltaf_) {
+      r->nodes_[d]->visited = false;
+    }
+    return false;
+  }
+  BackwardDFS(r, x, ny->rank);
+  Reorder(r);
+  return true;
+}
+
+static bool ForwardDFS(GraphCycles::Rep* r, int32_t n, int32_t upper_bound) {
+  // Avoid recursion since stack space might be limited.
+  // We instead keep a stack of nodes to visit.
+  r->deltaf_.clear();
+  r->stack_.clear();
+  r->stack_.push_back(n);
+  while (!r->stack_.empty()) {
+    n = r->stack_.back();
+    r->stack_.pop_back();
+    Node* nn = r->nodes_[n];
+    if (nn->visited) continue;
+
+    nn->visited = true;
+    r->deltaf_.push_back(n);
+
+    HASH_FOR_EACH(w, nn->out) {
+      Node* nw = r->nodes_[w];
+      if (nw->rank == upper_bound) {
+        return false;  // Cycle
+      }
+      if (!nw->visited && nw->rank < upper_bound) {
+        r->stack_.push_back(w);
+      }
+    }
+  }
+  return true;
+}
+
+static void BackwardDFS(GraphCycles::Rep* r, int32_t n, int32_t lower_bound) {
+  r->deltab_.clear();
+  r->stack_.clear();
+  r->stack_.push_back(n);
+  while (!r->stack_.empty()) {
+    n = r->stack_.back();
+    r->stack_.pop_back();
+    Node* nn = r->nodes_[n];
+    if (nn->visited) continue;
+
+    nn->visited = true;
+    r->deltab_.push_back(n);
+
+    HASH_FOR_EACH(w, nn->in) {
+      Node* nw = r->nodes_[w];
+      if (!nw->visited && lower_bound < nw->rank) {
+        r->stack_.push_back(w);
+      }
+    }
+  }
+}
+
+static void Reorder(GraphCycles::Rep* r) {
+  Sort(r->nodes_, &r->deltab_);
+  Sort(r->nodes_, &r->deltaf_);
+
+  // Adds contents of delta lists to list_ (backwards deltas first).
+  r->list_.clear();
+  MoveToList(r, &r->deltab_, &r->list_);
+  MoveToList(r, &r->deltaf_, &r->list_);
+
+  // Produce sorted list of all ranks that will be reassigned.
+  r->merged_.resize(r->deltab_.size() + r->deltaf_.size());
+  std::merge(r->deltab_.begin(), r->deltab_.end(),
+             r->deltaf_.begin(), r->deltaf_.end(),
+             r->merged_.begin());
+
+  // Assign the ranks in order to the collected list.
+  for (uint32_t i = 0; i < r->list_.size(); i++) {
+    r->nodes_[r->list_[i]]->rank = r->merged_[i];
+  }
+}
+
+static void Sort(const Vec<Node*>& nodes, Vec<int32_t>* delta) {
+  struct ByRank {
+    const Vec<Node*>* nodes;
+    bool operator()(int32_t a, int32_t b) const {
+      return (*nodes)[a]->rank < (*nodes)[b]->rank;
+    }
+  };
+  ByRank cmp;
+  cmp.nodes = &nodes;
+  std::sort(delta->begin(), delta->end(), cmp);
+}
+
+static void MoveToList(
+    GraphCycles::Rep* r, Vec<int32_t>* src, Vec<int32_t>* dst) {
+  for (auto& v : *src) {
+    int32_t w = v;
+    v = r->nodes_[w]->rank;         // Replace v entry with its rank
+    r->nodes_[w]->visited = false;  // Prepare for future DFS calls
+    dst->push_back(w);
+  }
+}
+
+int GraphCycles::FindPath(GraphId idx, GraphId idy, int max_path_len,
+                          GraphId path[]) const {
+  Rep* r = rep_;
+  if (FindNode(r, idx) == nullptr || FindNode(r, idy) == nullptr) return 0;
+  const int32_t x = NodeIndex(idx);
+  const int32_t y = NodeIndex(idy);
+
+  // Forward depth first search starting at x until we hit y.
+  // As we descend into a node, we push it onto the path.
+  // As we leave a node, we remove it from the path.
+  int path_len = 0;
+
+  NodeSet seen;
+  r->stack_.clear();
+  r->stack_.push_back(x);
+  while (!r->stack_.empty()) {
+    int32_t n = r->stack_.back();
+    r->stack_.pop_back();
+    if (n < 0) {
+      // Marker to indicate that we are leaving a node
+      path_len--;
+      continue;
+    }
+
+    if (path_len < max_path_len) {
+      path[path_len] = MakeId(n, rep_->nodes_[n]->version);
+    }
+    path_len++;
+    r->stack_.push_back(-1);  // Will remove tentative path entry
+
+    if (n == y) {
+      return path_len;
+    }
+
+    HASH_FOR_EACH(w, r->nodes_[n]->out) {
+      if (seen.insert(w)) {
+        r->stack_.push_back(w);
+      }
+    }
+  }
+
+  return 0;
+}
+
+bool GraphCycles::IsReachable(GraphId x, GraphId y) const {
+  return FindPath(x, y, 0, nullptr) > 0;
+}
+
+void GraphCycles::UpdateStackTrace(GraphId id, int priority,
+                                   int (*get_stack_trace)(void** stack, int)) {
+  Node* n = FindNode(rep_, id);
+  if (n == nullptr || n->priority >= priority) {
+    return;
+  }
+  n->nstack = (*get_stack_trace)(n->stack, ABSL_ARRAYSIZE(n->stack));
+  n->priority = priority;
+}
+
+int GraphCycles::GetStackTrace(GraphId id, void*** ptr) {
+  Node* n = FindNode(rep_, id);
+  if (n == nullptr) {
+    *ptr = nullptr;
+    return 0;
+  } else {
+    *ptr = n->stack;
+    return n->nstack;
+  }
+}
+
+}  // namespace synchronization_internal
+}  // namespace absl
+
+#endif  // ABSL_LOW_LEVEL_ALLOC_MISSING
diff --git a/absl/synchronization/internal/graphcycles.h b/absl/synchronization/internal/graphcycles.h
new file mode 100644
index 000000000000..53474b7b0b53
--- /dev/null
+++ b/absl/synchronization/internal/graphcycles.h
@@ -0,0 +1,136 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+
+#ifndef ABSL_SYNCHRONIZATION_INTERNAL_GRAPHCYCLES_H_
+#define ABSL_SYNCHRONIZATION_INTERNAL_GRAPHCYCLES_H_
+
+// GraphCycles detects the introduction of a cycle into a directed
+// graph that is being built up incrementally.
+//
+// Nodes are identified by small integers.  It is not possible to
+// record multiple edges with the same (source, destination) pair;
+// requests to add an edge where one already exists are silently
+// ignored.
+//
+// It is also not possible to introduce a cycle; an attempt to insert
+// an edge that would introduce a cycle fails and returns false.
+//
+// GraphCycles uses no internal locking; calls into it should be
+// serialized externally.
+
+// Performance considerations:
+//   Works well on sparse graphs, poorly on dense graphs.
+//   Extra information is maintained incrementally to detect cycles quickly.
+//   InsertEdge() is very fast when the edge already exists, and reasonably fast
+//   otherwise.
+//   FindPath() is linear in the size of the graph.
+// The current implemenation uses O(|V|+|E|) space.
+
+#include <cstdint>
+
+namespace absl {
+namespace synchronization_internal {
+
+// Opaque identifier for a graph node.
+struct GraphId {
+  uint64_t handle;
+
+  bool operator==(const GraphId& x) const { return handle == x.handle; }
+  bool operator!=(const GraphId& x) const { return handle != x.handle; }
+};
+
+// Return an invalid graph id that will never be assigned by GraphCycles.
+inline GraphId InvalidGraphId() {
+  return GraphId{0};
+}
+
+class GraphCycles {
+ public:
+  GraphCycles();
+  ~GraphCycles();
+
+  // Return the id to use for ptr, assigning one if necessary.
+  // Subsequent calls with the same ptr value will return the same id
+  // until Remove().
+  GraphId GetId(void* ptr);
+
+  // Remove "ptr" from the graph.  Its corresponding node and all
+  // edges to and from it are removed.
+  void RemoveNode(void* ptr);
+
+  // Return the pointer associated with id, or nullptr if id is not
+  // currently in the graph.
+  void* Ptr(GraphId id);
+
+  // Attempt to insert an edge from source_node to dest_node.  If the
+  // edge would introduce a cycle, return false without making any
+  // changes. Otherwise add the edge and return true.
+  bool InsertEdge(GraphId source_node, GraphId dest_node);
+
+  // Remove any edge that exists from source_node to dest_node.
+  void RemoveEdge(GraphId source_node, GraphId dest_node);
+
+  // Return whether node exists in the graph.
+  bool HasNode(GraphId node);
+
+  // Return whether there is an edge directly from source_node to dest_node.
+  bool HasEdge(GraphId source_node, GraphId dest_node) const;
+
+  // Return whether dest_node is reachable from source_node
+  // by following edges.
+  bool IsReachable(GraphId source_node, GraphId dest_node) const;
+
+  // Find a path from "source" to "dest".  If such a path exists,
+  // place the nodes on the path in the array path[], and return
+  // the number of nodes on the path.  If the path is longer than
+  // max_path_len nodes, only the first max_path_len nodes are placed
+  // in path[].  The client should compare the return value with
+  // max_path_len" to see when this occurs.  If no path exists, return
+  // 0.  Any valid path stored in path[] will start with "source" and
+  // end with "dest".  There is no guarantee that the path is the
+  // shortest, but no node will appear twice in the path, except the
+  // source and destination node if they are identical; therefore, the
+  // return value is at most one greater than the number of nodes in
+  // the graph.
+  int FindPath(GraphId source, GraphId dest, int max_path_len,
+               GraphId path[]) const;
+
+  // Update the stack trace recorded for id with the current stack
+  // trace if the last time it was updated had a smaller priority
+  // than the priority passed on this call.
+  //
+  // *get_stack_trace is called to get the stack trace.
+  void UpdateStackTrace(GraphId id, int priority,
+                        int (*get_stack_trace)(void**, int));
+
+  // Set *ptr to the beginning of the array that holds the recorded
+  // stack trace for id and return the depth of the stack trace.
+  int GetStackTrace(GraphId id, void*** ptr);
+
+  // Check internal invariants. Crashes on failure, returns true on success.
+  // Expensive: should only be called from graphcycles_test.cc.
+  bool CheckInvariants() const;
+
+  // ----------------------------------------------------
+  struct Rep;
+ private:
+  Rep *rep_;      // opaque representation
+  GraphCycles(const GraphCycles&) = delete;
+  GraphCycles& operator=(const GraphCycles&) = delete;
+};
+
+}  // namespace synchronization_internal
+}  // namespace absl
+#endif
diff --git a/absl/synchronization/internal/graphcycles_test.cc b/absl/synchronization/internal/graphcycles_test.cc
new file mode 100644
index 000000000000..734f2770b26e
--- /dev/null
+++ b/absl/synchronization/internal/graphcycles_test.cc
@@ -0,0 +1,471 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Copyright 2007 Google, Inc.
+// All rights reserved.
+
+// Author: Mike Burrows
+
+// A test for the GraphCycles interface.
+
+// This test is testing a component of //third_party/absl.  As written it
+// heavily uses logging, including VLOG, so this test can't ship with Abseil.
+// We're leaving it here until Abseil gets base/logging.h in a future release.
+#include "absl/synchronization/internal/graphcycles.h"
+
+#include <map>
+#include <random>
+#include <vector>
+#include <unordered_set>
+
+#include "gtest/gtest.h"
+#include "absl/base/internal/raw_logging.h"
+#include "absl/base/macros.h"
+
+namespace absl {
+namespace synchronization_internal {
+
+// We emulate a GraphCycles object with a node vector and an edge vector.
+// We then compare the two implementations.
+
+using Nodes = std::vector<int>;
+struct Edge {
+  int from;
+  int to;
+};
+using Edges = std::vector<Edge>;
+using RandomEngine = std::mt19937_64;
+
+// Mapping from integer index to GraphId.
+typedef std::map<int, GraphId> IdMap;
+static GraphId Get(const IdMap& id, int num) {
+  auto iter = id.find(num);
+  return (iter == id.end()) ? InvalidGraphId() : iter->second;
+}
+
+// Return whether "to" is reachable from "from".
+static bool IsReachable(Edges *edges, int from, int to,
+                        std::unordered_set<int> *seen) {
+  seen->insert(from);     // we are investigating "from"; don't do it again
+  if (from == to) return true;
+  for (const auto &edge : *edges) {
+    if (edge.from == from) {
+      if (edge.to == to) {  // success via edge directly
+        return true;
+      } else if (seen->find(edge.to) == seen->end() &&  // success via edge
+                 IsReachable(edges, edge.to, to, seen)) {
+        return true;
+      }
+    }
+  }
+  return false;
+}
+
+static void PrintEdges(Edges *edges) {
+  ABSL_RAW_LOG(INFO, "EDGES (%zu)", edges->size());
+  for (const auto &edge : *edges) {
+    int a = edge.from;
+    int b = edge.to;
+    ABSL_RAW_LOG(INFO, "%d %d", a, b);
+  }
+  ABSL_RAW_LOG(INFO, "---");
+}
+
+static void PrintGCEdges(Nodes *nodes, const IdMap &id, GraphCycles *gc) {
+  ABSL_RAW_LOG(INFO, "GC EDGES");
+  for (int a : *nodes) {
+    for (int b : *nodes) {
+      if (gc->HasEdge(Get(id, a), Get(id, b))) {
+        ABSL_RAW_LOG(INFO, "%d %d", a, b);
+      }
+    }
+  }
+  ABSL_RAW_LOG(INFO, "---");
+}
+
+static void PrintTransitiveClosure(Nodes *nodes, Edges *edges) {
+  ABSL_RAW_LOG(INFO, "Transitive closure");
+  for (int a : *nodes) {
+    for (int b : *nodes) {
+      std::unordered_set<int> seen;
+      if (IsReachable(edges, a, b, &seen)) {
+        ABSL_RAW_LOG(INFO, "%d %d", a, b);
+      }
+    }
+  }
+  ABSL_RAW_LOG(INFO, "---");
+}
+
+static void PrintGCTransitiveClosure(Nodes *nodes, const IdMap &id,
+                                     GraphCycles *gc) {
+  ABSL_RAW_LOG(INFO, "GC Transitive closure");
+  for (int a : *nodes) {
+    for (int b : *nodes) {
+      if (gc->IsReachable(Get(id, a), Get(id, b))) {
+        ABSL_RAW_LOG(INFO, "%d %d", a, b);
+      }
+    }
+  }
+  ABSL_RAW_LOG(INFO, "---");
+}
+
+static void CheckTransitiveClosure(Nodes *nodes, Edges *edges, const IdMap &id,
+                                   GraphCycles *gc) {
+  std::unordered_set<int> seen;
+  for (const auto &a : *nodes) {
+    for (const auto &b : *nodes) {
+      seen.clear();
+      bool gc_reachable = gc->IsReachable(Get(id, a), Get(id, b));
+      bool reachable = IsReachable(edges, a, b, &seen);
+      if (gc_reachable != reachable) {
+        PrintEdges(edges);
+        PrintGCEdges(nodes, id, gc);
+        PrintTransitiveClosure(nodes, edges);
+        PrintGCTransitiveClosure(nodes, id, gc);
+        ABSL_RAW_LOG(FATAL, "gc_reachable %s reachable %s a %d b %d",
+                     gc_reachable ? "true" : "false",
+                     reachable ? "true" : "false", a, b);
+      }
+    }
+  }
+}
+
+static void CheckEdges(Nodes *nodes, Edges *edges, const IdMap &id,
+                       GraphCycles *gc) {
+  int count = 0;
+  for (const auto &edge : *edges) {
+    int a = edge.from;
+    int b = edge.to;
+    if (!gc->HasEdge(Get(id, a), Get(id, b))) {
+      PrintEdges(edges);
+      PrintGCEdges(nodes, id, gc);
+      ABSL_RAW_LOG(FATAL, "!gc->HasEdge(%d, %d)", a, b);
+    }
+  }
+  for (const auto &a : *nodes) {
+    for (const auto &b : *nodes) {
+      if (gc->HasEdge(Get(id, a), Get(id, b))) {
+        count++;
+      }
+    }
+  }
+  if (count != edges->size()) {
+    PrintEdges(edges);
+    PrintGCEdges(nodes, id, gc);
+    ABSL_RAW_LOG(FATAL, "edges->size() %zu  count %d", edges->size(), count);
+  }
+}
+
+static void CheckInvariants(const GraphCycles &gc) {
+  if (ABSL_PREDICT_FALSE(!gc.CheckInvariants()))
+    ABSL_RAW_LOG(FATAL, "CheckInvariants");
+}
+
+// Returns the index of a randomly chosen node in *nodes.
+// Requires *nodes be non-empty.
+static int RandomNode(RandomEngine* rng, Nodes *nodes) {
+  std::uniform_int_distribution<int> uniform(0, nodes->size()-1);
+  return uniform(*rng);
+}
+
+// Returns the index of a randomly chosen edge in *edges.
+// Requires *edges be non-empty.
+static int RandomEdge(RandomEngine* rng, Edges *edges) {
+  std::uniform_int_distribution<int> uniform(0, edges->size()-1);
+  return uniform(*rng);
+}
+
+// Returns the index of edge (from, to) in *edges or -1 if it is not in *edges.
+static int EdgeIndex(Edges *edges, int from, int to) {
+  int i = 0;
+  while (i != edges->size() &&
+         ((*edges)[i].from != from || (*edges)[i].to != to)) {
+    i++;
+  }
+  return i == edges->size()? -1 : i;
+}
+
+TEST(GraphCycles, RandomizedTest) {
+  int next_node = 0;
+  Nodes nodes;
+  Edges edges;   // from, to
+  IdMap id;
+  GraphCycles graph_cycles;
+  static const int kMaxNodes = 7;  // use <= 7 nodes to keep test short
+  static const int kDataOffset = 17;  // an offset to the node-specific data
+  int n = 100000;
+  int op = 0;
+  RandomEngine rng(testing::UnitTest::GetInstance()->random_seed());
+  std::uniform_int_distribution<int> uniform(0, 5);
+
+  auto ptr = [](intptr_t i) {
+    return reinterpret_cast<void*>(i + kDataOffset);
+  };
+
+  for (int iter = 0; iter != n; iter++) {
+    for (const auto &node : nodes) {
+      ASSERT_EQ(graph_cycles.Ptr(Get(id, node)), ptr(node)) << " node " << node;
+    }
+    CheckEdges(&nodes, &edges, id, &graph_cycles);
+    CheckTransitiveClosure(&nodes, &edges, id, &graph_cycles);
+    op = uniform(rng);
+    switch (op) {
+    case 0:     // Add a node
+      if (nodes.size() < kMaxNodes) {
+        int new_node = next_node++;
+        GraphId new_gnode = graph_cycles.GetId(ptr(new_node));
+        ASSERT_NE(new_gnode, InvalidGraphId());
+        id[new_node] = new_gnode;
+        ASSERT_EQ(ptr(new_node), graph_cycles.Ptr(new_gnode));
+        nodes.push_back(new_node);
+      }
+      break;
+
+    case 1:    // Remove a node
+      if (nodes.size() > 0) {
+        int node_index = RandomNode(&rng, &nodes);
+        int node = nodes[node_index];
+        nodes[node_index] = nodes.back();
+        nodes.pop_back();
+        graph_cycles.RemoveNode(ptr(node));
+        ASSERT_EQ(graph_cycles.Ptr(Get(id, node)), nullptr);
+        id.erase(node);
+        int i = 0;
+        while (i != edges.size()) {
+          if (edges[i].from == node || edges[i].to == node) {
+            edges[i] = edges.back();
+            edges.pop_back();
+          } else {
+            i++;
+          }
+        }
+      }
+      break;
+
+    case 2:   // Add an edge
+      if (nodes.size() > 0) {
+        int from = RandomNode(&rng, &nodes);
+        int to = RandomNode(&rng, &nodes);
+        if (EdgeIndex(&edges, nodes[from], nodes[to]) == -1) {
+          if (graph_cycles.InsertEdge(id[nodes[from]], id[nodes[to]])) {
+            Edge new_edge;
+            new_edge.from = nodes[from];
+            new_edge.to = nodes[to];
+            edges.push_back(new_edge);
+          } else {
+            std::unordered_set<int> seen;
+            ASSERT_TRUE(IsReachable(&edges, nodes[to], nodes[from], &seen))
+                << "Edge " << nodes[to] << "->" << nodes[from];
+          }
+        }
+      }
+      break;
+
+    case 3:    // Remove an edge
+      if (edges.size() > 0) {
+        int i = RandomEdge(&rng, &edges);
+        int from = edges[i].from;
+        int to = edges[i].to;
+        ASSERT_EQ(i, EdgeIndex(&edges, from, to));
+        edges[i] = edges.back();
+        edges.pop_back();
+        ASSERT_EQ(-1, EdgeIndex(&edges, from, to));
+        graph_cycles.RemoveEdge(id[from], id[to]);
+      }
+      break;
+
+    case 4:   // Check a path
+      if (nodes.size() > 0) {
+        int from = RandomNode(&rng, &nodes);
+        int to = RandomNode(&rng, &nodes);
+        GraphId path[2*kMaxNodes];
+        int path_len = graph_cycles.FindPath(id[nodes[from]], id[nodes[to]],
+                                             ABSL_ARRAYSIZE(path), path);
+        std::unordered_set<int> seen;
+        bool reachable = IsReachable(&edges, nodes[from], nodes[to], &seen);
+        bool gc_reachable =
+            graph_cycles.IsReachable(Get(id, nodes[from]), Get(id, nodes[to]));
+        ASSERT_EQ(path_len != 0, reachable);
+        ASSERT_EQ(path_len != 0, gc_reachable);
+        // In the following line, we add one because a node can appear
+        // twice, if the path is from that node to itself, perhaps via
+        // every other node.
+        ASSERT_LE(path_len, kMaxNodes + 1);
+        if (path_len != 0) {
+          ASSERT_EQ(id[nodes[from]], path[0]);
+          ASSERT_EQ(id[nodes[to]], path[path_len-1]);
+          for (int i = 1; i < path_len; i++) {
+            ASSERT_TRUE(graph_cycles.HasEdge(path[i-1], path[i]));
+          }
+        }
+      }
+      break;
+
+    case 5:  // Check invariants
+      CheckInvariants(graph_cycles);
+      break;
+
+    default:
+      ABSL_RAW_LOG(FATAL, "op %d", op);
+    }
+
+    // Very rarely, test graph expansion by adding then removing many nodes.
+    std::bernoulli_distribution one_in_1024(1.0 / 1024);
+    if (one_in_1024(rng)) {
+      CheckEdges(&nodes, &edges, id, &graph_cycles);
+      CheckTransitiveClosure(&nodes, &edges, id, &graph_cycles);
+      for (int i = 0; i != 256; i++) {
+        int new_node = next_node++;
+        GraphId new_gnode = graph_cycles.GetId(ptr(new_node));
+        ASSERT_NE(InvalidGraphId(), new_gnode);
+        id[new_node] = new_gnode;
+        ASSERT_EQ(ptr(new_node), graph_cycles.Ptr(new_gnode));
+        for (const auto &node : nodes) {
+          ASSERT_NE(node, new_node);
+        }
+        nodes.push_back(new_node);
+      }
+      for (int i = 0; i != 256; i++) {
+        ASSERT_GT(nodes.size(), 0);
+        int node_index = RandomNode(&rng, &nodes);
+        int node = nodes[node_index];
+        nodes[node_index] = nodes.back();
+        nodes.pop_back();
+        graph_cycles.RemoveNode(ptr(node));
+        id.erase(node);
+        int j = 0;
+        while (j != edges.size()) {
+          if (edges[j].from == node || edges[j].to == node) {
+            edges[j] = edges.back();
+            edges.pop_back();
+          } else {
+            j++;
+          }
+        }
+      }
+      CheckInvariants(graph_cycles);
+    }
+  }
+}
+
+class GraphCyclesTest : public ::testing::Test {
+ public:
+  IdMap id_;
+  GraphCycles g_;
+
+  static void* Ptr(int i) {
+    return reinterpret_cast<void*>(static_cast<uintptr_t>(i));
+  }
+
+  static int Num(void* ptr) {
+    return static_cast<int>(reinterpret_cast<uintptr_t>(ptr));
+  }
+
+  // Test relies on ith NewNode() call returning Node numbered i
+  GraphCyclesTest() {
+    for (int i = 0; i < 100; i++) {
+      id_[i] = g_.GetId(Ptr(i));
+    }
+    CheckInvariants(g_);
+  }
+
+  bool AddEdge(int x, int y) {
+    return g_.InsertEdge(Get(id_, x), Get(id_, y));
+  }
+
+  void AddMultiples() {
+    // For every node x > 0: add edge to 2*x, 3*x
+    for (int x = 1; x < 25; x++) {
+      EXPECT_TRUE(AddEdge(x, 2*x)) << x;
+      EXPECT_TRUE(AddEdge(x, 3*x)) << x;
+    }
+    CheckInvariants(g_);
+  }
+
+  std::string Path(int x, int y) {
+    GraphId path[5];
+    int np = g_.FindPath(Get(id_, x), Get(id_, y), ABSL_ARRAYSIZE(path), path);
+    std::string result;
+    for (int i = 0; i < np; i++) {
+      if (i >= ABSL_ARRAYSIZE(path)) {
+        result += " ...";
+        break;
+      }
+      if (!result.empty()) result.push_back(' ');
+      char buf[20];
+      snprintf(buf, sizeof(buf), "%d", Num(g_.Ptr(path[i])));
+      result += buf;
+    }
+    return result;
+  }
+};
+
+TEST_F(GraphCyclesTest, NoCycle) {
+  AddMultiples();
+  CheckInvariants(g_);
+}
+
+TEST_F(GraphCyclesTest, SimpleCycle) {
+  AddMultiples();
+  EXPECT_FALSE(AddEdge(8, 4));
+  EXPECT_EQ("4 8", Path(4, 8));
+  CheckInvariants(g_);
+}
+
+TEST_F(GraphCyclesTest, IndirectCycle) {
+  AddMultiples();
+  EXPECT_TRUE(AddEdge(16, 9));
+  CheckInvariants(g_);
+  EXPECT_FALSE(AddEdge(9, 2));
+  EXPECT_EQ("2 4 8 16 9", Path(2, 9));
+  CheckInvariants(g_);
+}
+
+TEST_F(GraphCyclesTest, LongPath) {
+  ASSERT_TRUE(AddEdge(2, 4));
+  ASSERT_TRUE(AddEdge(4, 6));
+  ASSERT_TRUE(AddEdge(6, 8));
+  ASSERT_TRUE(AddEdge(8, 10));
+  ASSERT_TRUE(AddEdge(10, 12));
+  ASSERT_FALSE(AddEdge(12, 2));
+  EXPECT_EQ("2 4 6 8 10 ...", Path(2, 12));
+  CheckInvariants(g_);
+}
+
+TEST_F(GraphCyclesTest, RemoveNode) {
+  ASSERT_TRUE(AddEdge(1, 2));
+  ASSERT_TRUE(AddEdge(2, 3));
+  ASSERT_TRUE(AddEdge(3, 4));
+  ASSERT_TRUE(AddEdge(4, 5));
+  g_.RemoveNode(g_.Ptr(id_[3]));
+  id_.erase(3);
+  ASSERT_TRUE(AddEdge(5, 1));
+}
+
+TEST_F(GraphCyclesTest, ManyEdges) {
+  const int N = 50;
+  for (int i = 0; i < N; i++) {
+    for (int j = 1; j < N; j++) {
+      ASSERT_TRUE(AddEdge(i, i+j));
+    }
+  }
+  CheckInvariants(g_);
+  ASSERT_TRUE(AddEdge(2*N-1, 0));
+  CheckInvariants(g_);
+  ASSERT_FALSE(AddEdge(10, 9));
+  CheckInvariants(g_);
+}
+
+}  // namespace synchronization_internal
+}  // namespace absl
diff --git a/absl/synchronization/internal/kernel_timeout.h b/absl/synchronization/internal/kernel_timeout.h
new file mode 100644
index 000000000000..a83c427bf499
--- /dev/null
+++ b/absl/synchronization/internal/kernel_timeout.h
@@ -0,0 +1,147 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+
+// An optional absolute timeout, with nanosecond granularity,
+// compatible with absl::Time. Suitable for in-register
+// parameter-passing (e.g. syscalls.)
+// Constructible from a absl::Time (for a timeout to be respected) or {}
+// (for "no timeout".)
+// This is a private low-level API for use by a handful of low-level
+// components that are friends of this class. Higher-level components
+// should build APIs based on absl::Time and absl::Duration.
+
+#ifndef ABSL_SYNCHRONIZATION_INTERNAL_KERNEL_TIMEOUT_H_
+#define ABSL_SYNCHRONIZATION_INTERNAL_KERNEL_TIMEOUT_H_
+
+#ifdef _WIN32
+#include <intsafe.h>
+#endif
+#include <time.h>
+#include <algorithm>
+#include <limits>
+
+#include "absl/base/internal/raw_logging.h"
+#include "absl/time/clock.h"
+#include "absl/time/time.h"
+
+namespace absl {
+namespace synchronization_internal {
+
+class Waiter;
+
+class KernelTimeout {
+ public:
+  // A timeout that should expire at <t>.  Any value, in the full
+  // InfinitePast() to InfiniteFuture() range, is valid here and will be
+  // respected.
+  explicit KernelTimeout(absl::Time t) : ns_(MakeNs(t)) {}
+  // No timeout.
+  KernelTimeout() : ns_(0) {}
+
+  // A more explicit factory for those who prefer it.  Equivalent to {}.
+  static KernelTimeout Never() { return {}; }
+
+  // We explicitly do not support other custom formats: timespec, int64_t nanos.
+  // Unify on this and absl::Time, please.
+  bool has_timeout() const { return ns_ != 0; }
+
+ private:
+  // internal rep, not user visible: ns after unix epoch.
+  // zero = no timeout.
+  // Negative we treat as an unlikely (and certainly expired!) but valid
+  // timeout.
+  int64_t ns_;
+
+  static int64_t MakeNs(absl::Time t) {
+    // optimization--InfiniteFuture is common "no timeout" value
+    // and cheaper to compare than convert.
+    if (t == absl::InfiniteFuture()) return 0;
+    int64_t x = ToUnixNanos(t);
+
+    // A timeout that lands exactly on the epoch (x=0) needs to be respected,
+    // so we alter it unnoticably to 1.  Negative timeouts are in
+    // theory supported, but handled poorly by the kernel (long
+    // delays) so push them forward too; since all such times have
+    // already passed, it's indistinguishable.
+    if (x <= 0) x = 1;
+    // A time larger than what can be represented to the kernel is treated
+    // as no timeout.
+    if (x == std::numeric_limits<int64_t>::max()) x = 0;
+    return x;
+  }
+
+  // Convert to parameter for sem_timedwait/futex/similar.  Only for approved
+  // users.  Do not call if !has_timeout.
+  struct timespec MakeAbsTimespec() {
+    int64_t n = ns_;
+    static const int64_t kNanosPerSecond = 1000 * 1000 * 1000;
+    if (n == 0) {
+      ABSL_RAW_LOG(
+          ERROR,
+          "Tried to create a timespec from a non-timeout; never do this.");
+      // But we'll try to continue sanely.  no-timeout ~= saturated timeout.
+      n = std::numeric_limits<int64_t>::max();
+    }
+
+    // Kernel APIs validate timespecs as being at or after the epoch,
+    // despite the kernel time type being signed.  However, no one can
+    // tell the difference between a timeout at or before the epoch (since
+    // all such timeouts have expired!)
+    if (n < 0) n = 0;
+
+    struct timespec abstime;
+    int64_t seconds = std::min(n / kNanosPerSecond,
+                             int64_t{std::numeric_limits<time_t>::max()});
+    abstime.tv_sec = static_cast<time_t>(seconds);
+    abstime.tv_nsec =
+        static_cast<decltype(abstime.tv_nsec)>(n % kNanosPerSecond);
+    return abstime;
+  }
+
+#ifdef _WIN32
+  // Converts to milliseconds from now, or INFINITE when
+  // !has_timeout(). For use by SleepConditionVariableSRW on
+  // Windows. Callers should recognize that the return value is a
+  // relative duration (it should be recomputed by calling this method
+  // in the case of a spurious wakeup).
+  DWORD InMillisecondsFromNow() const {
+    if (!has_timeout()) {
+      return INFINITE;
+    }
+    // The use of absl::Now() to convert from absolute time to
+    // relative time means that absl::Now() cannot use anything that
+    // depends on KernelTimeout (for example, Mutex) on Windows.
+    int64_t now = ToUnixNanos(absl::Now());
+    if (ns_ >= now) {
+      // Round up so that Now() + ms_from_now >= ns_.
+      constexpr uint64_t max_nanos =
+          std::numeric_limits<int64_t>::max() - 999999u;
+      uint64_t ms_from_now =
+          (std::min<uint64_t>(max_nanos, ns_ - now) + 999999u) / 1000000u;
+      if (ms_from_now > std::numeric_limits<DWORD>::max()) {
+        return INFINITE;
+      }
+      return static_cast<DWORD>(ms_from_now);
+    }
+    return 0;
+  }
+#endif
+
+  friend class Waiter;
+};
+
+}  // namespace synchronization_internal
+}  // namespace absl
+#endif  // ABSL_SYNCHRONIZATION_INTERNAL_KERNEL_TIMEOUT_H_
diff --git a/absl/synchronization/internal/mutex_nonprod.cc b/absl/synchronization/internal/mutex_nonprod.cc
new file mode 100644
index 000000000000..94be54b88ff5
--- /dev/null
+++ b/absl/synchronization/internal/mutex_nonprod.cc
@@ -0,0 +1,311 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// Implementation of a small subset of Mutex and CondVar functionality
+// for platforms where the production implementation hasn't been fully
+// ported yet.
+
+#include "absl/synchronization/mutex.h"
+
+#if defined(_WIN32)
+#include <chrono>  // NOLINT(build/c++11)
+#else
+#include <sys/time.h>
+#include <time.h>
+#endif
+
+#include <algorithm>
+
+#include "absl/base/internal/raw_logging.h"
+#include "absl/time/time.h"
+
+namespace absl {
+namespace synchronization_internal {
+
+namespace {
+
+// Return the current time plus the timeout.
+absl::Time DeadlineFromTimeout(absl::Duration timeout) {
+  return absl::Now() + timeout;
+}
+
+// Limit the deadline to a positive, 32-bit time_t value to accommodate
+// implementation restrictions.  This also deals with InfinitePast and
+// InfiniteFuture.
+absl::Time LimitedDeadline(absl::Time deadline) {
+  deadline = std::max(absl::FromTimeT(0), deadline);
+  deadline = std::min(deadline, absl::FromTimeT(0x7fffffff));
+  return deadline;
+}
+
+}  // namespace
+
+#if defined(_WIN32)
+
+MutexImpl::MutexImpl() {}
+
+MutexImpl::~MutexImpl() {
+  if (locked_) {
+    std_mutex_.unlock();
+  }
+}
+
+void MutexImpl::Lock() {
+  std_mutex_.lock();
+  locked_ = true;
+}
+
+bool MutexImpl::TryLock() {
+  bool locked = std_mutex_.try_lock();
+  if (locked) locked_ = true;
+  return locked;
+}
+
+void MutexImpl::Unlock() {
+  locked_ = false;
+  released_.SignalAll();
+  std_mutex_.unlock();
+}
+
+CondVarImpl::CondVarImpl() {}
+
+CondVarImpl::~CondVarImpl() {}
+
+void CondVarImpl::Signal() { std_cv_.notify_one(); }
+
+void CondVarImpl::SignalAll() { std_cv_.notify_all(); }
+
+void CondVarImpl::Wait(MutexImpl* mu) {
+  mu->released_.SignalAll();
+  std_cv_.wait(mu->std_mutex_);
+}
+
+bool CondVarImpl::WaitWithDeadline(MutexImpl* mu, absl::Time deadline) {
+  mu->released_.SignalAll();
+  time_t when = ToTimeT(deadline);
+  int64_t nanos = ToInt64Nanoseconds(deadline - absl::FromTimeT(when));
+  std::chrono::system_clock::time_point deadline_tp =
+      std::chrono::system_clock::from_time_t(when) +
+      std::chrono::duration_cast<std::chrono::system_clock::duration>(
+          std::chrono::nanoseconds(nanos));
+  auto deadline_since_epoch =
+      std::chrono::duration_cast<std::chrono::duration<double>>(
+          deadline_tp - std::chrono::system_clock::from_time_t(0));
+  return std_cv_.wait_until(mu->std_mutex_, deadline_tp) ==
+         std::cv_status::timeout;
+}
+
+#else  // ! _WIN32
+
+MutexImpl::MutexImpl() {
+  ABSL_RAW_CHECK(pthread_mutex_init(&pthread_mutex_, nullptr) == 0,
+                 "pthread error");
+}
+
+MutexImpl::~MutexImpl() {
+  if (locked_) {
+    ABSL_RAW_CHECK(pthread_mutex_unlock(&pthread_mutex_) == 0, "pthread error");
+  }
+  ABSL_RAW_CHECK(pthread_mutex_destroy(&pthread_mutex_) == 0, "pthread error");
+}
+
+void MutexImpl::Lock() {
+  ABSL_RAW_CHECK(pthread_mutex_lock(&pthread_mutex_) == 0, "pthread error");
+  locked_ = true;
+}
+
+bool MutexImpl::TryLock() {
+  bool locked = (0 == pthread_mutex_trylock(&pthread_mutex_));
+  if (locked) locked_ = true;
+  return locked;
+}
+
+void MutexImpl::Unlock() {
+  locked_ = false;
+  released_.SignalAll();
+  ABSL_RAW_CHECK(pthread_mutex_unlock(&pthread_mutex_) == 0, "pthread error");
+}
+
+CondVarImpl::CondVarImpl() {
+  ABSL_RAW_CHECK(pthread_cond_init(&pthread_cv_, nullptr) == 0,
+                 "pthread error");
+}
+
+CondVarImpl::~CondVarImpl() {
+  ABSL_RAW_CHECK(pthread_cond_destroy(&pthread_cv_) == 0, "pthread error");
+}
+
+void CondVarImpl::Signal() {
+  ABSL_RAW_CHECK(pthread_cond_signal(&pthread_cv_) == 0, "pthread error");
+}
+
+void CondVarImpl::SignalAll() {
+  ABSL_RAW_CHECK(pthread_cond_broadcast(&pthread_cv_) == 0, "pthread error");
+}
+
+void CondVarImpl::Wait(MutexImpl* mu) {
+  mu->released_.SignalAll();
+  ABSL_RAW_CHECK(pthread_cond_wait(&pthread_cv_, &mu->pthread_mutex_) == 0,
+                 "pthread error");
+}
+
+bool CondVarImpl::WaitWithDeadline(MutexImpl* mu, absl::Time deadline) {
+  mu->released_.SignalAll();
+  struct timespec ts = ToTimespec(deadline);
+  int rc = pthread_cond_timedwait(&pthread_cv_, &mu->pthread_mutex_, &ts);
+  if (rc == ETIMEDOUT) return true;
+  ABSL_RAW_CHECK(rc == 0, "pthread error");
+  return false;
+}
+
+#endif  // ! _WIN32
+
+void MutexImpl::Await(const Condition& cond) {
+  if (cond.Eval()) return;
+  released_.SignalAll();
+  do {
+    released_.Wait(this);
+  } while (!cond.Eval());
+}
+
+bool MutexImpl::AwaitWithDeadline(const Condition& cond, absl::Time deadline) {
+  if (cond.Eval()) return true;
+  released_.SignalAll();
+  while (true) {
+    if (released_.WaitWithDeadline(this, deadline)) return false;
+    if (cond.Eval()) return true;
+  }
+}
+
+}  // namespace synchronization_internal
+
+Mutex::Mutex() {}
+
+Mutex::~Mutex() {}
+
+void Mutex::Lock() { impl()->Lock(); }
+
+void Mutex::Unlock() { impl()->Unlock(); }
+
+bool Mutex::TryLock() { return impl()->TryLock(); }
+
+void Mutex::ReaderLock() { Lock(); }
+
+void Mutex::ReaderUnlock() { Unlock(); }
+
+void Mutex::Await(const Condition& cond) { impl()->Await(cond); }
+
+void Mutex::LockWhen(const Condition& cond) {
+  Lock();
+  Await(cond);
+}
+
+bool Mutex::AwaitWithDeadline(const Condition& cond, absl::Time deadline) {
+  return impl()->AwaitWithDeadline(
+      cond, synchronization_internal::LimitedDeadline(deadline));
+}
+
+bool Mutex::AwaitWithTimeout(const Condition& cond, absl::Duration timeout) {
+  return AwaitWithDeadline(
+      cond, synchronization_internal::DeadlineFromTimeout(timeout));
+}
+
+bool Mutex::LockWhenWithDeadline(const Condition& cond, absl::Time deadline) {
+  Lock();
+  return AwaitWithDeadline(cond, deadline);
+}
+
+bool Mutex::LockWhenWithTimeout(const Condition& cond, absl::Duration timeout) {
+  return LockWhenWithDeadline(
+      cond, synchronization_internal::DeadlineFromTimeout(timeout));
+}
+
+bool Mutex::ReaderLockWhenWithTimeout(const Condition& cond,
+                                      absl::Duration timeout) {
+  return LockWhenWithTimeout(cond, timeout);
+}
+bool Mutex::ReaderLockWhenWithDeadline(const Condition& cond,
+                                       absl::Time deadline) {
+  return LockWhenWithDeadline(cond, deadline);
+}
+
+void Mutex::EnableDebugLog(const char*) {}
+void Mutex::EnableInvariantDebugging(void (*)(void*), void*) {}
+void Mutex::ForgetDeadlockInfo() {}
+void Mutex::AssertHeld() const {}
+void Mutex::AssertReaderHeld() const {}
+void Mutex::AssertNotHeld() const {}
+
+CondVar::CondVar() {}
+
+CondVar::~CondVar() {}
+
+void CondVar::Signal() { impl()->Signal(); }
+
+void CondVar::SignalAll() { impl()->SignalAll(); }
+
+void CondVar::Wait(Mutex* mu) { return impl()->Wait(mu->impl()); }
+
+bool CondVar::WaitWithDeadline(Mutex* mu, absl::Time deadline) {
+  return impl()->WaitWithDeadline(
+      mu->impl(), synchronization_internal::LimitedDeadline(deadline));
+}
+
+bool CondVar::WaitWithTimeout(Mutex* mu, absl::Duration timeout) {
+  return WaitWithDeadline(mu, absl::Now() + timeout);
+}
+
+void CondVar::EnableDebugLog(const char*) {}
+
+#ifdef THREAD_SANITIZER
+extern "C" void __tsan_read1(void *addr);
+#else
+#define __tsan_read1(addr)  // do nothing if TSan not enabled
+#endif
+
+// A function that just returns its argument, dereferenced
+static bool Dereference(void *arg) {
+  // ThreadSanitizer does not instrument this file for memory accesses.
+  // This function dereferences a user variable that can participate
+  // in a data race, so we need to manually tell TSan about this memory access.
+  __tsan_read1(arg);
+  return *(static_cast<bool *>(arg));
+}
+
+Condition::Condition() {}   // null constructor, used for kTrue only
+const Condition Condition::kTrue;
+
+Condition::Condition(bool (*func)(void *), void *arg)
+    : eval_(&CallVoidPtrFunction),
+      function_(func),
+      method_(nullptr),
+      arg_(arg) {}
+
+bool Condition::CallVoidPtrFunction(const Condition *c) {
+  return (*c->function_)(c->arg_);
+}
+
+Condition::Condition(const bool *cond)
+    : eval_(CallVoidPtrFunction),
+      function_(Dereference),
+      method_(nullptr),
+      // const_cast is safe since Dereference does not modify arg
+      arg_(const_cast<bool *>(cond)) {}
+
+bool Condition::Eval() const {
+  // eval_ == null for kTrue
+  return (this->eval_ == nullptr) || (*this->eval_)(this);
+}
+
+}  // namespace absl
diff --git a/absl/synchronization/internal/mutex_nonprod.inc b/absl/synchronization/internal/mutex_nonprod.inc
new file mode 100644
index 000000000000..51441b2577fc
--- /dev/null
+++ b/absl/synchronization/internal/mutex_nonprod.inc
@@ -0,0 +1,256 @@
+// Do not include.  This is an implementation detail of base/mutex.h.
+//
+// Declares three classes:
+//
+// base::internal::MutexImpl - implementation helper for Mutex
+// base::internal::CondVarImpl - implementation helper for CondVar
+// base::internal::SynchronizationStorage<T> - implementation helper for
+//                                             Mutex, CondVar
+
+#include <type_traits>
+
+#if defined(_WIN32)
+#include <condition_variable>
+#include <mutex>
+#else
+#include <pthread.h>
+#endif
+
+#include "absl/base/call_once.h"
+#include "absl/time/time.h"
+
+// Declare that Mutex::ReaderLock is actually Lock().  Intended primarily
+// for tests, and even then as a last resort.
+#ifdef ABSL_MUTEX_READER_LOCK_IS_EXCLUSIVE
+#error ABSL_MUTEX_READER_LOCK_IS_EXCLUSIVE cannot be directly set
+#else
+#define ABSL_MUTEX_READER_LOCK_IS_EXCLUSIVE 1
+#endif
+
+// Declare that Mutex::EnableInvariantDebugging is not implemented.
+// Intended primarily for tests, and even then as a last resort.
+#ifdef ABSL_MUTEX_ENABLE_INVARIANT_DEBUGGING_NOT_IMPLEMENTED
+#error ABSL_MUTEX_ENABLE_INVARIANT_DEBUGGING_NOT_IMPLEMENTED cannot be directly set
+#else
+#define ABSL_MUTEX_ENABLE_INVARIANT_DEBUGGING_NOT_IMPLEMENTED 1
+#endif
+
+namespace absl {
+class Condition;
+
+namespace synchronization_internal {
+
+class MutexImpl;
+
+// Do not use this implementation detail of CondVar. Provides most of the
+// implementation, but should not be placed directly in static storage
+// because it will not linker initialize properly. See
+// SynchronizationStorage<T> below for what we mean by linker
+// initialization.
+class CondVarImpl {
+ public:
+  CondVarImpl();
+  CondVarImpl(const CondVarImpl&) = delete;
+  CondVarImpl& operator=(const CondVarImpl&) = delete;
+  ~CondVarImpl();
+
+  void Signal();
+  void SignalAll();
+  void Wait(MutexImpl* mutex);
+  bool WaitWithDeadline(MutexImpl* mutex, absl::Time deadline);
+
+ private:
+#if defined(_WIN32)
+  std::condition_variable_any std_cv_;
+#else
+  pthread_cond_t pthread_cv_;
+#endif
+};
+
+// Do not use this implementation detail of Mutex. Provides most of the
+// implementation, but should not be placed directly in static storage
+// because it will not linker initialize properly. See
+// SynchronizationStorage<T> below for what we mean by linker
+// initialization.
+class MutexImpl {
+ public:
+  MutexImpl();
+  MutexImpl(const MutexImpl&) = delete;
+  MutexImpl& operator=(const MutexImpl&) = delete;
+  ~MutexImpl();
+
+  void Lock();
+  bool TryLock();
+  void Unlock();
+  void Await(const Condition& cond);
+  bool AwaitWithDeadline(const Condition& cond, absl::Time deadline);
+
+ private:
+  friend class CondVarImpl;
+
+#if defined(_WIN32)
+  std::mutex std_mutex_;
+#else
+  pthread_mutex_t pthread_mutex_;
+#endif
+
+  // True if the underlying mutex is locked.  If the destructor is entered
+  // while locked_, the underlying mutex is unlocked.  Mutex supports
+  // destruction while locked, but the same is undefined behavior for both
+  // pthread_mutex_t and std::mutex.
+  bool locked_ = false;
+
+  // Signaled before releasing the lock, in support of Await.
+  CondVarImpl released_;
+};
+
+// Do not use this implementation detail of CondVar and Mutex.  A storage
+// space for T that supports a base::LinkerInitialized constructor. T must
+// have a default constructor, which is called by the first call to
+// get(). T's destructor is never called if the base::LinkerInitialized
+// constructor is called.
+//
+// Objects constructed with the default constructor are constructed and
+// destructed like any other object, and should never be allocated in
+// static storage.
+//
+// Objects constructed with the base::LinkerInitialized constructor should
+// always be in static storage. For such objects, calls to get() are always
+// valid, except from signal handlers.
+//
+// Note that this implementation relies on undefined language behavior that
+// are known to hold for the set of supported compilers. An analysis
+// follows.
+//
+// From the C++11 standard:
+//
+// [basic.life] says an object has non-trivial initialization if it is of
+// class type and it is initialized by a constructor other than a trivial
+// default constructor.  (the base::LinkerInitialized constructor is
+// non-trivial)
+//
+// [basic.life] says the lifetime of an object with a non-trivial
+// constructor begins when the call to the constructor is complete.
+//
+// [basic.life] says the lifetime of an object with non-trivial destructor
+// ends when the call to the destructor begins.
+//
+// [basic.life] p5 specifies undefined behavior when accessing non-static
+// members of an instance outside its
+// lifetime. (SynchronizationStorage::get() access non-static members)
+//
+// So, base::LinkerInitialized object of SynchronizationStorage uses a
+// non-trivial constructor, which is called at some point during dynamic
+// initialization, and is therefore subject to order of dynamic
+// initialization bugs, where get() is called before the object's
+// constructor is, resulting in undefined behavior.
+//
+// Similarly, a base::LinkerInitialized SynchronizationStorage object has a
+// non-trivial destructor, and so its lifetime ends at some point during
+// destruction of objects with static storage duration [basic.start.term]
+// p4. There is a window where other exit code could call get() after this
+// occurs, resulting in undefined behavior.
+//
+// Combined, these statements imply that base::LinkerInitialized instances
+// of SynchronizationStorage<T> rely on undefined behavior.
+//
+// However, in practice, the implementation works on all supported
+// compilers. Specifically, we rely on:
+//
+// a) zero-initialization being sufficient to initialize
+// base::LinkerInitialized instances for the purposes of calling
+// get(), regardless of when the constructor is called. This is
+// because the is_dynamic_ boolean is correctly zero-initialized to
+// false.
+//
+// b) the base::LinkerInitialized constructor is a NOP, and immaterial to
+// even to concurrent calls to get().
+//
+// c) the destructor being a NOP for base::LinkerInitialized objects
+// (guaranteed by a check for !is_dynamic_), and so any concurrent and
+// subsequent calls to get() functioning as if the destructor were not
+// called, by virtue of the instances' storage remaining valid after the
+// destructor runs.
+//
+// d) That a-c apply transitively when SynchronizationStorage<T> is the
+// only member of a class allocated in static storage.
+//
+// Nothing in the language standard guarantees that a-d hold.  In practice,
+// these hold in all supported compilers.
+//
+// Future direction:
+//
+// Ideally, we would simply use std::mutex or a similar class, which when
+// allocated statically would support use immediately after static
+// initialization up until static storage is reclaimed (i.e. the properties
+// we require of all "linker initialized" instances).
+//
+// Regarding construction in static storage, std::mutex is required to
+// provide a constexpr default constructor [thread.mutex.class], which
+// ensures the instance's lifetime begins with static initialization
+// [basic.start.init], and so is immune to any problems caused by the order
+// of dynamic initialization. However, as of this writing Microsoft's
+// Visual Studio does not provide a constexpr constructor for std::mutex.
+// See
+// https://blogs.msdn.microsoft.com/vcblog/2015/06/02/constexpr-complete-for-vs-2015-rtm-c11-compiler-c17-stl/
+//
+// Regarding destruction of instances in static storage, [basic.life] does
+// say an object ends when storage in which the occupies is released, in
+// the case of non-trivial destructor. However, std::mutex is not specified
+// to have a trivial destructor.
+//
+// So, we would need a class with a constexpr default constructor and a
+// trivial destructor. Today, we can achieve neither desired property using
+// std::mutex directly.
+template <typename T>
+class SynchronizationStorage {
+ public:
+  // Instances allocated on the heap or on the stack should use the default
+  // constructor.
+  SynchronizationStorage()
+      : is_dynamic_(true), once_() {}
+
+  // Instances allocated in static storage (not on the heap, not on the
+  // stack) should use this constructor.
+  explicit SynchronizationStorage(base::LinkerInitialized) {}
+
+  SynchronizationStorage(SynchronizationStorage&) = delete;
+  SynchronizationStorage& operator=(SynchronizationStorage&) = delete;
+
+  ~SynchronizationStorage() {
+    if (is_dynamic_) {
+      get()->~T();
+    }
+  }
+
+  // Retrieve the object in storage. This is fast and thread safe, but does
+  // incur the cost of absl::call_once().
+  //
+  // For instances in static storage constructed with the
+  // base::LinkerInitialized constructor, may be called at any time without
+  // regard for order of dynamic initialization or destruction of objects
+  // in static storage. See the class comment for caveats.
+  T* get() {
+    absl::call_once(once_, SynchronizationStorage::Construct, this);
+    return reinterpret_cast<T*>(&space_);
+  }
+
+ private:
+  static void Construct(SynchronizationStorage<T>* self) {
+    new (&self->space_) T();
+  }
+
+  // When true, T's destructor is run when this is destructed.
+  //
+  // The base::LinkerInitialized constructor assumes this value will be set
+  // false by static initialization.
+  bool is_dynamic_;
+
+  absl::once_flag once_;
+
+  // An aligned space for T.
+  typename std::aligned_storage<sizeof(T), alignof(T)>::type space_;
+};
+
+}  // namespace synchronization_internal
+}  // namespace absl
diff --git a/absl/synchronization/internal/per_thread_sem.cc b/absl/synchronization/internal/per_thread_sem.cc
new file mode 100644
index 000000000000..af87222816d5
--- /dev/null
+++ b/absl/synchronization/internal/per_thread_sem.cc
@@ -0,0 +1,106 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+// This file is a no-op if the required LowLevelAlloc support is missing.
+#include "absl/base/internal/low_level_alloc.h"
+#ifndef ABSL_LOW_LEVEL_ALLOC_MISSING
+
+#include "absl/synchronization/internal/per_thread_sem.h"
+
+#include <atomic>
+
+#include "absl/base/attributes.h"
+#include "absl/base/internal/malloc_extension.h"
+#include "absl/base/internal/thread_identity.h"
+#include "absl/synchronization/internal/waiter.h"
+
+namespace absl {
+namespace synchronization_internal {
+
+void PerThreadSem::SetThreadBlockedCounter(std::atomic<int> *counter) {
+  base_internal::ThreadIdentity *identity;
+  identity = GetOrCreateCurrentThreadIdentity();
+  identity->blocked_count_ptr = counter;
+}
+
+std::atomic<int> *PerThreadSem::GetThreadBlockedCounter() {
+  base_internal::ThreadIdentity *identity;
+  identity = GetOrCreateCurrentThreadIdentity();
+  return identity->blocked_count_ptr;
+}
+
+void PerThreadSem::Init(base_internal::ThreadIdentity *identity) {
+  Waiter::GetWaiter(identity)->Init();
+  identity->ticker.store(0, std::memory_order_relaxed);
+  identity->wait_start.store(0, std::memory_order_relaxed);
+  identity->is_idle.store(false, std::memory_order_relaxed);
+}
+
+void PerThreadSem::Tick(base_internal::ThreadIdentity *identity) {
+  const int ticker =
+      identity->ticker.fetch_add(1, std::memory_order_relaxed) + 1;
+  const int wait_start = identity->wait_start.load(std::memory_order_relaxed);
+  const bool is_idle = identity->is_idle.load(std::memory_order_relaxed);
+  if (wait_start && (ticker - wait_start > Waiter::kIdlePeriods) && !is_idle) {
+    // Wakeup the waiting thread since it is time for it to become idle.
+    Waiter::GetWaiter(identity)->Poke();
+  }
+}
+
+}  // namespace synchronization_internal
+}  // namespace absl
+
+extern "C" {
+
+ABSL_ATTRIBUTE_WEAK void AbslInternalPerThreadSemPost(
+    absl::base_internal::ThreadIdentity *identity) {
+  absl::synchronization_internal::Waiter::GetWaiter(identity)->Post();
+}
+
+ABSL_ATTRIBUTE_WEAK bool AbslInternalPerThreadSemWait(
+    absl::synchronization_internal::KernelTimeout t) {
+  bool timeout = false;
+  absl::base_internal::ThreadIdentity *identity;
+  identity = absl::synchronization_internal::GetOrCreateCurrentThreadIdentity();
+
+  // Ensure wait_start != 0.
+  int ticker = identity->ticker.load(std::memory_order_relaxed);
+  identity->wait_start.store(ticker ? ticker : 1, std::memory_order_relaxed);
+  identity->is_idle.store(false, std::memory_order_relaxed);
+
+  if (identity->blocked_count_ptr != nullptr) {
+    // Increment count of threads blocked in a given thread pool.
+    identity->blocked_count_ptr->fetch_add(1, std::memory_order_relaxed);
+  }
+
+  timeout =
+      !absl::synchronization_internal::Waiter::GetWaiter(identity)->Wait(t);
+
+  if (identity->blocked_count_ptr != nullptr) {
+    identity->blocked_count_ptr->fetch_sub(1, std::memory_order_relaxed);
+  }
+
+  if (identity->is_idle.load(std::memory_order_relaxed)) {
+    // We became idle during the wait; become non-idle again so that
+    // performance of deallocations done from now on does not suffer.
+    absl::base_internal::MallocExtension::instance()->MarkThreadBusy();
+  }
+  identity->is_idle.store(false, std::memory_order_relaxed);
+  identity->wait_start.store(0, std::memory_order_relaxed);
+  return !timeout;
+}
+
+}  // extern "C"
+
+#endif  // ABSL_LOW_LEVEL_ALLOC_MISSING
diff --git a/absl/synchronization/internal/per_thread_sem.h b/absl/synchronization/internal/per_thread_sem.h
new file mode 100644
index 000000000000..678b69e487a5
--- /dev/null
+++ b/absl/synchronization/internal/per_thread_sem.h
@@ -0,0 +1,107 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+
+// PerThreadSem is a low-level synchronization primitive controlling the
+// runnability of a single thread, used internally by Mutex and CondVar.
+//
+// This is NOT a general-purpose synchronization mechanism, and should not be
+// used directly by applications.  Applications should use Mutex and CondVar.
+//
+// The semantics of PerThreadSem are the same as that of a counting semaphore.
+// Each thread maintains an abstract "count" value associated with its identity.
+
+#ifndef ABSL_SYNCHRONIZATION_INTERNAL_PER_THREAD_SEM_H_
+#define ABSL_SYNCHRONIZATION_INTERNAL_PER_THREAD_SEM_H_
+
+#include <atomic>
+
+#include "absl/base/internal/thread_identity.h"
+#include "absl/synchronization/internal/create_thread_identity.h"
+#include "absl/synchronization/internal/kernel_timeout.h"
+
+namespace absl {
+
+class Mutex;
+
+namespace synchronization_internal {
+
+class PerThreadSem {
+ public:
+  PerThreadSem() = delete;
+  PerThreadSem(const PerThreadSem&) = delete;
+  PerThreadSem& operator=(const PerThreadSem&) = delete;
+
+  // Routine invoked periodically (once a second) by a background thread.
+  // Has no effect on user-visible state.
+  static void Tick(base_internal::ThreadIdentity* identity);
+
+  // ---------------------------------------------------------------------------
+  // Routines used by autosizing threadpools to detect when threads are
+  // blocked.  Each thread has a counter pointer, initially zero.  If non-zero,
+  // the implementation atomically increments the counter when it blocks on a
+  // semaphore, a decrements it again when it wakes.  This allows a threadpool
+  // to keep track of how many of its threads are blocked.
+  // SetThreadBlockedCounter() should be used only by threadpool
+  // implementations.  GetThreadBlockedCounter() should be used by modules that
+  // block threads; if the pointer returned is non-zero, the location should be
+  // incremented before the thread blocks, and decremented after it wakes.
+  static void SetThreadBlockedCounter(std::atomic<int> *counter);
+  static std::atomic<int> *GetThreadBlockedCounter();
+
+ private:
+  // Create the PerThreadSem associated with "identity".  Initializes count=0.
+  // REQUIRES: May only be called by ThreadIdentity.
+  static void Init(base_internal::ThreadIdentity* identity);
+
+  // Increments "identity"'s count.
+  static inline void Post(base_internal::ThreadIdentity* identity);
+
+  // Waits until either our count > 0 or t has expired.
+  // If count > 0, decrements count and returns true.  Otherwise returns false.
+  // !t.has_timeout() => Wait(t) will return true.
+  static inline bool Wait(KernelTimeout t);
+
+  // White-listed callers.
+  friend class PerThreadSemTest;
+  friend class absl::Mutex;
+  friend absl::base_internal::ThreadIdentity* CreateThreadIdentity();
+};
+
+}  // namespace synchronization_internal
+}  // namespace absl
+
+// In some build configurations we pass --detect-odr-violations to the
+// gold linker.  This causes it to flag weak symbol overrides as ODR
+// violations.  Because ODR only applies to C++ and not C,
+// --detect-odr-violations ignores symbols not mangled with C++ names.
+// By changing our extension points to be extern "C", we dodge this
+// check.
+extern "C" {
+void AbslInternalPerThreadSemPost(
+    absl::base_internal::ThreadIdentity* identity);
+bool AbslInternalPerThreadSemWait(
+    absl::synchronization_internal::KernelTimeout t);
+}  // extern "C"
+
+void absl::synchronization_internal::PerThreadSem::Post(
+    absl::base_internal::ThreadIdentity* identity) {
+  AbslInternalPerThreadSemPost(identity);
+}
+
+bool absl::synchronization_internal::PerThreadSem::Wait(
+    absl::synchronization_internal::KernelTimeout t) {
+  return AbslInternalPerThreadSemWait(t);
+}
+#endif  // ABSL_SYNCHRONIZATION_INTERNAL_PER_THREAD_SEM_H_
diff --git a/absl/synchronization/internal/per_thread_sem_test.cc b/absl/synchronization/internal/per_thread_sem_test.cc
new file mode 100644
index 000000000000..1d072a79e3d4
--- /dev/null
+++ b/absl/synchronization/internal/per_thread_sem_test.cc
@@ -0,0 +1,246 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/synchronization/internal/per_thread_sem.h"
+
+#include <atomic>
+#include <condition_variable>  // NOLINT(build/c++11)
+#include <functional>
+#include <limits>
+#include <mutex>               // NOLINT(build/c++11)
+#include <string>
+#include <thread>              // NOLINT(build/c++11)
+
+#include "absl/base/internal/cycleclock.h"
+#include "absl/base/internal/malloc_extension.h"
+#include "absl/base/internal/thread_identity.h"
+#include "absl/strings/str_cat.h"
+#include "absl/time/clock.h"
+#include "absl/time/time.h"
+#include "gtest/gtest.h"
+
+// In this test we explicitly avoid the use of synchronization
+// primitives which might use PerThreadSem, most notably absl::Mutex.
+
+namespace absl {
+namespace synchronization_internal {
+
+class SimpleSemaphore {
+ public:
+  SimpleSemaphore() : count_(0) {}
+
+  // Decrements (locks) the semaphore. If the semaphore's value is
+  // greater than zero, then the decrement proceeds, and the function
+  // returns, immediately. If the semaphore currently has the value
+  // zero, then the call blocks until it becomes possible to perform
+  // the decrement.
+  void Wait() {
+    std::unique_lock<std::mutex> lock(mu_);
+    cv_.wait(lock, [this]() { return count_ > 0; });
+    --count_;
+    cv_.notify_one();
+  }
+
+  // Increments (unlocks) the semaphore. If the semaphore's value
+  // consequently becomes greater than zero, then another thread
+  // blocked Wait() call will be woken up and proceed to lock the
+  // semaphore.
+  void Post() {
+    std::lock_guard<std::mutex> lock(mu_);
+    ++count_;
+    cv_.notify_one();
+  }
+
+ private:
+  std::mutex mu_;
+  std::condition_variable cv_;
+  int count_;
+};
+
+struct ThreadData {
+  int num_iterations;                 // Number of replies to send.
+  SimpleSemaphore identity2_written;  // Posted by thread writing identity2.
+  base_internal::ThreadIdentity *identity1;  // First Post()-er.
+  base_internal::ThreadIdentity *identity2;  // First Wait()-er.
+  KernelTimeout timeout;
+};
+
+// Need friendship with PerThreadSem.
+class PerThreadSemTest : public testing::Test {
+ public:
+  static void TimingThread(ThreadData* t) {
+    t->identity2 = GetOrCreateCurrentThreadIdentity();
+    t->identity2_written.Post();
+    while (t->num_iterations--) {
+      Wait(t->timeout);
+      Post(t->identity1);
+    }
+  }
+
+  void TestTiming(const char *msg, bool timeout) {
+    static const int kNumIterations = 100;
+    ThreadData t;
+    t.num_iterations = kNumIterations;
+    t.timeout = timeout ?
+        KernelTimeout(absl::Now() + absl::Seconds(10000))  // far in the future
+        : KernelTimeout::Never();
+    t.identity1 = GetOrCreateCurrentThreadIdentity();
+
+    // We can't use the Thread class here because it uses the Mutex
+    // class which will invoke PerThreadSem, so we use std::thread instead.
+    std::thread partner_thread(std::bind(TimingThread, &t));
+
+    // Wait for our partner thread to register their identity.
+    t.identity2_written.Wait();
+
+    int64_t min_cycles = std::numeric_limits<int64_t>::max();
+    int64_t total_cycles = 0;
+    for (int i = 0; i < kNumIterations; ++i) {
+      absl::SleepFor(absl::Milliseconds(20));
+      int64_t cycles = base_internal::CycleClock::Now();
+      Post(t.identity2);
+      Wait(t.timeout);
+      cycles = base_internal::CycleClock::Now() - cycles;
+      min_cycles = std::min(min_cycles, cycles);
+      total_cycles += cycles;
+    }
+    std::string out =
+        StrCat(msg, "min cycle count=", min_cycles, " avg cycle count=",
+               absl::SixDigits(static_cast<double>(total_cycles) /
+                               kNumIterations));
+    printf("%s\n", out.c_str());
+
+    partner_thread.join();
+  }
+
+ protected:
+  static void Post(base_internal::ThreadIdentity *id) {
+    PerThreadSem::Post(id);
+  }
+  static bool Wait(KernelTimeout t) {
+    return PerThreadSem::Wait(t);
+  }
+
+  // convenience overload
+  static bool Wait(absl::Time t) {
+    return Wait(KernelTimeout(t));
+  }
+
+  static void Tick(base_internal::ThreadIdentity *identity) {
+    PerThreadSem::Tick(identity);
+  }
+};
+
+namespace {
+
+TEST_F(PerThreadSemTest, WithoutTimeout) {
+  PerThreadSemTest::TestTiming("Without timeout: ", false);
+}
+
+TEST_F(PerThreadSemTest, WithTimeout) {
+  PerThreadSemTest::TestTiming("With timeout:    ", true);
+}
+
+TEST_F(PerThreadSemTest, Timeouts) {
+  absl::Time timeout = absl::Now() + absl::Milliseconds(50);
+  EXPECT_FALSE(Wait(timeout));
+  EXPECT_LE(timeout, absl::Now());
+
+  absl::Time negative_timeout = absl::UnixEpoch() - absl::Milliseconds(100);
+  EXPECT_FALSE(Wait(negative_timeout));
+  EXPECT_LE(negative_timeout, absl::Now());  // trivially true :)
+
+  Post(GetOrCreateCurrentThreadIdentity());
+  // The wait here has an expired timeout, but we have a wake to consume,
+  // so this should succeed
+  EXPECT_TRUE(Wait(negative_timeout));
+}
+
+// Test that idle threads properly register themselves as such with malloc.
+TEST_F(PerThreadSemTest, Idle) {
+  // We can't use gmock because it might use synch calls.  So we do it
+  // by hand, messily.  I don't bother hitting every one of the
+  // MallocExtension calls because most of them won't get made
+  // anyway--if they do we can add them.
+  class MockMallocExtension : public base_internal::MallocExtension {
+   public:
+    MockMallocExtension(base_internal::MallocExtension *real,
+                        base_internal::ThreadIdentity *id,
+                        std::atomic<int> *idles, std::atomic<int> *busies)
+        : real_(real), id_(id), idles_(idles), busies_(busies) {}
+    void MarkThreadIdle() override {
+      if (base_internal::CurrentThreadIdentityIfPresent() != id_) {
+        return;
+      }
+      idles_->fetch_add(1, std::memory_order_relaxed);
+    }
+
+    void MarkThreadBusy() override {
+      if (base_internal::CurrentThreadIdentityIfPresent() != id_) {
+        return;
+      }
+      busies_->fetch_add(1, std::memory_order_relaxed);
+    }
+    size_t GetAllocatedSize(const void* p) override {
+      return real_->GetAllocatedSize(p);
+    }
+
+   private:
+    MallocExtension *real_;
+    base_internal::ThreadIdentity *id_;
+    std::atomic<int>* idles_;
+    std::atomic<int>* busies_;
+  };
+
+  base_internal::ThreadIdentity *id = GetOrCreateCurrentThreadIdentity();
+  std::atomic<int> idles(0);
+  std::atomic<int> busies(0);
+  base_internal::MallocExtension *old =
+      base_internal::MallocExtension::instance();
+  MockMallocExtension mock(old, id, &idles, &busies);
+  base_internal::MallocExtension::Register(&mock);
+  std::atomic<int> sync(0);
+
+  std::thread t([id, &idles, &sync]() {
+    // Wait for the main thread to begin the wait process
+    while (0 == sync.load(std::memory_order_relaxed)) {
+      SleepFor(absl::Milliseconds(1));
+    }
+    // Wait for main thread to become idle, then wake it
+    // pretend time is passing--enough of these should cause an idling.
+    for (int i = 0; i < 100; ++i) {
+      Tick(id);
+    }
+    while (0 == idles.load(std::memory_order_relaxed)) {
+      // Keep ticking, just in case.
+      Tick(id);
+      SleepFor(absl::Milliseconds(1));
+    }
+    Post(id);
+  });
+
+  idles.store(0, std::memory_order_relaxed);  // In case we slept earlier.
+  sync.store(1, std::memory_order_relaxed);
+  Wait(KernelTimeout::Never());
+
+  // t will wake us once we become idle.
+  EXPECT_LT(0, busies.load(std::memory_order_relaxed));
+  t.join();
+  base_internal::MallocExtension::Register(old);
+}
+
+}  // namespace
+
+}  // namespace synchronization_internal
+}  // namespace absl
diff --git a/absl/synchronization/internal/thread_pool.h b/absl/synchronization/internal/thread_pool.h
new file mode 100644
index 000000000000..846404277a03
--- /dev/null
+++ b/absl/synchronization/internal/thread_pool.h
@@ -0,0 +1,90 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#ifndef ABSL_SYNCHRONIZATION_INTERNAL_THREAD_POOL_H_
+#define ABSL_SYNCHRONIZATION_INTERNAL_THREAD_POOL_H_
+
+#include <cassert>
+#include <functional>
+#include <queue>
+#include <thread>  // NOLINT(build/c++11)
+#include <vector>
+
+#include "absl/base/thread_annotations.h"
+#include "absl/synchronization/mutex.h"
+
+namespace absl {
+namespace synchronization_internal {
+
+// A simple ThreadPool implementation for tests.
+class ThreadPool {
+ public:
+  explicit ThreadPool(int num_threads) {
+    for (int i = 0; i < num_threads; ++i) {
+      threads_.push_back(std::thread(&ThreadPool::WorkLoop, this));
+    }
+  }
+
+  ThreadPool(const ThreadPool &) = delete;
+  ThreadPool &operator=(const ThreadPool &) = delete;
+
+  ~ThreadPool() {
+    {
+      absl::MutexLock l(&mu_);
+      for (int i = 0; i < threads_.size(); ++i) {
+        queue_.push(nullptr);  // Shutdown signal.
+      }
+    }
+    for (auto &t : threads_) {
+      t.join();
+    }
+  }
+
+  // Schedule a function to be run on a ThreadPool thread immediately.
+  void Schedule(std::function<void()> func) {
+    assert(func != nullptr);
+    absl::MutexLock l(&mu_);
+    queue_.push(std::move(func));
+  }
+
+ private:
+  bool WorkAvailable() const EXCLUSIVE_LOCKS_REQUIRED(mu_) {
+    return !queue_.empty();
+  }
+
+  void WorkLoop() {
+    while (true) {
+      std::function<void()> func;
+      {
+        absl::MutexLock l(&mu_);
+        mu_.Await(absl::Condition(this, &ThreadPool::WorkAvailable));
+        func = std::move(queue_.front());
+        queue_.pop();
+      }
+      if (func == nullptr) {  // Shutdown signal.
+        break;
+      }
+      func();
+    }
+  }
+
+  absl::Mutex mu_;
+  std::queue<std::function<void()>> queue_ GUARDED_BY(mu_);
+  std::vector<std::thread> threads_;
+};
+
+}  // namespace synchronization_internal
+}  // namespace absl
+
+#endif  // ABSL_SYNCHRONIZATION_INTERNAL_THREAD_POOL_H_
diff --git a/absl/synchronization/internal/waiter.cc b/absl/synchronization/internal/waiter.cc
new file mode 100644
index 000000000000..cd16c7887c08
--- /dev/null
+++ b/absl/synchronization/internal/waiter.cc
@@ -0,0 +1,394 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/synchronization/internal/waiter.h"
+
+#include "absl/base/config.h"
+
+#ifdef _WIN32
+#include <windows.h>
+#else
+#include <pthread.h>
+#include <sys/time.h>
+#include <unistd.h>
+#endif
+
+#ifdef __linux__
+#include <linux/futex.h>
+#include <sys/syscall.h>
+#endif
+
+#ifdef ABSL_HAVE_SEMAPHORE_H
+#include <semaphore.h>
+#endif
+
+#include <errno.h>
+#include <stdio.h>
+#include <time.h>
+
+#include <atomic>
+#include <cassert>
+
+#include "absl/base/internal/malloc_extension.h"
+#include "absl/base/internal/raw_logging.h"
+#include "absl/base/internal/thread_identity.h"
+#include "absl/synchronization/internal/kernel_timeout.h"
+
+namespace absl {
+namespace synchronization_internal {
+
+static void MaybeBecomeIdle() {
+  base_internal::ThreadIdentity *identity =
+      base_internal::CurrentThreadIdentityIfPresent();
+  assert(identity != nullptr);
+  const bool is_idle = identity->is_idle.load(std::memory_order_relaxed);
+  const int ticker = identity->ticker.load(std::memory_order_relaxed);
+  const int wait_start = identity->wait_start.load(std::memory_order_relaxed);
+  if (!is_idle && ticker - wait_start > Waiter::kIdlePeriods) {
+    identity->is_idle.store(true, std::memory_order_relaxed);
+    base_internal::MallocExtension::instance()->MarkThreadIdle();
+  }
+}
+
+#if ABSL_WAITER_MODE == ABSL_WAITER_MODE_FUTEX
+
+// Some Android headers are missing these definitions even though they
+// support these futex operations.
+#ifdef __BIONIC__
+#ifndef SYS_futex
+#define SYS_futex __NR_futex
+#endif
+#ifndef FUTEX_WAIT_BITSET
+#define FUTEX_WAIT_BITSET 9
+#endif
+#ifndef FUTEX_PRIVATE_FLAG
+#define FUTEX_PRIVATE_FLAG 128
+#endif
+#ifndef FUTEX_CLOCK_REALTIME
+#define FUTEX_CLOCK_REALTIME 256
+#endif
+#ifndef FUTEX_BITSET_MATCH_ANY
+#define FUTEX_BITSET_MATCH_ANY 0xFFFFFFFF
+#endif
+#endif
+
+void Waiter::Init() {
+  futex_.store(0, std::memory_order_relaxed);
+}
+
+bool Waiter::Wait(KernelTimeout t) {
+  // Loop until we can atomically decrement futex from a positive
+  // value, waiting on a futex while we believe it is zero.
+  while (true) {
+    int x = futex_.load(std::memory_order_relaxed);
+    if (x != 0) {
+      if (!futex_.compare_exchange_weak(x, x - 1,
+                                        std::memory_order_acquire,
+                                        std::memory_order_relaxed)) {
+        continue;  // Raced with someone, retry.
+      }
+      return true;  // Consumed a wakeup, we are done.
+    }
+
+    int err = 0;
+    if (t.has_timeout()) {
+      // https://locklessinc.com/articles/futex_cheat_sheet/
+      // Unlike FUTEX_WAIT, FUTEX_WAIT_BITSET uses absolute time.
+      struct timespec abs_timeout = t.MakeAbsTimespec();
+      // Atomically check that the futex value is still 0, and if it
+      // is, sleep until abs_timeout or until woken by FUTEX_WAKE.
+      err = syscall(
+          SYS_futex, reinterpret_cast<int *>(&futex_),
+          FUTEX_WAIT_BITSET | FUTEX_PRIVATE_FLAG | FUTEX_CLOCK_REALTIME, 0,
+          &abs_timeout, nullptr, FUTEX_BITSET_MATCH_ANY);
+    } else {
+      // Atomically check that the futex value is still 0, and if it
+      // is, sleep until woken by FUTEX_WAKE.
+      err = syscall(SYS_futex, reinterpret_cast<int *>(&futex_),
+                    FUTEX_WAIT | FUTEX_PRIVATE_FLAG, 0, nullptr);
+    }
+    if (err != 0) {
+      if (errno == EINTR || errno == EWOULDBLOCK) {
+        // Do nothing, the loop will retry.
+      } else if (errno == ETIMEDOUT) {
+        return false;  // Timeout.
+      } else {
+        ABSL_RAW_LOG(FATAL, "Futex operation failed with errno %d\n", errno);
+      }
+    }
+
+    MaybeBecomeIdle();
+  }
+}
+
+void Waiter::Post() {
+  if (futex_.fetch_add(1, std::memory_order_release) == 0) {
+    // We incremented from 0, need to wake a potential waker.
+    Poke();
+  }
+}
+
+void Waiter::Poke() {
+  // Wake one thread waiting on the futex.
+  int err = syscall(SYS_futex, reinterpret_cast<int *>(&futex_),
+                    FUTEX_WAKE | FUTEX_PRIVATE_FLAG, 1);
+  if (err < 0) {
+    ABSL_RAW_LOG(FATAL, "FUTEX_WAKE failed with errno %d\n", errno);
+  }
+}
+
+#elif ABSL_WAITER_MODE == ABSL_WAITER_MODE_CONDVAR
+
+class PthreadMutexHolder {
+ public:
+  explicit PthreadMutexHolder(pthread_mutex_t *mu) : mu_(mu) {
+    const int err = pthread_mutex_lock(mu_);
+    if (err != 0) {
+      ABSL_RAW_LOG(FATAL, "pthread_mutex_lock failed: %d", err);
+    }
+  }
+
+  PthreadMutexHolder(const PthreadMutexHolder &rhs) = delete;
+  PthreadMutexHolder &operator=(const PthreadMutexHolder &rhs) = delete;
+
+  ~PthreadMutexHolder() {
+    const int err = pthread_mutex_unlock(mu_);
+    if (err != 0) {
+      ABSL_RAW_LOG(FATAL, "pthread_mutex_unlock failed: %d", err);
+    }
+  }
+
+ private:
+  pthread_mutex_t *mu_;
+};
+
+void Waiter::Init() {
+  const int err = pthread_mutex_init(&mu_, 0);
+  if (err != 0) {
+    ABSL_RAW_LOG(FATAL, "pthread_mutex_init failed: %d", err);
+  }
+
+  const int err2 = pthread_cond_init(&cv_, 0);
+  if (err2 != 0) {
+    ABSL_RAW_LOG(FATAL, "pthread_cond_init failed: %d", err2);
+  }
+
+  waiter_count_.store(0, std::memory_order_relaxed);
+  wakeup_count_.store(0, std::memory_order_relaxed);
+}
+
+bool Waiter::Wait(KernelTimeout t) {
+  struct timespec abs_timeout;
+  if (t.has_timeout()) {
+    abs_timeout = t.MakeAbsTimespec();
+  }
+
+  PthreadMutexHolder h(&mu_);
+  waiter_count_.fetch_add(1, std::memory_order_relaxed);
+  // Loop until we find a wakeup to consume or timeout.
+  while (true) {
+    int x = wakeup_count_.load(std::memory_order_relaxed);
+    if (x != 0) {
+      if (!wakeup_count_.compare_exchange_weak(x, x - 1,
+                                               std::memory_order_acquire,
+                                               std::memory_order_relaxed)) {
+        continue;  // Raced with someone, retry.
+      }
+      // Successfully consumed a wakeup, we're done.
+      waiter_count_.fetch_sub(1, std::memory_order_relaxed);
+      return true;
+    }
+
+    // No wakeups available, time to wait.
+    if (!t.has_timeout()) {
+      const int err = pthread_cond_wait(&cv_, &mu_);
+      if (err != 0) {
+        ABSL_RAW_LOG(FATAL, "pthread_cond_wait failed: %d", err);
+      }
+    } else {
+      const int err = pthread_cond_timedwait(&cv_, &mu_, &abs_timeout);
+      if (err == ETIMEDOUT) {
+        waiter_count_.fetch_sub(1, std::memory_order_relaxed);
+        return false;
+      }
+      if (err != 0) {
+        ABSL_RAW_LOG(FATAL, "pthread_cond_wait failed: %d", err);
+      }
+    }
+    MaybeBecomeIdle();
+  }
+}
+
+void Waiter::Post() {
+  wakeup_count_.fetch_add(1, std::memory_order_release);
+  Poke();
+}
+
+void Waiter::Poke() {
+  if (waiter_count_.load(std::memory_order_relaxed) == 0) {
+    return;
+  }
+  // Potentially a waker. Take the lock and check again.
+  PthreadMutexHolder h(&mu_);
+  if (waiter_count_.load(std::memory_order_relaxed) == 0) {
+    return;
+  }
+  const int err = pthread_cond_signal(&cv_);
+  if (err != 0) {
+    ABSL_RAW_LOG(FATAL, "pthread_cond_signal failed: %d", err);
+  }
+}
+
+#elif ABSL_WAITER_MODE == ABSL_WAITER_MODE_SEM
+
+void Waiter::Init() {
+  if (sem_init(&sem_, 0, 0) != 0) {
+    ABSL_RAW_LOG(FATAL, "sem_init failed with errno %d\n", errno);
+  }
+  wakeups_.store(0, std::memory_order_relaxed);
+}
+
+bool Waiter::Wait(KernelTimeout t) {
+  struct timespec abs_timeout;
+  if (t.has_timeout()) {
+    abs_timeout = t.MakeAbsTimespec();
+  }
+
+  // Loop until we timeout or consume a wakeup.
+  while (true) {
+    int x = wakeups_.load(std::memory_order_relaxed);
+    if (x != 0) {
+      if (!wakeups_.compare_exchange_weak(x, x - 1,
+                                          std::memory_order_acquire,
+                                          std::memory_order_relaxed)) {
+        continue;  // Raced with someone, retry.
+      }
+      // Successfully consumed a wakeup, we're done.
+      return true;
+    }
+
+    // Nothing to consume, wait (looping on EINTR).
+    while (true) {
+      if (!t.has_timeout()) {
+        if (sem_wait(&sem_) == 0) break;
+        if (errno == EINTR) continue;
+        ABSL_RAW_LOG(FATAL, "sem_wait failed: %d", errno);
+      } else {
+        if (sem_timedwait(&sem_, &abs_timeout) == 0) break;
+        if (errno == EINTR) continue;
+        if (errno == ETIMEDOUT) return false;
+        ABSL_RAW_LOG(FATAL, "sem_timedwait failed: %d", errno);
+      }
+    }
+    MaybeBecomeIdle();
+  }
+}
+
+void Waiter::Post() {
+  wakeups_.fetch_add(1, std::memory_order_release);  // Post a wakeup.
+  Poke();
+}
+
+void Waiter::Poke() {
+  if (sem_post(&sem_) != 0) {  // Wake any semaphore waiter.
+    ABSL_RAW_LOG(FATAL, "sem_post failed with errno %d\n", errno);
+  }
+}
+
+#elif ABSL_WAITER_MODE == ABSL_WAITER_MODE_WIN32
+
+class LockHolder {
+ public:
+  explicit LockHolder(SRWLOCK* mu) : mu_(mu) {
+    AcquireSRWLockExclusive(mu_);
+  }
+
+  LockHolder(const LockHolder&) = delete;
+  LockHolder& operator=(const LockHolder&) = delete;
+
+  ~LockHolder() {
+    ReleaseSRWLockExclusive(mu_);
+  }
+
+ private:
+  SRWLOCK* mu_;
+};
+
+void Waiter::Init() {
+  InitializeSRWLock(&mu_);
+  InitializeConditionVariable(&cv_);
+  waiter_count_.store(0, std::memory_order_relaxed);
+  wakeup_count_.store(0, std::memory_order_relaxed);
+}
+
+bool Waiter::Wait(KernelTimeout t) {
+  LockHolder h(&mu_);
+  waiter_count_.fetch_add(1, std::memory_order_relaxed);
+
+  // Loop until we find a wakeup to consume or timeout.
+  while (true) {
+    int x = wakeup_count_.load(std::memory_order_relaxed);
+    if (x != 0) {
+      if (!wakeup_count_.compare_exchange_weak(x, x - 1,
+                                               std::memory_order_acquire,
+                                               std::memory_order_relaxed)) {
+        continue;  // Raced with someone, retry.
+      }
+      // Successfully consumed a wakeup, we're done.
+      waiter_count_.fetch_sub(1, std::memory_order_relaxed);
+      return true;
+    }
+
+    // No wakeups available, time to wait.
+    if (!SleepConditionVariableSRW(
+            &cv_, &mu_, t.InMillisecondsFromNow(), 0)) {
+      // GetLastError() returns a Win32 DWORD, but we assign to
+      // unsigned long to simplify the ABSL_RAW_LOG case below.  The uniform
+      // initialization guarantees this is not a narrowing conversion.
+      const unsigned long err{GetLastError()};  // NOLINT(runtime/int)
+      if (err == ERROR_TIMEOUT) {
+        waiter_count_.fetch_sub(1, std::memory_order_relaxed);
+        return false;
+      } else {
+        ABSL_RAW_LOG(FATAL, "SleepConditionVariableSRW failed: %lu", err);
+      }
+    }
+
+    MaybeBecomeIdle();
+  }
+}
+
+void Waiter::Post() {
+  wakeup_count_.fetch_add(1, std::memory_order_release);
+  Poke();
+}
+
+void Waiter::Poke() {
+  if (waiter_count_.load(std::memory_order_relaxed) == 0) {
+    return;
+  }
+  // Potentially a waker. Take the lock and check again.
+  LockHolder h(&mu_);
+  if (waiter_count_.load(std::memory_order_relaxed) == 0) {
+    return;
+  }
+  WakeConditionVariable(&cv_);
+}
+
+#else
+#error Unknown ABSL_WAITER_MODE
+#endif
+
+}  // namespace synchronization_internal
+}  // namespace absl
diff --git a/absl/synchronization/internal/waiter.h b/absl/synchronization/internal/waiter.h
new file mode 100644
index 000000000000..025ace42e261
--- /dev/null
+++ b/absl/synchronization/internal/waiter.h
@@ -0,0 +1,138 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+
+#ifndef ABSL_SYNCHRONIZATION_INTERNAL_WAITER_H_
+#define ABSL_SYNCHRONIZATION_INTERNAL_WAITER_H_
+
+#include "absl/base/config.h"
+
+#ifdef _WIN32
+#include <windows.h>
+#else
+#include <pthread.h>
+#endif
+
+#ifdef ABSL_HAVE_SEMAPHORE_H
+#include <semaphore.h>
+#endif
+
+#include <atomic>
+
+#include "absl/base/internal/thread_identity.h"
+#include "absl/synchronization/internal/kernel_timeout.h"
+
+// May be chosen at compile time via -DABSL_FORCE_WAITER_MODE=<index>
+#define ABSL_WAITER_MODE_FUTEX 0
+#define ABSL_WAITER_MODE_SEM 1
+#define ABSL_WAITER_MODE_CONDVAR 2
+#define ABSL_WAITER_MODE_WIN32 3
+
+#if defined(ABSL_FORCE_WAITER_MODE)
+#define ABSL_WAITER_MODE ABSL_FORCE_WAITER_MODE
+#elif defined(_WIN32)
+#define ABSL_WAITER_MODE ABSL_WAITER_MODE_WIN32
+#elif defined(__linux__)
+#define ABSL_WAITER_MODE ABSL_WAITER_MODE_FUTEX
+#elif defined(ABSL_HAVE_SEMAPHORE_H)
+#define ABSL_WAITER_MODE ABSL_WAITER_MODE_SEM
+#else
+#define ABSL_WAITER_MODE ABSL_WAITER_MODE_CONDVAR
+#endif
+
+namespace absl {
+namespace synchronization_internal {
+
+// Waiter is an OS-specific semaphore.
+class Waiter {
+ public:
+  // No constructor, instances use the reserved space in ThreadIdentity.
+  // All initialization logic belongs in `Init()`.
+  Waiter() = delete;
+  Waiter(const Waiter&) = delete;
+  Waiter& operator=(const Waiter&) = delete;
+
+  // Prepare any data to track waits.
+  void Init();
+
+  // Blocks the calling thread until a matching call to `Post()` or
+  // `t` has passed. Returns `true` if woken (`Post()` called),
+  // `false` on timeout.
+  bool Wait(KernelTimeout t);
+
+  // Restart the caller of `Wait()` as with a normal semaphore.
+  void Post();
+
+  // If anyone is waiting, wake them up temporarily and cause them to
+  // call `MaybeBecomeIdle()`. They will then return to waiting for a
+  // `Post()` or timeout.
+  void Poke();
+
+  // Returns the Waiter associated with the identity.
+  static Waiter* GetWaiter(base_internal::ThreadIdentity* identity) {
+    static_assert(
+        sizeof(Waiter) <= sizeof(base_internal::ThreadIdentity::WaiterState),
+        "Insufficient space for Waiter");
+    return reinterpret_cast<Waiter*>(identity->waiter_state.data);
+  }
+
+  // How many periods to remain idle before releasing resources
+#ifndef THREAD_SANITIZER
+  static const int kIdlePeriods = 60;
+#else
+  // Memory consumption under ThreadSanitizer is a serious concern,
+  // so we release resources sooner. The value of 1 leads to 1 to 2 second
+  // delay before marking a thread as idle.
+  static const int kIdlePeriods = 1;
+#endif
+
+ private:
+#if ABSL_WAITER_MODE == ABSL_WAITER_MODE_FUTEX
+  // Futexes are defined by specification to be ints.
+  // Thus std::atomic<int> must be just an int with lockfree methods.
+  std::atomic<int> futex_;
+  static_assert(sizeof(int) == sizeof(futex_), "Wrong size for futex");
+
+#elif ABSL_WAITER_MODE == ABSL_WAITER_MODE_CONDVAR
+  pthread_mutex_t mu_;
+  pthread_cond_t cv_;
+  std::atomic<int> waiter_count_;
+  std::atomic<int> wakeup_count_;  // Unclaimed wakeups, written under lock.
+
+#elif ABSL_WAITER_MODE == ABSL_WAITER_MODE_SEM
+  sem_t sem_;
+  // This seems superfluous, but for Poke() we need to cause spurious
+  // wakeups on the semaphore. Hence we can't actually use the
+  // semaphore's count.
+  std::atomic<int> wakeups_;
+
+#elif ABSL_WAITER_MODE == ABSL_WAITER_MODE_WIN32
+  // The Windows API has lots of choices for synchronization
+  // primivitives.  We are using SRWLOCK and CONDITION_VARIABLE
+  // because they don't require a destructor to release system
+  // resources.
+  SRWLOCK mu_;
+  CONDITION_VARIABLE cv_;
+  std::atomic<int> waiter_count_;
+  std::atomic<int> wakeup_count_;
+
+#else
+  #error Unknown ABSL_WAITER_MODE
+#endif
+};
+
+}  // namespace synchronization_internal
+}  // namespace absl
+
+#endif  // ABSL_SYNCHRONIZATION_INTERNAL_WAITER_H_
diff --git a/absl/synchronization/mutex.cc b/absl/synchronization/mutex.cc
new file mode 100644
index 000000000000..cb0a3a10a012
--- /dev/null
+++ b/absl/synchronization/mutex.cc
@@ -0,0 +1,2680 @@
+#include "absl/synchronization/mutex.h"
+
+#ifdef _WIN32
+#include <windows.h>
+#ifdef ERROR
+#undef ERROR
+#endif
+#else
+#include <fcntl.h>
+#include <pthread.h>
+#include <sched.h>
+#include <sys/time.h>
+#endif
+
+#include <assert.h>
+#include <errno.h>
+#include <stdio.h>
+#include <stdlib.h>
+#include <string.h>
+#include <time.h>
+
+#include <algorithm>
+#include <atomic>
+#include <cinttypes>
+#include <thread>  // NOLINT(build/c++11)
+
+#include "absl/base/attributes.h"
+#include "absl/base/config.h"
+#include "absl/base/dynamic_annotations.h"
+#include "absl/base/internal/atomic_hook.h"
+#include "absl/base/internal/cycleclock.h"
+#include "absl/base/internal/low_level_alloc.h"
+#include "absl/base/internal/raw_logging.h"
+#include "absl/base/internal/spinlock.h"
+#include "absl/base/internal/sysinfo.h"
+#include "absl/base/internal/thread_identity.h"
+#include "absl/base/internal/tsan_mutex_interface.h"
+#include "absl/base/port.h"
+#include "absl/debugging/stacktrace.h"
+#include "absl/synchronization/internal/graphcycles.h"
+#include "absl/synchronization/internal/per_thread_sem.h"
+#include "absl/time/time.h"
+
+using absl::base_internal::CurrentThreadIdentityIfPresent;
+using absl::base_internal::PerThreadSynch;
+using absl::base_internal::ThreadIdentity;
+using absl::synchronization_internal::GetOrCreateCurrentThreadIdentity;
+using absl::synchronization_internal::GraphCycles;
+using absl::synchronization_internal::GraphId;
+using absl::synchronization_internal::InvalidGraphId;
+using absl::synchronization_internal::KernelTimeout;
+using absl::synchronization_internal::PerThreadSem;
+
+extern "C" {
+ABSL_ATTRIBUTE_WEAK void AbslInternalMutexYield() { std::this_thread::yield(); }
+}  // extern "C"
+
+namespace absl {
+
+namespace {
+
+#if defined(THREAD_SANITIZER)
+constexpr OnDeadlockCycle kDeadlockDetectionDefault = OnDeadlockCycle::kIgnore;
+#else
+constexpr OnDeadlockCycle kDeadlockDetectionDefault = OnDeadlockCycle::kAbort;
+#endif
+
+ABSL_CONST_INIT std::atomic<OnDeadlockCycle> synch_deadlock_detection(
+    kDeadlockDetectionDefault);
+ABSL_CONST_INIT std::atomic<bool> synch_check_invariants(false);
+
+// ------------------------------------------ spinlock support
+
+// Make sure read-only globals used in the Mutex code are contained on the
+// same cacheline and cacheline aligned to eliminate any false sharing with
+// other globals from this and other modules.
+static struct MutexGlobals {
+  MutexGlobals() {
+    // Find machine-specific data needed for Delay() and
+    // TryAcquireWithSpinning(). This runs in the global constructor
+    // sequence, and before that zeros are safe values.
+    num_cpus = absl::base_internal::NumCPUs();
+    spinloop_iterations = num_cpus > 1 ? 1500 : 0;
+  }
+  int num_cpus;
+  int spinloop_iterations;
+  // Pad this struct to a full cacheline to prevent false sharing.
+  char padding[ABSL_CACHELINE_SIZE - 2 * sizeof(int)];
+} ABSL_CACHELINE_ALIGNED mutex_globals;
+static_assert(
+    sizeof(MutexGlobals) == ABSL_CACHELINE_SIZE,
+    "MutexGlobals must occupy an entire cacheline to prevent false sharing");
+
+ABSL_CONST_INIT absl::base_internal::AtomicHook<void (*)(int64_t wait_cycles)>
+    submit_profile_data;
+ABSL_CONST_INIT absl::base_internal::AtomicHook<
+    void (*)(const char *msg, const void *obj, int64_t wait_cycles)> mutex_tracer;
+ABSL_CONST_INIT absl::base_internal::AtomicHook<
+    void (*)(const char *msg, const void *cv)> cond_var_tracer;
+ABSL_CONST_INIT absl::base_internal::AtomicHook<
+    bool (*)(const void *pc, char *out, int out_size)> symbolizer;
+
+}  // namespace
+
+void RegisterMutexProfiler(void (*fn)(int64_t wait_timestamp)) {
+  submit_profile_data.Store(fn);
+}
+
+void RegisterMutexTracer(void (*fn)(const char *msg, const void *obj,
+                                    int64_t wait_cycles)) {
+  mutex_tracer.Store(fn);
+}
+
+void RegisterCondVarTracer(void (*fn)(const char *msg, const void *cv)) {
+  cond_var_tracer.Store(fn);
+}
+
+void RegisterSymbolizer(bool (*fn)(const void *pc, char *out, int out_size)) {
+  symbolizer.Store(fn);
+}
+
+// spinlock delay on iteration c.  Returns new c.
+namespace {
+  enum DelayMode { AGGRESSIVE, GENTLE };
+};
+static int Delay(int32_t c, DelayMode mode) {
+  // If this a uniprocessor, only yield/sleep.  Otherwise, if the mode is
+  // aggressive then spin many times before yielding.  If the mode is
+  // gentle then spin only a few times before yielding.  Aggressive spinning is
+  // used to ensure that an Unlock() call, which  must get the spin lock for
+  // any thread to make progress gets it without undue delay.
+  int32_t limit = (mutex_globals.num_cpus > 1) ?
+      ((mode == AGGRESSIVE) ? 5000 : 250) : 0;
+  if (c < limit) {
+    c++;               // spin
+  } else {
+    ABSL_TSAN_MUTEX_PRE_DIVERT(0, 0);
+    if (c == limit) {  // yield once
+      AbslInternalMutexYield();
+      c++;
+    } else {           // then wait
+      absl::SleepFor(absl::Microseconds(10));
+      c = 0;
+    }
+    ABSL_TSAN_MUTEX_POST_DIVERT(0, 0);
+  }
+  return (c);
+}
+
+// --------------------------Generic atomic ops
+// Ensure that "(*pv & bits) == bits" by doing an atomic update of "*pv" to
+// "*pv | bits" if necessary.  Wait until (*pv & wait_until_clear)==0
+// before making any change.
+// This is used to set flags in mutex and condition variable words.
+static void AtomicSetBits(std::atomic<intptr_t>* pv, intptr_t bits,
+                          intptr_t wait_until_clear) {
+  intptr_t v;
+  do {
+    v = pv->load(std::memory_order_relaxed);
+  } while ((v & bits) != bits &&
+           ((v & wait_until_clear) != 0 ||
+            !pv->compare_exchange_weak(v, v | bits,
+                                       std::memory_order_release,
+                                       std::memory_order_relaxed)));
+}
+
+// Ensure that "(*pv & bits) == 0" by doing an atomic update of "*pv" to
+// "*pv & ~bits" if necessary.  Wait until (*pv & wait_until_clear)==0
+// before making any change.
+// This is used to unset flags in mutex and condition variable words.
+static void AtomicClearBits(std::atomic<intptr_t>* pv, intptr_t bits,
+                            intptr_t wait_until_clear) {
+  intptr_t v;
+  do {
+    v = pv->load(std::memory_order_relaxed);
+  } while ((v & bits) != 0 &&
+           ((v & wait_until_clear) != 0 ||
+            !pv->compare_exchange_weak(v, v & ~bits,
+                                       std::memory_order_release,
+                                       std::memory_order_relaxed)));
+}
+
+//------------------------------------------------------------------
+
+// Data for doing deadlock detection.
+static absl::base_internal::SpinLock deadlock_graph_mu(
+    absl::base_internal::kLinkerInitialized);
+
+// graph used to detect deadlocks.
+static GraphCycles *deadlock_graph GUARDED_BY(deadlock_graph_mu)
+    PT_GUARDED_BY(deadlock_graph_mu);
+
+//------------------------------------------------------------------
+// An event mechanism for debugging mutex use.
+// It also allows mutexes to be given names for those who can't handle
+// addresses, and instead like to give their data structures names like
+// "Henry", "Fido", or "Rupert IV, King of Yondavia".
+
+namespace {  // to prevent name pollution
+enum {       // Mutex and CondVar events passed as "ev" to PostSynchEvent
+             // Mutex events
+  SYNCH_EV_TRYLOCK_SUCCESS,
+  SYNCH_EV_TRYLOCK_FAILED,
+  SYNCH_EV_READERTRYLOCK_SUCCESS,
+  SYNCH_EV_READERTRYLOCK_FAILED,
+  SYNCH_EV_LOCK,
+  SYNCH_EV_LOCK_RETURNING,
+  SYNCH_EV_READERLOCK,
+  SYNCH_EV_READERLOCK_RETURNING,
+  SYNCH_EV_UNLOCK,
+  SYNCH_EV_READERUNLOCK,
+
+  // CondVar events
+  SYNCH_EV_WAIT,
+  SYNCH_EV_WAIT_RETURNING,
+  SYNCH_EV_SIGNAL,
+  SYNCH_EV_SIGNALALL,
+};
+
+enum {                 // Event flags
+  SYNCH_F_R = 0x01,    // reader event
+  SYNCH_F_LCK = 0x02,  // PostSynchEvent called with mutex held
+  SYNCH_F_ACQ = 0x04,  // event is an acquire
+
+  SYNCH_F_LCK_W = SYNCH_F_LCK,
+  SYNCH_F_LCK_R = SYNCH_F_LCK | SYNCH_F_R,
+  SYNCH_F_ACQ_W = SYNCH_F_ACQ,
+  SYNCH_F_ACQ_R = SYNCH_F_ACQ | SYNCH_F_R,
+};
+}  // anonymous namespace
+
+// Properties of the events.
+static const struct {
+  int flags;
+  const char *msg;
+} event_properties[] = {
+  { SYNCH_F_LCK_W|SYNCH_F_ACQ_W, "TryLock succeeded " },
+  { 0,                           "TryLock failed " },
+  { SYNCH_F_LCK_R|SYNCH_F_ACQ_R, "ReaderTryLock succeeded " },
+  { 0,                           "ReaderTryLock failed " },
+  {               SYNCH_F_ACQ_W, "Lock blocking " },
+  { SYNCH_F_LCK_W,               "Lock returning " },
+  {               SYNCH_F_ACQ_R, "ReaderLock blocking " },
+  { SYNCH_F_LCK_R,               "ReaderLock returning " },
+  { SYNCH_F_LCK_W,               "Unlock " },
+  { SYNCH_F_LCK_R,               "ReaderUnlock " },
+  { 0,                           "Wait on " },
+  { 0,                           "Wait unblocked " },
+  { 0,                           "Signal on " },
+  { 0,                           "SignalAll on " },
+};
+
+static absl::base_internal::SpinLock synch_event_mu(
+    absl::base_internal::kLinkerInitialized);
+// protects synch_event
+
+// Hash table size; should be prime > 2.
+// Can't be too small, as it's used for deadlock detection information.
+static const uint32_t kNSynchEvent = 1031;
+
+// We need to hide Mutexes (or other deadlock detection's pointers)
+// from the leak detector.
+static const uintptr_t kHideMask = static_cast<uintptr_t>(0xF03A5F7BF03A5F7BLL);
+static uintptr_t MaskMu(const void *mu) {
+  return reinterpret_cast<uintptr_t>(mu) ^ kHideMask;
+}
+
+static struct SynchEvent {     // this is a trivial hash table for the events
+  // struct is freed when refcount reaches 0
+  int refcount GUARDED_BY(synch_event_mu);
+
+  // buckets have linear, 0-terminated  chains
+  SynchEvent *next GUARDED_BY(synch_event_mu);
+
+  // Constant after initialization
+  uintptr_t masked_addr;  // object at this address is called "name"
+
+  // No explicit synchronization used.  Instead we assume that the
+  // client who enables/disables invariants/logging on a Mutex does so
+  // while the Mutex is not being concurrently accessed by others.
+  void (*invariant)(void *arg);  // called on each event
+  void *arg;            // first arg to (*invariant)()
+  bool log;             // logging turned on
+
+  // Constant after initialization
+  char name[1];         // actually longer---null-terminated std::string
+} *synch_event[kNSynchEvent] GUARDED_BY(synch_event_mu);
+
+// Ensure that the object at "addr" has a SynchEvent struct associated with it,
+// set "bits" in the word there (waiting until lockbit is clear before doing
+// so), and return a refcounted reference that will remain valid until
+// UnrefSynchEvent() is called.  If a new SynchEvent is allocated,
+// the std::string name is copied into it.
+// When used with a mutex, the caller should also ensure that kMuEvent
+// is set in the mutex word, and similarly for condition variables and kCVEvent.
+static SynchEvent *EnsureSynchEvent(std::atomic<intptr_t> *addr,
+                                    const char *name, intptr_t bits,
+                                    intptr_t lockbit) {
+  uint32_t h = reinterpret_cast<intptr_t>(addr) % kNSynchEvent;
+  SynchEvent *e;
+  // first look for existing SynchEvent struct..
+  synch_event_mu.Lock();
+  for (e = synch_event[h]; e != nullptr && e->masked_addr != MaskMu(addr);
+       e = e->next) {
+  }
+  if (e == nullptr) {  // no SynchEvent struct found; make one.
+    if (name == nullptr) {
+      name = "";
+    }
+    size_t l = strlen(name);
+    e = reinterpret_cast<SynchEvent *>(
+        base_internal::LowLevelAlloc::Alloc(sizeof(*e) + l));
+    e->refcount = 2;    // one for return value, one for linked list
+    e->masked_addr = MaskMu(addr);
+    e->invariant = nullptr;
+    e->arg = nullptr;
+    e->log = false;
+    strcpy(e->name, name);  // NOLINT(runtime/printf)
+    e->next = synch_event[h];
+    AtomicSetBits(addr, bits, lockbit);
+    synch_event[h] = e;
+  } else {
+    e->refcount++;      // for return value
+  }
+  synch_event_mu.Unlock();
+  return e;
+}
+
+// Deallocate the SynchEvent *e, whose refcount has fallen to zero.
+static void DeleteSynchEvent(SynchEvent *e) {
+  base_internal::LowLevelAlloc::Free(e);
+}
+
+// Decrement the reference count of *e, or do nothing if e==null.
+static void UnrefSynchEvent(SynchEvent *e) {
+  if (e != nullptr) {
+    synch_event_mu.Lock();
+    bool del = (--(e->refcount) == 0);
+    synch_event_mu.Unlock();
+    if (del) {
+      DeleteSynchEvent(e);
+    }
+  }
+}
+
+// Forget the mapping from the object (Mutex or CondVar) at address addr
+// to SynchEvent object, and clear "bits" in its word (waiting until lockbit
+// is clear before doing so).
+static void ForgetSynchEvent(std::atomic<intptr_t> *addr, intptr_t bits,
+                             intptr_t lockbit) {
+  uint32_t h = reinterpret_cast<intptr_t>(addr) % kNSynchEvent;
+  SynchEvent **pe;
+  SynchEvent *e;
+  synch_event_mu.Lock();
+  for (pe = &synch_event[h];
+       (e = *pe) != nullptr && e->masked_addr != MaskMu(addr); pe = &e->next) {
+  }
+  bool del = false;
+  if (e != nullptr) {
+    *pe = e->next;
+    del = (--(e->refcount) == 0);
+  }
+  AtomicClearBits(addr, bits, lockbit);
+  synch_event_mu.Unlock();
+  if (del) {
+    DeleteSynchEvent(e);
+  }
+}
+
+// Return a refcounted reference to the SynchEvent of the object at address
+// "addr", if any.  The pointer returned is valid until the UnrefSynchEvent() is
+// called.
+static SynchEvent *GetSynchEvent(const void *addr) {
+  uint32_t h = reinterpret_cast<intptr_t>(addr) % kNSynchEvent;
+  SynchEvent *e;
+  synch_event_mu.Lock();
+  for (e = synch_event[h]; e != nullptr && e->masked_addr != MaskMu(addr);
+       e = e->next) {
+  }
+  if (e != nullptr) {
+    e->refcount++;
+  }
+  synch_event_mu.Unlock();
+  return e;
+}
+
+// Called when an event "ev" occurs on a Mutex of CondVar "obj"
+// if event recording is on
+static void PostSynchEvent(void *obj, int ev) {
+  SynchEvent *e = GetSynchEvent(obj);
+  // logging is on if event recording is on and either there's no event struct,
+  // or it explicitly says to log
+  if (e == nullptr || e->log) {
+    void *pcs[40];
+    int n = absl::GetStackTrace(pcs, ABSL_ARRAYSIZE(pcs), 1);
+    // A buffer with enough space for the ASCII for all the PCs, even on a
+    // 64-bit machine.
+    char buffer[ABSL_ARRAYSIZE(pcs) * 24];
+    int pos = snprintf(buffer, sizeof (buffer), " @");
+    for (int i = 0; i != n; i++) {
+      pos += snprintf(&buffer[pos], sizeof (buffer) - pos, " %p", pcs[i]);
+    }
+    ABSL_RAW_LOG(INFO, "%s%p %s %s", event_properties[ev].msg, obj,
+                 (e == nullptr ? "" : e->name), buffer);
+  }
+  if ((event_properties[ev].flags & SYNCH_F_LCK) != 0 && e != nullptr &&
+      e->invariant != nullptr) {
+    (*e->invariant)(e->arg);
+  }
+  UnrefSynchEvent(e);
+}
+
+//------------------------------------------------------------------
+
+// The SynchWaitParams struct encapsulates the way in which a thread is waiting:
+// whether it has a timeout, the condition, exclusive/shared, and whether a
+// condition variable wait has an associated Mutex (as opposed to another
+// type of lock).  It also points to the PerThreadSynch struct of its thread.
+// cv_word tells Enqueue() to enqueue on a CondVar using CondVarEnqueue().
+//
+// This structure is held on the stack rather than directly in
+// PerThreadSynch because a thread can be waiting on multiple Mutexes if,
+// while waiting on one Mutex, the implementation calls a client callback
+// (such as a Condition function) that acquires another Mutex. We don't
+// strictly need to allow this, but programmers become confused if we do not
+// allow them to use functions such a LOG() within Condition functions.  The
+// PerThreadSynch struct points at the most recent SynchWaitParams struct when
+// the thread is on a Mutex's waiter queue.
+struct SynchWaitParams {
+  SynchWaitParams(Mutex::MuHow how_arg, const Condition *cond_arg,
+                  KernelTimeout timeout_arg, Mutex *cvmu_arg,
+                  PerThreadSynch *thread_arg,
+                  std::atomic<intptr_t> *cv_word_arg)
+      : how(how_arg),
+        cond(cond_arg),
+        timeout(timeout_arg),
+        cvmu(cvmu_arg),
+        thread(thread_arg),
+        cv_word(cv_word_arg),
+        contention_start_cycles(base_internal::CycleClock::Now()) {}
+
+  const Mutex::MuHow how;  // How this thread needs to wait.
+  const Condition *cond;  // The condition that this thread is waiting for.
+                          // In Mutex, this field is set to zero if a timeout
+                          // expires.
+  KernelTimeout timeout;  // timeout expiry---absolute time
+                          // In Mutex, this field is set to zero if a timeout
+                          // expires.
+  Mutex *const cvmu;      // used for transfer from cond var to mutex
+  PerThreadSynch *const thread;  // thread that is waiting
+
+  // If not null, thread should be enqueued on the CondVar whose state
+  // word is cv_word instead of queueing normally on the Mutex.
+  std::atomic<intptr_t> *cv_word;
+
+  int64_t contention_start_cycles;  // Time (in cycles) when this thread started
+                                  // to contend for the mutex.
+};
+
+struct SynchLocksHeld {
+  int n;              // number of valid entries in locks[]
+  bool overflow;      // true iff we overflowed the array at some point
+  struct {
+    Mutex *mu;        // lock acquired
+    int32_t count;      // times acquired
+    GraphId id;       // deadlock_graph id of acquired lock
+  } locks[40];
+  // If a thread overfills the array during deadlock detection, we
+  // continue, discarding information as needed.  If no overflow has
+  // taken place, we can provide more error checking, such as
+  // detecting when a thread releases a lock it does not hold.
+};
+
+// A sentinel value in lists that is not 0.
+// A 0 value is used to mean "not on a list".
+static PerThreadSynch *const kPerThreadSynchNull =
+  reinterpret_cast<PerThreadSynch *>(1);
+
+static SynchLocksHeld *LocksHeldAlloc() {
+  SynchLocksHeld *ret = reinterpret_cast<SynchLocksHeld *>(
+      base_internal::LowLevelAlloc::Alloc(sizeof(SynchLocksHeld)));
+  ret->n = 0;
+  ret->overflow = false;
+  return ret;
+}
+
+// Return the PerThreadSynch-struct for this thread.
+static PerThreadSynch *Synch_GetPerThread() {
+  ThreadIdentity *identity = GetOrCreateCurrentThreadIdentity();
+  return &identity->per_thread_synch;
+}
+
+static PerThreadSynch *Synch_GetPerThreadAnnotated(Mutex *mu) {
+  if (mu) {
+    ABSL_TSAN_MUTEX_PRE_DIVERT(mu, 0);
+  }
+  PerThreadSynch *w = Synch_GetPerThread();
+  if (mu) {
+    ABSL_TSAN_MUTEX_POST_DIVERT(mu, 0);
+  }
+  return w;
+}
+
+static SynchLocksHeld *Synch_GetAllLocks() {
+  PerThreadSynch *s = Synch_GetPerThread();
+  if (s->all_locks == nullptr) {
+    s->all_locks = LocksHeldAlloc();  // Freed by ReclaimThreadIdentity.
+  }
+  return s->all_locks;
+}
+
+// Post on "w"'s associated PerThreadSem.
+inline void Mutex::IncrementSynchSem(Mutex *mu, PerThreadSynch *w) {
+  if (mu) {
+    ABSL_TSAN_MUTEX_PRE_DIVERT(mu, 0);
+  }
+  PerThreadSem::Post(w->thread_identity());
+  if (mu) {
+    ABSL_TSAN_MUTEX_POST_DIVERT(mu, 0);
+  }
+}
+
+// Wait on "w"'s associated PerThreadSem; returns false if timeout expired.
+bool Mutex::DecrementSynchSem(Mutex *mu, PerThreadSynch *w, KernelTimeout t) {
+  if (mu) {
+    ABSL_TSAN_MUTEX_PRE_DIVERT(mu, 0);
+  }
+  assert(w == Synch_GetPerThread());
+  static_cast<void>(w);
+  bool res = PerThreadSem::Wait(t);
+  if (mu) {
+    ABSL_TSAN_MUTEX_POST_DIVERT(mu, 0);
+  }
+  return res;
+}
+
+// We're in a fatal signal handler that hopes to use Mutex and to get
+// lucky by not deadlocking.  We try to improve its chances of success
+// by effectively disabling some of the consistency checks.  This will
+// prevent certain ABSL_RAW_CHECK() statements from being triggered when
+// re-rentry is detected.  The ABSL_RAW_CHECK() statements are those in the
+// Mutex code checking that the "waitp" field has not been reused.
+void Mutex::InternalAttemptToUseMutexInFatalSignalHandler() {
+  // Fix the per-thread state only if it exists.
+  ThreadIdentity *identity = CurrentThreadIdentityIfPresent();
+  if (identity != nullptr) {
+    identity->per_thread_synch.suppress_fatal_errors = true;
+  }
+  // Don't do deadlock detection when we are already failing.
+  synch_deadlock_detection.store(OnDeadlockCycle::kIgnore,
+                                 std::memory_order_release);
+}
+
+// --------------------------time support
+
+// Return the current time plus the timeout.  Use the same clock as
+// PerThreadSem::Wait() for consistency.  Unfortunately, we don't have
+// such a choice when a deadline is given directly.
+static absl::Time DeadlineFromTimeout(absl::Duration timeout) {
+#ifndef _WIN32
+  struct timeval tv;
+  gettimeofday(&tv, nullptr);
+  return absl::TimeFromTimeval(tv) + timeout;
+#else
+  return absl::Now() + timeout;
+#endif
+}
+
+// --------------------------Mutexes
+
+// In the layout below, the msb of the bottom byte is currently unused.  Also,
+// the following constraints were considered in choosing the layout:
+//  o Both the debug allocator's "uninitialized" and "freed" patterns (0xab and
+//    0xcd) are illegal: reader and writer lock both held.
+//  o kMuWriter and kMuEvent should exceed kMuDesig and kMuWait, to enable the
+//    bit-twiddling trick in Mutex::Unlock().
+//  o kMuWriter / kMuReader == kMuWrWait / kMuWait,
+//    to enable the bit-twiddling trick in CheckForMutexCorruption().
+static const intptr_t kMuReader      = 0x0001L;  // a reader holds the lock
+static const intptr_t kMuDesig       = 0x0002L;  // there's a designated waker
+static const intptr_t kMuWait        = 0x0004L;  // threads are waiting
+static const intptr_t kMuWriter      = 0x0008L;  // a writer holds the lock
+static const intptr_t kMuEvent       = 0x0010L;  // record this mutex's events
+// INVARIANT1:  there's a thread that was blocked on the mutex, is
+// no longer, yet has not yet acquired the mutex.  If there's a
+// designated waker, all threads can avoid taking the slow path in
+// unlock because the designated waker will subsequently acquire
+// the lock and wake someone.  To maintain INVARIANT1 the bit is
+// set when a thread is unblocked(INV1a), and threads that were
+// unblocked reset the bit when they either acquire or re-block
+// (INV1b).
+static const intptr_t kMuWrWait      = 0x0020L;  // runnable writer is waiting
+                                                 // for a reader
+static const intptr_t kMuSpin        = 0x0040L;  // spinlock protects wait list
+static const intptr_t kMuLow         = 0x00ffL;  // mask all mutex bits
+static const intptr_t kMuHigh        = ~kMuLow;  // mask pointer/reader count
+
+// Hack to make constant values available to gdb pretty printer
+enum {
+  kGdbMuSpin = kMuSpin,
+  kGdbMuEvent = kMuEvent,
+  kGdbMuWait = kMuWait,
+  kGdbMuWriter = kMuWriter,
+  kGdbMuDesig = kMuDesig,
+  kGdbMuWrWait = kMuWrWait,
+  kGdbMuReader = kMuReader,
+  kGdbMuLow = kMuLow,
+};
+
+// kMuWrWait implies kMuWait.
+// kMuReader and kMuWriter are mutually exclusive.
+// If kMuReader is zero, there are no readers.
+// Otherwise, if kMuWait is zero, the high order bits contain a count of the
+// number of readers.  Otherwise, the reader count is held in
+// PerThreadSynch::readers of the most recently queued waiter, again in the
+// bits above kMuLow.
+static const intptr_t kMuOne = 0x0100;  // a count of one reader
+
+// flags passed to Enqueue and LockSlow{,WithTimeout,Loop}
+static const int kMuHasBlocked = 0x01;  // already blocked (MUST == 1)
+static const int kMuIsCond = 0x02;      // conditional waiter (CV or Condition)
+
+static_assert(PerThreadSynch::kAlignment > kMuLow,
+              "PerThreadSynch::kAlignment must be greater than kMuLow");
+
+// This struct contains various bitmasks to be used in
+// acquiring and releasing a mutex in a particular mode.
+struct MuHowS {
+  // if all the bits in fast_need_zero are zero, the lock can be acquired by
+  // adding fast_add and oring fast_or.  The bit kMuDesig should be reset iff
+  // this is the designated waker.
+  intptr_t fast_need_zero;
+  intptr_t fast_or;
+  intptr_t fast_add;
+
+  intptr_t slow_need_zero;  // fast_need_zero with events (e.g. logging)
+
+  intptr_t slow_inc_need_zero;  // if all the bits in slow_inc_need_zero are
+                                // zero a reader can acquire a read share by
+                                // setting the reader bit and incrementing
+                                // the reader count (in last waiter since
+                                // we're now slow-path).  kMuWrWait be may
+                                // be ignored if we already waited once.
+};
+
+static const MuHowS kSharedS = {
+    // shared or read lock
+    kMuWriter | kMuWait | kMuEvent,   // fast_need_zero
+    kMuReader,                        // fast_or
+    kMuOne,                           // fast_add
+    kMuWriter | kMuWait,              // slow_need_zero
+    kMuSpin | kMuWriter | kMuWrWait,  // slow_inc_need_zero
+};
+static const MuHowS kExclusiveS = {
+    // exclusive or write lock
+    kMuWriter | kMuReader | kMuEvent,  // fast_need_zero
+    kMuWriter,                         // fast_or
+    0,                                 // fast_add
+    kMuWriter | kMuReader,             // slow_need_zero
+    ~static_cast<intptr_t>(0),         // slow_inc_need_zero
+};
+static const Mutex::MuHow kShared = &kSharedS;        // shared lock
+static const Mutex::MuHow kExclusive = &kExclusiveS;  // exclusive lock
+
+#ifdef NDEBUG
+static constexpr bool kDebugMode = false;
+#else
+static constexpr bool kDebugMode = true;
+#endif
+
+#ifdef THREAD_SANITIZER
+static unsigned TsanFlags(Mutex::MuHow how) {
+  return how == kShared ? __tsan_mutex_read_lock : 0;
+}
+#endif
+
+Mutex::Mutex() : mu_(0) {
+  ABSL_TSAN_MUTEX_CREATE(this, 0);
+}
+
+static bool DebugOnlyIsExiting() {
+  return false;
+}
+
+Mutex::~Mutex() {
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+  if ((v & kMuEvent) != 0 && !DebugOnlyIsExiting()) {
+    ForgetSynchEvent(&this->mu_, kMuEvent, kMuSpin);
+  }
+  if (kDebugMode) {
+    this->ForgetDeadlockInfo();
+  }
+  ABSL_TSAN_MUTEX_DESTROY(this, 0);
+}
+
+void Mutex::EnableDebugLog(const char *name) {
+  SynchEvent *e = EnsureSynchEvent(&this->mu_, name, kMuEvent, kMuSpin);
+  e->log = true;
+  UnrefSynchEvent(e);
+}
+
+void EnableMutexInvariantDebugging(bool enabled) {
+  synch_check_invariants.store(enabled, std::memory_order_release);
+}
+
+void Mutex::EnableInvariantDebugging(void (*invariant)(void *),
+                                     void *arg) {
+  if (synch_check_invariants.load(std::memory_order_acquire) &&
+      invariant != nullptr) {
+    SynchEvent *e = EnsureSynchEvent(&this->mu_, nullptr, kMuEvent, kMuSpin);
+    e->invariant = invariant;
+    e->arg = arg;
+    UnrefSynchEvent(e);
+  }
+}
+
+void SetMutexDeadlockDetectionMode(OnDeadlockCycle mode) {
+  synch_deadlock_detection.store(mode, std::memory_order_release);
+}
+
+// Return true iff threads x and y are waiting on the same condition for the
+// same type of lock.  Requires that x and y be waiting on the same Mutex
+// queue.
+static bool MuSameCondition(PerThreadSynch *x, PerThreadSynch *y) {
+  return x->waitp->how == y->waitp->how &&
+         Condition::GuaranteedEqual(x->waitp->cond, y->waitp->cond);
+}
+
+// Given the contents of a mutex word containing a PerThreadSynch pointer,
+// return the pointer.
+static inline PerThreadSynch *GetPerThreadSynch(intptr_t v) {
+  return reinterpret_cast<PerThreadSynch *>(v & kMuHigh);
+}
+
+// The next several routines maintain the per-thread next and skip fields
+// used in the Mutex waiter queue.
+// The queue is a circular singly-linked list, of which the "head" is the
+// last element, and head->next if the first element.
+// The skip field has the invariant:
+//   For thread x, x->skip is one of:
+//     - invalid (iff x is not in a Mutex wait queue),
+//     - null, or
+//     - a pointer to a distinct thread waiting later in the same Mutex queue
+//       such that all threads in [x, x->skip] have the same condition and
+//       lock type (MuSameCondition() is true for all pairs in [x, x->skip]).
+// In addition, if x->skip is  valid, (x->may_skip || x->skip == null)
+//
+// By the spec of MuSameCondition(), it is not necessary when removing the
+// first runnable thread y from the front a Mutex queue to adjust the skip
+// field of another thread x because if x->skip==y, x->skip must (have) become
+// invalid before y is removed.  The function TryRemove can remove a specified
+// thread from an arbitrary position in the queue whether runnable or not, so
+// it fixes up skip fields that would otherwise be left dangling.
+// The statement
+//     if (x->may_skip && MuSameCondition(x, x->next)) { x->skip = x->next; }
+// maintains the invariant provided x is not the last waiter in a Mutex queue
+// The statement
+//          if (x->skip != null) { x->skip = x->skip->skip; }
+// maintains the invariant.
+
+// Returns the last thread y in a mutex waiter queue such that all threads in
+// [x, y] inclusive share the same condition.  Sets skip fields of some threads
+// in that range to optimize future evaluation of Skip() on x values in
+// the range.  Requires thread x is in a mutex waiter queue.
+// The locking is unusual.  Skip() is called under these conditions:
+//   - spinlock is held in call from Enqueue(), with maybe_unlocking == false
+//   - Mutex is held in call from UnlockSlow() by last unlocker, with
+//     maybe_unlocking == true
+//   - both Mutex and spinlock are held in call from DequeueAllWakeable() (from
+//     UnlockSlow()) and TryRemove()
+// These cases are mutually exclusive, so Skip() never runs concurrently
+// with itself on the same Mutex.   The skip chain is used in these other places
+// that cannot occur concurrently:
+//   - FixSkip() (from TryRemove()) - spinlock and Mutex are held)
+//   - Dequeue() (with spinlock and Mutex held)
+//   - UnlockSlow() (with spinlock and Mutex held)
+// A more complex case is Enqueue()
+//   - Enqueue() (with spinlock held and maybe_unlocking == false)
+//               This is the first case in which Skip is called, above.
+//   - Enqueue() (without spinlock held; but queue is empty and being freshly
+//                formed)
+//   - Enqueue() (with spinlock held and maybe_unlocking == true)
+// The first case has mutual exclusion, and the second isolation through
+// working on an otherwise unreachable data structure.
+// In the last case, Enqueue() is required to change no skip/next pointers
+// except those in the added node and the former "head" node.  This implies
+// that the new node is added after head, and so must be the new head or the
+// new front of the queue.
+static PerThreadSynch *Skip(PerThreadSynch *x) {
+  PerThreadSynch *x0 = nullptr;
+  PerThreadSynch *x1 = x;
+  PerThreadSynch *x2 = x->skip;
+  if (x2 != nullptr) {
+    // Each iteration attempts to advance sequence (x0,x1,x2) to next sequence
+    // such that   x1 == x0->skip && x2 == x1->skip
+    while ((x0 = x1, x1 = x2, x2 = x2->skip) != nullptr) {
+      x0->skip = x2;      // short-circuit skip from x0 to x2
+    }
+    x->skip = x1;         // short-circuit skip from x to result
+  }
+  return x1;
+}
+
+// "ancestor" appears before "to_be_removed" in the same Mutex waiter queue.
+// The latter is going to be removed out of order, because of a timeout.
+// Check whether "ancestor" has a skip field pointing to "to_be_removed",
+// and fix it if it does.
+static void FixSkip(PerThreadSynch *ancestor, PerThreadSynch *to_be_removed) {
+  if (ancestor->skip == to_be_removed) {  // ancestor->skip left dangling
+    if (to_be_removed->skip != nullptr) {
+      ancestor->skip = to_be_removed->skip;  // can skip past to_be_removed
+    } else if (ancestor->next != to_be_removed) {  // they are not adjacent
+      ancestor->skip = ancestor->next;             // can skip one past ancestor
+    } else {
+      ancestor->skip = nullptr;  // can't skip at all
+    }
+  }
+}
+
+static void CondVarEnqueue(SynchWaitParams *waitp);
+
+// Enqueue thread "waitp->thread" on a waiter queue.
+// Called with mutex spinlock held if head != nullptr
+// If head==nullptr and waitp->cv_word==nullptr, then Enqueue() is
+// idempotent; it alters no state associated with the existing (empty)
+// queue.
+//
+// If waitp->cv_word == nullptr, queue the thread at either the front or
+// the end (according to its priority) of the circular mutex waiter queue whose
+// head is "head", and return the new head.  mu is the previous mutex state,
+// which contains the reader count (perhaps adjusted for the operation in
+// progress) if the list was empty and a read lock held, and the holder hint if
+// the list was empty and a write lock held.  (flags & kMuIsCond) indicates
+// whether this thread was transferred from a CondVar or is waiting for a
+// non-trivial condition.  In this case, Enqueue() never returns nullptr
+//
+// If waitp->cv_word != nullptr, CondVarEnqueue() is called, and "head" is
+// returned. This mechanism is used by CondVar to queue a thread on the
+// condition variable queue instead of the mutex queue in implementing Wait().
+// In this case, Enqueue() can return nullptr (if head==nullptr).
+static PerThreadSynch *Enqueue(PerThreadSynch *head,
+                               SynchWaitParams *waitp, intptr_t mu, int flags) {
+  // If we have been given a cv_word, call CondVarEnqueue() and return
+  // the previous head of the Mutex waiter queue.
+  if (waitp->cv_word != nullptr) {
+    CondVarEnqueue(waitp);
+    return head;
+  }
+
+  PerThreadSynch *s = waitp->thread;
+  ABSL_RAW_CHECK(
+      s->waitp == nullptr ||    // normal case
+          s->waitp == waitp ||  // Fer()---transfer from condition variable
+          s->suppress_fatal_errors,
+      "detected illegal recursion into Mutex code");
+  s->waitp = waitp;
+  s->skip = nullptr;             // maintain skip invariant (see above)
+  s->may_skip = true;            // always true on entering queue
+  s->wake = false;               // not being woken
+  s->cond_waiter = ((flags & kMuIsCond) != 0);
+  if (head == nullptr) {         // s is the only waiter
+    s->next = s;                 // it's the only entry in the cycle
+    s->readers = mu;             // reader count is from mu word
+    s->maybe_unlocking = false;  // no one is searching an empty list
+    head = s;                    // s is new head
+  } else {
+    PerThreadSynch *enqueue_after = nullptr;  // we'll put s after this element
+#ifdef ABSL_HAVE_PTHREAD_GETSCHEDPARAM
+    int64_t now_cycles = base_internal::CycleClock::Now();
+    if (s->next_priority_read_cycles < now_cycles) {
+      // Every so often, update our idea of the thread's priority.
+      // pthread_getschedparam() is 5% of the block/wakeup time;
+      // base_internal::CycleClock::Now() is 0.5%.
+      int policy;
+      struct sched_param param;
+      pthread_getschedparam(pthread_self(), &policy, &param);
+      s->priority = param.sched_priority;
+      s->next_priority_read_cycles =
+          now_cycles +
+          static_cast<int64_t>(base_internal::CycleClock::Frequency());
+    }
+    if (s->priority > head->priority) {  // s's priority is above head's
+      // try to put s in priority-fifo order, or failing that at the front.
+      if (!head->maybe_unlocking) {
+        // No unlocker can be scanning the queue, so we can insert between
+        // skip-chains, and within a skip-chain if it has the same condition as
+        // s.  We insert in priority-fifo order, examining the end of every
+        // skip-chain, plus every element with the same condition as s.
+        PerThreadSynch *advance_to = head;    // next value of enqueue_after
+        PerThreadSynch *cur;                  // successor of enqueue_after
+        do {
+          enqueue_after = advance_to;
+          cur = enqueue_after->next;  // this advance ensures progress
+          advance_to = Skip(cur);   // normally, advance to end of skip chain
+                                    // (side-effect: optimizes skip chain)
+          if (advance_to != cur && s->priority > advance_to->priority &&
+              MuSameCondition(s, cur)) {
+            // but this skip chain is not a singleton, s has higher priority
+            // than its tail and has the same condition as the chain,
+            // so we can insert within the skip-chain
+            advance_to = cur;         // advance by just one
+          }
+        } while (s->priority <= advance_to->priority);
+              // termination guaranteed because s->priority > head->priority
+              // and head is the end of a skip chain
+      } else if (waitp->how == kExclusive &&
+                 Condition::GuaranteedEqual(waitp->cond, nullptr)) {
+        // An unlocker could be scanning the queue, but we know it will recheck
+        // the queue front for writers that have no condition, which is what s
+        // is, so an insert at front is safe.
+        enqueue_after = head;       // add after head, at front
+      }
+    }
+#endif
+    if (enqueue_after != nullptr) {
+      s->next = enqueue_after->next;
+      enqueue_after->next = s;
+
+      // enqueue_after can be: head, Skip(...), or cur.
+      // The first two imply enqueue_after->skip == nullptr, and
+      // the last is used only if MuSameCondition(s, cur).
+      // We require this because clearing enqueue_after->skip
+      // is impossible; enqueue_after's predecessors might also
+      // incorrectly skip over s if we were to allow other
+      // insertion points.
+      ABSL_RAW_CHECK(
+          enqueue_after->skip == nullptr || MuSameCondition(enqueue_after, s),
+          "Mutex Enqueue failure");
+
+      if (enqueue_after != head && enqueue_after->may_skip &&
+          MuSameCondition(enqueue_after, enqueue_after->next)) {
+        // enqueue_after can skip to its new successor, s
+        enqueue_after->skip = enqueue_after->next;
+      }
+      if (MuSameCondition(s, s->next)) {  // s->may_skip is known to be true
+        s->skip = s->next;                // s may skip to its successor
+      }
+    } else {   // enqueue not done any other way, so
+               // we're inserting s at the back
+      // s will become new head; copy data from head into it
+      s->next = head->next;        // add s after head
+      head->next = s;
+      s->readers = head->readers;  // reader count is from previous head
+      s->maybe_unlocking = head->maybe_unlocking;  // same for unlock hint
+      if (head->may_skip && MuSameCondition(head, s)) {
+        // head now has successor; may skip
+        head->skip = s;
+      }
+      head = s;  // s is new head
+    }
+  }
+  s->state.store(PerThreadSynch::kQueued, std::memory_order_relaxed);
+  return head;
+}
+
+// Dequeue the successor pw->next of thread pw from the Mutex waiter queue
+// whose last element is head.  The new head element is returned, or null
+// if the list is made empty.
+// Dequeue is called with both spinlock and Mutex held.
+static PerThreadSynch *Dequeue(PerThreadSynch *head, PerThreadSynch *pw) {
+  PerThreadSynch *w = pw->next;
+  pw->next = w->next;         // snip w out of list
+  if (head == w) {            // we removed the head
+    head = (pw == w) ? nullptr : pw;  // either emptied list, or pw is new head
+  } else if (pw != head && MuSameCondition(pw, pw->next)) {
+    // pw can skip to its new successor
+    if (pw->next->skip !=
+        nullptr) {  // either skip to its successors skip target
+      pw->skip = pw->next->skip;
+    } else {                   // or to pw's successor
+      pw->skip = pw->next;
+    }
+  }
+  return head;
+}
+
+// Traverse the elements [ pw->next, h] of the circular list whose last element
+// is head.
+// Remove all elements with wake==true and place them in the
+// singly-linked list wake_list in the order found.   Assumes that
+// there is only one such element if the element has how == kExclusive.
+// Return the new head.
+static PerThreadSynch *DequeueAllWakeable(PerThreadSynch *head,
+                                          PerThreadSynch *pw,
+                                          PerThreadSynch **wake_tail) {
+  PerThreadSynch *orig_h = head;
+  PerThreadSynch *w = pw->next;
+  bool skipped = false;
+  do {
+    if (w->wake) {                    // remove this element
+      ABSL_RAW_CHECK(pw->skip == nullptr, "bad skip in DequeueAllWakeable");
+      // we're removing pw's successor so either pw->skip is zero or we should
+      // already have removed pw since if pw->skip!=null, pw has the same
+      // condition as w.
+      head = Dequeue(head, pw);
+      w->next = *wake_tail;           // keep list terminated
+      *wake_tail = w;                 // add w to wake_list;
+      wake_tail = &w->next;           // next addition to end
+      if (w->waitp->how == kExclusive) {  // wake at most 1 writer
+        break;
+      }
+    } else {                // not waking this one; skip
+      pw = Skip(w);       // skip as much as possible
+      skipped = true;
+    }
+    w = pw->next;
+    // We want to stop processing after we've considered the original head,
+    // orig_h.  We can't test for w==orig_h in the loop because w may skip over
+    // it; we are guaranteed only that w's predecessor will not skip over
+    // orig_h.  When we've considered orig_h, either we've processed it and
+    // removed it (so orig_h != head), or we considered it and skipped it (so
+    // skipped==true && pw == head because skipping from head always skips by
+    // just one, leaving pw pointing at head).  So we want to
+    // continue the loop with the negation of that expression.
+  } while (orig_h == head && (pw != head || !skipped));
+  return head;
+}
+
+// Try to remove thread s from the list of waiters on this mutex.
+// Does nothing if s is not on the waiter list.
+void Mutex::TryRemove(PerThreadSynch *s) {
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+  // acquire spinlock & lock
+  if ((v & (kMuWait | kMuSpin | kMuWriter | kMuReader)) == kMuWait &&
+      mu_.compare_exchange_strong(v, v | kMuSpin | kMuWriter,
+                                  std::memory_order_acquire,
+                                  std::memory_order_relaxed)) {
+    PerThreadSynch *h = GetPerThreadSynch(v);
+    if (h != nullptr) {
+      PerThreadSynch *pw = h;   // pw is w's predecessor
+      PerThreadSynch *w;
+      if ((w = pw->next) != s) {  // search for thread,
+        do {                      // processing at least one element
+          if (!MuSameCondition(s, w)) {  // seeking different condition
+            pw = Skip(w);                // so skip all that won't match
+            // we don't have to worry about dangling skip fields
+            // in the threads we skipped; none can point to s
+            // because their condition differs from s
+          } else {          // seeking same condition
+            FixSkip(w, s);  // fix up any skip pointer from w to s
+            pw = w;
+          }
+          // don't search further if we found the thread, or we're about to
+          // process the first thread again.
+        } while ((w = pw->next) != s && pw != h);
+      }
+      if (w == s) {                 // found thread; remove it
+        // pw->skip may be non-zero here; the loop above ensured that
+        // no ancestor of s can skip to s, so removal is safe anyway.
+        h = Dequeue(h, pw);
+        s->next = nullptr;
+        s->state.store(PerThreadSynch::kAvailable, std::memory_order_release);
+      }
+    }
+    intptr_t nv;
+    do {                        // release spinlock and lock
+      v = mu_.load(std::memory_order_relaxed);
+      nv = v & (kMuDesig | kMuEvent);
+      if (h != nullptr) {
+        nv |= kMuWait | reinterpret_cast<intptr_t>(h);
+        h->readers = 0;            // we hold writer lock
+        h->maybe_unlocking = false;  // finished unlocking
+      }
+    } while (!mu_.compare_exchange_weak(v, nv,
+                                        std::memory_order_release,
+                                        std::memory_order_relaxed));
+  }
+}
+
+// Wait until thread "s", which must be the current thread, is removed from the
+// this mutex's waiter queue.  If "s->waitp->timeout" has a timeout, wake up
+// if the wait extends past the absolute time specified, even if "s" is still
+// on the mutex queue.  In this case, remove "s" from the queue and return
+// true, otherwise return false.
+void Mutex::Block(PerThreadSynch *s) {
+  while (s->state.load(std::memory_order_acquire) == PerThreadSynch::kQueued) {
+    if (!DecrementSynchSem(this, s, s->waitp->timeout)) {
+      // After a timeout, we go into a spin loop until we remove ourselves
+      // from the queue, or someone else removes us.  We can't be sure to be
+      // able to remove ourselves in a single lock acquisition because this
+      // mutex may be held, and the holder has the right to read the centre
+      // of the waiter queue without holding the spinlock.
+      this->TryRemove(s);
+      int c = 0;
+      while (s->next != nullptr) {
+        c = Delay(c, GENTLE);
+        this->TryRemove(s);
+      }
+      if (kDebugMode) {
+        // This ensures that we test the case that TryRemove() is called when s
+        // is not on the queue.
+        this->TryRemove(s);
+      }
+      s->waitp->timeout = KernelTimeout::Never();      // timeout is satisfied
+      s->waitp->cond = nullptr;  // condition no longer relevant for wakeups
+    }
+  }
+  ABSL_RAW_CHECK(s->waitp != nullptr || s->suppress_fatal_errors,
+                 "detected illegal recursion in Mutex code");
+  s->waitp = nullptr;
+}
+
+// Wake thread w, and return the next thread in the list.
+PerThreadSynch *Mutex::Wakeup(PerThreadSynch *w) {
+  PerThreadSynch *next = w->next;
+  w->next = nullptr;
+  w->state.store(PerThreadSynch::kAvailable, std::memory_order_release);
+  IncrementSynchSem(this, w);
+
+  return next;
+}
+
+static GraphId GetGraphIdLocked(Mutex *mu)
+    EXCLUSIVE_LOCKS_REQUIRED(deadlock_graph_mu) {
+  if (!deadlock_graph) {  // (re)create the deadlock graph.
+    deadlock_graph =
+        new (base_internal::LowLevelAlloc::Alloc(sizeof(*deadlock_graph)))
+            GraphCycles;
+  }
+  return deadlock_graph->GetId(mu);
+}
+
+static GraphId GetGraphId(Mutex *mu) LOCKS_EXCLUDED(deadlock_graph_mu) {
+  deadlock_graph_mu.Lock();
+  GraphId id = GetGraphIdLocked(mu);
+  deadlock_graph_mu.Unlock();
+  return id;
+}
+
+// Record a lock acquisition.  This is used in debug mode for deadlock
+// detection.  The held_locks pointer points to the relevant data
+// structure for each case.
+static void LockEnter(Mutex* mu, GraphId id, SynchLocksHeld *held_locks) {
+  int n = held_locks->n;
+  int i = 0;
+  while (i != n && held_locks->locks[i].id != id) {
+    i++;
+  }
+  if (i == n) {
+    if (n == ABSL_ARRAYSIZE(held_locks->locks)) {
+      held_locks->overflow = true;  // lost some data
+    } else {                        // we have room for lock
+      held_locks->locks[i].mu = mu;
+      held_locks->locks[i].count = 1;
+      held_locks->locks[i].id = id;
+      held_locks->n = n + 1;
+    }
+  } else {
+    held_locks->locks[i].count++;
+  }
+}
+
+// Record a lock release.  Each call to LockEnter(mu, id, x) should be
+// eventually followed by a call to LockLeave(mu, id, x) by the same thread.
+// It does not process the event if is not needed when deadlock detection is
+// disabled.
+static void LockLeave(Mutex* mu, GraphId id, SynchLocksHeld *held_locks) {
+  int n = held_locks->n;
+  int i = 0;
+  while (i != n && held_locks->locks[i].id != id) {
+    i++;
+  }
+  if (i == n) {
+    if (!held_locks->overflow) {
+      // The deadlock id may have been reassigned after ForgetDeadlockInfo,
+      // but in that case mu should still be present.
+      i = 0;
+      while (i != n && held_locks->locks[i].mu != mu) {
+        i++;
+      }
+      if (i == n) {  // mu missing means releasing unheld lock
+        SynchEvent *mu_events = GetSynchEvent(mu);
+        ABSL_RAW_LOG(FATAL,
+                     "thread releasing lock it does not hold: %p %s; "
+                     ,
+                     static_cast<void *>(mu),
+                     mu_events == nullptr ? "" : mu_events->name);
+      }
+    }
+  } else if (held_locks->locks[i].count == 1) {
+    held_locks->n = n - 1;
+    held_locks->locks[i] = held_locks->locks[n - 1];
+    held_locks->locks[n - 1].id = InvalidGraphId();
+    held_locks->locks[n - 1].mu =
+        nullptr;  // clear mu to please the leak detector.
+  } else {
+    assert(held_locks->locks[i].count > 0);
+    held_locks->locks[i].count--;
+  }
+}
+
+// Call LockEnter() if in debug mode and deadlock detection is enabled.
+static inline void DebugOnlyLockEnter(Mutex *mu) {
+  if (kDebugMode) {
+    if (synch_deadlock_detection.load(std::memory_order_acquire) !=
+        OnDeadlockCycle::kIgnore) {
+      LockEnter(mu, GetGraphId(mu), Synch_GetAllLocks());
+    }
+  }
+}
+
+// Call LockEnter() if in debug mode and deadlock detection is enabled.
+static inline void DebugOnlyLockEnter(Mutex *mu, GraphId id) {
+  if (kDebugMode) {
+    if (synch_deadlock_detection.load(std::memory_order_acquire) !=
+        OnDeadlockCycle::kIgnore) {
+      LockEnter(mu, id, Synch_GetAllLocks());
+    }
+  }
+}
+
+// Call LockLeave() if in debug mode and deadlock detection is enabled.
+static inline void DebugOnlyLockLeave(Mutex *mu) {
+  if (kDebugMode) {
+    if (synch_deadlock_detection.load(std::memory_order_acquire) !=
+        OnDeadlockCycle::kIgnore) {
+      LockLeave(mu, GetGraphId(mu), Synch_GetAllLocks());
+    }
+  }
+}
+
+static char *StackString(void **pcs, int n, char *buf, int maxlen,
+                         bool symbolize) {
+  static const int kSymLen = 200;
+  char sym[kSymLen];
+  int len = 0;
+  for (int i = 0; i != n; i++) {
+    if (symbolize) {
+      if (!symbolizer(pcs[i], sym, kSymLen)) {
+        sym[0] = '\0';
+      }
+      snprintf(buf + len, maxlen - len, "%s\t@ %p %s\n",
+               (i == 0 ? "\n" : ""),
+               pcs[i], sym);
+    } else {
+      snprintf(buf + len, maxlen - len, " %p", pcs[i]);
+    }
+    len += strlen(&buf[len]);
+  }
+  return buf;
+}
+
+static char *CurrentStackString(char *buf, int maxlen, bool symbolize) {
+  void *pcs[40];
+  return StackString(pcs, absl::GetStackTrace(pcs, ABSL_ARRAYSIZE(pcs), 2), buf,
+                     maxlen, symbolize);
+}
+
+namespace {
+enum { kMaxDeadlockPathLen = 10 };  // maximum length of a deadlock cycle;
+                                    // a path this long would be remarkable
+// Buffers required to report a deadlock.
+// We do not allocate them on stack to avoid large stack frame.
+struct DeadlockReportBuffers {
+  char buf[6100];
+  GraphId path[kMaxDeadlockPathLen];
+};
+
+struct ScopedDeadlockReportBuffers {
+  ScopedDeadlockReportBuffers() {
+    b = reinterpret_cast<DeadlockReportBuffers *>(
+        base_internal::LowLevelAlloc::Alloc(sizeof(*b)));
+  }
+  ~ScopedDeadlockReportBuffers() { base_internal::LowLevelAlloc::Free(b); }
+  DeadlockReportBuffers *b;
+};
+
+// Helper to pass to GraphCycles::UpdateStackTrace.
+int GetStack(void** stack, int max_depth) {
+  return absl::GetStackTrace(stack, max_depth, 3);
+}
+}  // anonymous namespace
+
+// Called in debug mode when a thread is about to acquire a lock in a way that
+// may block.
+static GraphId DeadlockCheck(Mutex *mu) {
+  if (synch_deadlock_detection.load(std::memory_order_acquire) ==
+      OnDeadlockCycle::kIgnore) {
+    return InvalidGraphId();
+  }
+
+  SynchLocksHeld *all_locks = Synch_GetAllLocks();
+
+  absl::base_internal::SpinLockHolder lock(&deadlock_graph_mu);
+  const GraphId mu_id = GetGraphIdLocked(mu);
+
+  if (all_locks->n == 0) {
+    // There are no other locks held. Return now so that we don't need to
+    // call GetSynchEvent(). This way we do not record the stack trace
+    // for this Mutex. It's ok, since if this Mutex is involved in a deadlock,
+    // it can't always be the first lock acquired by a thread.
+    return mu_id;
+  }
+
+  // We prefer to keep stack traces that show a thread holding and acquiring
+  // as many locks as possible.  This increases the chances that a given edge
+  // in the acquires-before graph will be represented in the stack traces
+  // recorded for the locks.
+  deadlock_graph->UpdateStackTrace(mu_id, all_locks->n + 1, GetStack);
+
+  // For each other mutex already held by this thread:
+  for (int i = 0; i != all_locks->n; i++) {
+    const GraphId other_node_id = all_locks->locks[i].id;
+    const Mutex *other =
+        static_cast<const Mutex *>(deadlock_graph->Ptr(other_node_id));
+    if (other == nullptr) {
+      // Ignore stale lock
+      continue;
+    }
+
+    // Add the acquired-before edge to the graph.
+    if (!deadlock_graph->InsertEdge(other_node_id, mu_id)) {
+      ScopedDeadlockReportBuffers scoped_buffers;
+      DeadlockReportBuffers *b = scoped_buffers.b;
+      static int number_of_reported_deadlocks = 0;
+      number_of_reported_deadlocks++;
+      // Symbolize only 2 first deadlock report to avoid huge slowdowns.
+      bool symbolize = number_of_reported_deadlocks <= 2;
+      ABSL_RAW_LOG(ERROR, "Potential Mutex deadlock: %s",
+                   CurrentStackString(b->buf, sizeof (b->buf), symbolize));
+      int len = 0;
+      for (int j = 0; j != all_locks->n; j++) {
+        void* pr = deadlock_graph->Ptr(all_locks->locks[j].id);
+        if (pr != nullptr) {
+          snprintf(b->buf + len, sizeof (b->buf) - len, " %p", pr);
+          len += static_cast<int>(strlen(&b->buf[len]));
+        }
+      }
+      ABSL_RAW_LOG(ERROR, "Acquiring %p    Mutexes held: %s",
+                   static_cast<void *>(mu), b->buf);
+      ABSL_RAW_LOG(ERROR, "Cycle: ");
+      int path_len = deadlock_graph->FindPath(
+          mu_id, other_node_id, ABSL_ARRAYSIZE(b->path), b->path);
+      for (int j = 0; j != path_len; j++) {
+        GraphId id = b->path[j];
+        Mutex *path_mu = static_cast<Mutex *>(deadlock_graph->Ptr(id));
+        if (path_mu == nullptr) continue;
+        void** stack;
+        int depth = deadlock_graph->GetStackTrace(id, &stack);
+        snprintf(b->buf, sizeof(b->buf),
+                 "mutex@%p stack: ", static_cast<void *>(path_mu));
+        StackString(stack, depth, b->buf + strlen(b->buf),
+                    static_cast<int>(sizeof(b->buf) - strlen(b->buf)),
+                    symbolize);
+        ABSL_RAW_LOG(ERROR, "%s", b->buf);
+      }
+      if (synch_deadlock_detection.load(std::memory_order_acquire) ==
+          OnDeadlockCycle::kAbort) {
+        deadlock_graph_mu.Unlock();  // avoid deadlock in fatal sighandler
+        ABSL_RAW_LOG(FATAL, "dying due to potential deadlock");
+        return mu_id;
+      }
+      break;   // report at most one potential deadlock per acquisition
+    }
+  }
+
+  return mu_id;
+}
+
+// Invoke DeadlockCheck() iff we're in debug mode and
+// deadlock checking has been enabled.
+static inline GraphId DebugOnlyDeadlockCheck(Mutex *mu) {
+  if (kDebugMode && synch_deadlock_detection.load(std::memory_order_acquire) !=
+                        OnDeadlockCycle::kIgnore) {
+    return DeadlockCheck(mu);
+  } else {
+    return InvalidGraphId();
+  }
+}
+
+void Mutex::ForgetDeadlockInfo() {
+  if (kDebugMode && synch_deadlock_detection.load(std::memory_order_acquire) !=
+                        OnDeadlockCycle::kIgnore) {
+    deadlock_graph_mu.Lock();
+    if (deadlock_graph != nullptr) {
+      deadlock_graph->RemoveNode(this);
+    }
+    deadlock_graph_mu.Unlock();
+  }
+}
+
+void Mutex::AssertNotHeld() const {
+  // We have the data to allow this check only if in debug mode and deadlock
+  // detection is enabled.
+  if (kDebugMode &&
+      (mu_.load(std::memory_order_relaxed) & (kMuWriter | kMuReader)) != 0 &&
+      synch_deadlock_detection.load(std::memory_order_acquire) !=
+          OnDeadlockCycle::kIgnore) {
+    GraphId id = GetGraphId(const_cast<Mutex *>(this));
+    SynchLocksHeld *locks = Synch_GetAllLocks();
+    for (int i = 0; i != locks->n; i++) {
+      if (locks->locks[i].id == id) {
+        SynchEvent *mu_events = GetSynchEvent(this);
+        ABSL_RAW_LOG(FATAL, "thread should not hold mutex %p %s",
+                     static_cast<const void *>(this),
+                     (mu_events == nullptr ? "" : mu_events->name));
+      }
+    }
+  }
+}
+
+// Attempt to acquire *mu, and return whether successful.  The implementation
+// may spin for a short while if the lock cannot be acquired immediately.
+static bool TryAcquireWithSpinning(std::atomic<intptr_t>* mu) {
+  int c = mutex_globals.spinloop_iterations;
+  int result = -1;  // result of operation:  0=false, 1=true, -1=unknown
+
+  do {  // do/while somewhat faster on AMD
+    intptr_t v = mu->load(std::memory_order_relaxed);
+    if ((v & (kMuReader|kMuEvent)) != 0) {  // a reader or tracing -> give up
+      result = 0;
+    } else if (((v & kMuWriter) == 0) &&  // no holder -> try to acquire
+               mu->compare_exchange_strong(v, kMuWriter | v,
+                                           std::memory_order_acquire,
+                                           std::memory_order_relaxed)) {
+      result = 1;
+    }
+  } while (result == -1 && --c > 0);
+  return result == 1;
+}
+
+ABSL_XRAY_LOG_ARGS(1) void Mutex::Lock() {
+  ABSL_TSAN_MUTEX_PRE_LOCK(this, 0);
+  GraphId id = DebugOnlyDeadlockCheck(this);
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+  // try fast acquire, then spin loop
+  if ((v & (kMuWriter | kMuReader | kMuEvent)) != 0 ||
+      !mu_.compare_exchange_strong(v, kMuWriter | v,
+                                   std::memory_order_acquire,
+                                   std::memory_order_relaxed)) {
+    // try spin acquire, then slow loop
+    if (!TryAcquireWithSpinning(&this->mu_)) {
+      this->LockSlow(kExclusive, nullptr, 0);
+    }
+  }
+  DebugOnlyLockEnter(this, id);
+  ABSL_TSAN_MUTEX_POST_LOCK(this, 0, 0);
+}
+
+ABSL_XRAY_LOG_ARGS(1) void Mutex::ReaderLock() {
+  ABSL_TSAN_MUTEX_PRE_LOCK(this, __tsan_mutex_read_lock);
+  GraphId id = DebugOnlyDeadlockCheck(this);
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+  // try fast acquire, then slow loop
+  if ((v & (kMuWriter | kMuWait | kMuEvent)) != 0 ||
+      !mu_.compare_exchange_strong(v, (kMuReader | v) + kMuOne,
+                                   std::memory_order_acquire,
+                                   std::memory_order_relaxed)) {
+    this->LockSlow(kShared, nullptr, 0);
+  }
+  DebugOnlyLockEnter(this, id);
+  ABSL_TSAN_MUTEX_POST_LOCK(this, __tsan_mutex_read_lock, 0);
+}
+
+void Mutex::LockWhen(const Condition &cond) {
+  ABSL_TSAN_MUTEX_PRE_LOCK(this, 0);
+  GraphId id = DebugOnlyDeadlockCheck(this);
+  this->LockSlow(kExclusive, &cond, 0);
+  DebugOnlyLockEnter(this, id);
+  ABSL_TSAN_MUTEX_POST_LOCK(this, 0, 0);
+}
+
+bool Mutex::LockWhenWithTimeout(const Condition &cond, absl::Duration timeout) {
+  return LockWhenWithDeadline(cond, DeadlineFromTimeout(timeout));
+}
+
+bool Mutex::LockWhenWithDeadline(const Condition &cond, absl::Time deadline) {
+  ABSL_TSAN_MUTEX_PRE_LOCK(this, 0);
+  GraphId id = DebugOnlyDeadlockCheck(this);
+  bool res = LockSlowWithDeadline(kExclusive, &cond,
+                                  KernelTimeout(deadline), 0);
+  DebugOnlyLockEnter(this, id);
+  ABSL_TSAN_MUTEX_POST_LOCK(this, 0, 0);
+  return res;
+}
+
+void Mutex::ReaderLockWhen(const Condition &cond) {
+  ABSL_TSAN_MUTEX_PRE_LOCK(this, __tsan_mutex_read_lock);
+  GraphId id = DebugOnlyDeadlockCheck(this);
+  this->LockSlow(kShared, &cond, 0);
+  DebugOnlyLockEnter(this, id);
+  ABSL_TSAN_MUTEX_POST_LOCK(this, __tsan_mutex_read_lock, 0);
+}
+
+bool Mutex::ReaderLockWhenWithTimeout(const Condition &cond,
+                                      absl::Duration timeout) {
+  return ReaderLockWhenWithDeadline(cond, DeadlineFromTimeout(timeout));
+}
+
+bool Mutex::ReaderLockWhenWithDeadline(const Condition &cond,
+                                       absl::Time deadline) {
+  ABSL_TSAN_MUTEX_PRE_LOCK(this, __tsan_mutex_read_lock);
+  GraphId id = DebugOnlyDeadlockCheck(this);
+  bool res = LockSlowWithDeadline(kShared, &cond, KernelTimeout(deadline), 0);
+  DebugOnlyLockEnter(this, id);
+  ABSL_TSAN_MUTEX_POST_LOCK(this, __tsan_mutex_read_lock, 0);
+  return res;
+}
+
+void Mutex::Await(const Condition &cond) {
+  if (cond.Eval()) {    // condition already true; nothing to do
+    if (kDebugMode) {
+      this->AssertReaderHeld();
+    }
+  } else {              // normal case
+    ABSL_RAW_CHECK(this->AwaitCommon(cond, KernelTimeout::Never()),
+                   "condition untrue on return from Await");
+  }
+}
+
+bool Mutex::AwaitWithTimeout(const Condition &cond, absl::Duration timeout) {
+  return AwaitWithDeadline(cond, DeadlineFromTimeout(timeout));
+}
+
+bool Mutex::AwaitWithDeadline(const Condition &cond, absl::Time deadline) {
+  if (cond.Eval()) {      // condition already true; nothing to do
+    if (kDebugMode) {
+      this->AssertReaderHeld();
+    }
+    return true;
+  }
+
+  KernelTimeout t{deadline};
+  bool res = this->AwaitCommon(cond, t);
+  ABSL_RAW_CHECK(res || t.has_timeout(),
+                 "condition untrue on return from Await");
+  return res;
+}
+
+bool Mutex::AwaitCommon(const Condition &cond, KernelTimeout t) {
+  this->AssertReaderHeld();
+  MuHow how =
+      (mu_.load(std::memory_order_relaxed) & kMuWriter) ? kExclusive : kShared;
+  ABSL_TSAN_MUTEX_PRE_UNLOCK(this, TsanFlags(how));
+  SynchWaitParams waitp(
+      how, &cond, t, nullptr /*no cvmu*/, Synch_GetPerThreadAnnotated(this),
+      nullptr /*no cv_word*/);
+  int flags = kMuHasBlocked;
+  if (!Condition::GuaranteedEqual(&cond, nullptr)) {
+    flags |= kMuIsCond;
+  }
+  this->UnlockSlow(&waitp);
+  this->Block(waitp.thread);
+  ABSL_TSAN_MUTEX_POST_UNLOCK(this, TsanFlags(how));
+  ABSL_TSAN_MUTEX_PRE_LOCK(this, TsanFlags(how));
+  this->LockSlowLoop(&waitp, flags);
+  bool res = waitp.cond != nullptr ||  // => cond known true from LockSlowLoop
+             cond.Eval();
+  ABSL_TSAN_MUTEX_POST_LOCK(this, TsanFlags(how), 0);
+  return res;
+}
+
+ABSL_XRAY_LOG_ARGS(1) bool Mutex::TryLock() {
+  ABSL_TSAN_MUTEX_PRE_LOCK(this, __tsan_mutex_try_lock);
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+  if ((v & (kMuWriter | kMuReader | kMuEvent)) == 0 &&  // try fast acquire
+      mu_.compare_exchange_strong(v, kMuWriter | v,
+                                  std::memory_order_acquire,
+                                  std::memory_order_relaxed)) {
+    DebugOnlyLockEnter(this);
+    ABSL_TSAN_MUTEX_POST_LOCK(this, __tsan_mutex_try_lock, 0);
+    return true;
+  }
+  if ((v & kMuEvent) != 0) {              // we're recording events
+    if ((v & kExclusive->slow_need_zero) == 0 &&  // try fast acquire
+        mu_.compare_exchange_strong(
+            v, (kExclusive->fast_or | v) + kExclusive->fast_add,
+            std::memory_order_acquire, std::memory_order_relaxed)) {
+      DebugOnlyLockEnter(this);
+      PostSynchEvent(this, SYNCH_EV_TRYLOCK_SUCCESS);
+      ABSL_TSAN_MUTEX_POST_LOCK(this, __tsan_mutex_try_lock, 0);
+      return true;
+    } else {
+      PostSynchEvent(this, SYNCH_EV_TRYLOCK_FAILED);
+    }
+  }
+  ABSL_TSAN_MUTEX_POST_LOCK(
+      this, __tsan_mutex_try_lock | __tsan_mutex_try_lock_failed, 0);
+  return false;
+}
+
+ABSL_XRAY_LOG_ARGS(1) bool Mutex::ReaderTryLock() {
+  ABSL_TSAN_MUTEX_PRE_LOCK(this,
+                           __tsan_mutex_read_lock | __tsan_mutex_try_lock);
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+  // The while-loops (here and below) iterate only if the mutex word keeps
+  // changing (typically because the reader count changes) under the CAS.  We
+  // limit the number of attempts to avoid having to think about livelock.
+  int loop_limit = 5;
+  while ((v & (kMuWriter|kMuWait|kMuEvent)) == 0 && loop_limit != 0) {
+    if (mu_.compare_exchange_strong(v, (kMuReader | v) + kMuOne,
+                                    std::memory_order_acquire,
+                                    std::memory_order_relaxed)) {
+      DebugOnlyLockEnter(this);
+      ABSL_TSAN_MUTEX_POST_LOCK(
+          this, __tsan_mutex_read_lock | __tsan_mutex_try_lock, 0);
+      return true;
+    }
+    loop_limit--;
+    v = mu_.load(std::memory_order_relaxed);
+  }
+  if ((v & kMuEvent) != 0) {   // we're recording events
+    loop_limit = 5;
+    while ((v & kShared->slow_need_zero) == 0 && loop_limit != 0) {
+      if (mu_.compare_exchange_strong(v, (kMuReader | v) + kMuOne,
+                                      std::memory_order_acquire,
+                                      std::memory_order_relaxed)) {
+        DebugOnlyLockEnter(this);
+        PostSynchEvent(this, SYNCH_EV_READERTRYLOCK_SUCCESS);
+        ABSL_TSAN_MUTEX_POST_LOCK(
+            this, __tsan_mutex_read_lock | __tsan_mutex_try_lock, 0);
+        return true;
+      }
+      loop_limit--;
+      v = mu_.load(std::memory_order_relaxed);
+    }
+    if ((v & kMuEvent) != 0) {
+      PostSynchEvent(this, SYNCH_EV_READERTRYLOCK_FAILED);
+    }
+  }
+  ABSL_TSAN_MUTEX_POST_LOCK(this,
+                            __tsan_mutex_read_lock | __tsan_mutex_try_lock |
+                                __tsan_mutex_try_lock_failed,
+                            0);
+  return false;
+}
+
+ABSL_XRAY_LOG_ARGS(1) void Mutex::Unlock() {
+  ABSL_TSAN_MUTEX_PRE_UNLOCK(this, 0);
+  DebugOnlyLockLeave(this);
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+
+  if (kDebugMode && ((v & (kMuWriter | kMuReader)) != kMuWriter)) {
+    ABSL_RAW_LOG(FATAL, "Mutex unlocked when destroyed or not locked: v=0x%x",
+                 static_cast<unsigned>(v));
+  }
+
+  // should_try_cas is whether we'll try a compare-and-swap immediately.
+  // NOTE: optimized out when kDebugMode is false.
+  bool should_try_cas = ((v & (kMuEvent | kMuWriter)) == kMuWriter &&
+                          (v & (kMuWait | kMuDesig)) != kMuWait);
+  // But, we can use an alternate computation of it, that compilers
+  // currently don't find on their own.  When that changes, this function
+  // can be simplified.
+  intptr_t x = (v ^ (kMuWriter | kMuWait)) & (kMuWriter | kMuEvent);
+  intptr_t y = (v ^ (kMuWriter | kMuWait)) & (kMuWait | kMuDesig);
+  // Claim: "x == 0 && y > 0" is equal to should_try_cas.
+  // Also, because kMuWriter and kMuEvent exceed kMuDesig and kMuWait,
+  // all possible non-zero values for x exceed all possible values for y.
+  // Therefore, (x == 0 && y > 0) == (x < y).
+  if (kDebugMode && should_try_cas != (x < y)) {
+    // We would usually use PRIdPTR here, but is not correctly implemented
+    // within the android toolchain.
+    ABSL_RAW_LOG(FATAL, "internal logic error %llx %llx %llx\n",
+                 static_cast<long long>(v), static_cast<long long>(x),
+                 static_cast<long long>(y));
+  }
+  if (x < y &&
+      mu_.compare_exchange_strong(v, v & ~(kMuWrWait | kMuWriter),
+                                  std::memory_order_release,
+                                  std::memory_order_relaxed)) {
+    // fast writer release (writer with no waiters or with designated waker)
+  } else {
+    this->UnlockSlow(nullptr /*no waitp*/);  // take slow path
+  }
+  ABSL_TSAN_MUTEX_POST_UNLOCK(this, 0);
+}
+
+// Requires v to represent a reader-locked state.
+static bool ExactlyOneReader(intptr_t v) {
+  assert((v & (kMuWriter|kMuReader)) == kMuReader);
+  assert((v & kMuHigh) != 0);
+  // The more straightforward "(v & kMuHigh) == kMuOne" also works, but
+  // on some architectures the following generates slightly smaller code.
+  // It may be faster too.
+  constexpr intptr_t kMuMultipleWaitersMask = kMuHigh ^ kMuOne;
+  return (v & kMuMultipleWaitersMask) == 0;
+}
+
+ABSL_XRAY_LOG_ARGS(1) void Mutex::ReaderUnlock() {
+  ABSL_TSAN_MUTEX_PRE_UNLOCK(this, __tsan_mutex_read_lock);
+  DebugOnlyLockLeave(this);
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+  assert((v & (kMuWriter|kMuReader)) == kMuReader);
+  if ((v & (kMuReader|kMuWait|kMuEvent)) == kMuReader) {
+    // fast reader release (reader with no waiters)
+    intptr_t clear = ExactlyOneReader(v) ? kMuReader|kMuOne : kMuOne;
+    if (mu_.compare_exchange_strong(v, v - clear,
+                                    std::memory_order_release,
+                                    std::memory_order_relaxed)) {
+      ABSL_TSAN_MUTEX_POST_UNLOCK(this, __tsan_mutex_read_lock);
+      return;
+    }
+  }
+  this->UnlockSlow(nullptr /*no waitp*/);  // take slow path
+  ABSL_TSAN_MUTEX_POST_UNLOCK(this, __tsan_mutex_read_lock);
+}
+
+// The zap_desig_waker bitmask is used to clear the designated waker flag in
+// the mutex if this thread has blocked, and therefore may be the designated
+// waker.
+static const intptr_t zap_desig_waker[] = {
+    ~static_cast<intptr_t>(0),  // not blocked
+    ~static_cast<intptr_t>(
+        kMuDesig)  // blocked; turn off the designated waker bit
+};
+
+// The ignore_waiting_writers bitmask is used to ignore the existence
+// of waiting writers if a reader that has already blocked once
+// wakes up.
+static const intptr_t ignore_waiting_writers[] = {
+    ~static_cast<intptr_t>(0),  // not blocked
+    ~static_cast<intptr_t>(
+        kMuWrWait)  // blocked; pretend there are no waiting writers
+};
+
+// Internal version of LockWhen().  See LockSlowWithDeadline()
+void Mutex::LockSlow(MuHow how, const Condition *cond, int flags) {
+  ABSL_RAW_CHECK(
+      this->LockSlowWithDeadline(how, cond, KernelTimeout::Never(), flags),
+      "condition untrue on return from LockSlow");
+}
+
+// Compute cond->Eval() and tell race detectors that we do it under mutex mu.
+static inline bool EvalConditionAnnotated(const Condition *cond, Mutex *mu,
+                                          bool locking, Mutex::MuHow how) {
+  // Delicate annotation dance.
+  // We are currently inside of read/write lock/unlock operation.
+  // All memory accesses are ignored inside of mutex operations + for unlock
+  // operation tsan considers that we've already released the mutex.
+  bool res = false;
+  if (locking) {
+    // For lock we pretend that we have finished the operation,
+    // evaluate the predicate, then unlock the mutex and start locking it again
+    // to match the annotation at the end of outer lock operation.
+    // Note: we can't simply do POST_LOCK, Eval, PRE_LOCK, because then tsan
+    // will think the lock acquisition is recursive which will trigger
+    // deadlock detector.
+    ABSL_TSAN_MUTEX_POST_LOCK(mu, TsanFlags(how), 0);
+    res = cond->Eval();
+    ABSL_TSAN_MUTEX_PRE_UNLOCK(mu, TsanFlags(how));
+    ABSL_TSAN_MUTEX_POST_UNLOCK(mu, TsanFlags(how));
+    ABSL_TSAN_MUTEX_PRE_LOCK(mu, TsanFlags(how));
+  } else {
+    // Similarly, for unlock we pretend that we have unlocked the mutex,
+    // lock the mutex, evaluate the predicate, and start unlocking it again
+    // to match the annotation at the end of outer unlock operation.
+    ABSL_TSAN_MUTEX_POST_UNLOCK(mu, TsanFlags(how));
+    ABSL_TSAN_MUTEX_PRE_LOCK(mu, TsanFlags(how));
+    ABSL_TSAN_MUTEX_POST_LOCK(mu, TsanFlags(how), 0);
+    res = cond->Eval();
+    ABSL_TSAN_MUTEX_PRE_UNLOCK(mu, TsanFlags(how));
+  }
+  // Prevent unused param warnings in non-TSAN builds.
+  static_cast<void>(mu);
+  static_cast<void>(how);
+  return res;
+}
+
+// Compute cond->Eval() hiding it from race detectors.
+// We are hiding it because inside of UnlockSlow we can evaluate a predicate
+// that was just added by a concurrent Lock operation; Lock adds the predicate
+// to the internal Mutex list without actually acquiring the Mutex
+// (it only acquires the internal spinlock, which is rightfully invisible for
+// tsan). As the result there is no tsan-visible synchronization between the
+// addition and this thread. So if we would enable race detection here,
+// it would race with the predicate initialization.
+static inline bool EvalConditionIgnored(Mutex *mu, const Condition *cond) {
+  // Memory accesses are already ignored inside of lock/unlock operations,
+  // but synchronization operations are also ignored. When we evaluate the
+  // predicate we must ignore only memory accesses but not synchronization,
+  // because missed synchronization can lead to false reports later.
+  // So we "divert" (which un-ignores both memory accesses and synchronization)
+  // and then separately turn on ignores of memory accesses.
+  ABSL_TSAN_MUTEX_PRE_DIVERT(mu, 0);
+  ANNOTATE_IGNORE_READS_AND_WRITES_BEGIN();
+  bool res = cond->Eval();
+  ANNOTATE_IGNORE_READS_AND_WRITES_END();
+  ABSL_TSAN_MUTEX_POST_DIVERT(mu, 0);
+  static_cast<void>(mu);  // Prevent unused param warning in non-TSAN builds.
+  return res;
+}
+
+// Internal equivalent of *LockWhenWithDeadline(), where
+//   "t" represents the absolute timeout; !t.has_timeout() means "forever".
+//   "how" is "kShared" (for ReaderLockWhen) or "kExclusive" (for LockWhen)
+// In flags, bits are ored together:
+// - kMuHasBlocked indicates that the client has already blocked on the call so
+//   the designated waker bit must be cleared and waiting writers should not
+//   obstruct this call
+// - kMuIsCond indicates that this is a conditional acquire (condition variable,
+//   Await,  LockWhen) so contention profiling should be suppressed.
+bool Mutex::LockSlowWithDeadline(MuHow how, const Condition *cond,
+                                 KernelTimeout t, int flags) {
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+  bool unlock = false;
+  if ((v & how->fast_need_zero) == 0 &&  // try fast acquire
+      mu_.compare_exchange_strong(
+          v, (how->fast_or | (v & zap_desig_waker[flags & kMuHasBlocked])) +
+                 how->fast_add,
+          std::memory_order_acquire, std::memory_order_relaxed)) {
+    if (cond == nullptr || EvalConditionAnnotated(cond, this, true, how)) {
+      return true;
+    }
+    unlock = true;
+  }
+  SynchWaitParams waitp(
+      how, cond, t, nullptr /*no cvmu*/, Synch_GetPerThreadAnnotated(this),
+      nullptr /*no cv_word*/);
+  if (!Condition::GuaranteedEqual(cond, nullptr)) {
+    flags |= kMuIsCond;
+  }
+  if (unlock) {
+    this->UnlockSlow(&waitp);
+    this->Block(waitp.thread);
+    flags |= kMuHasBlocked;
+  }
+  this->LockSlowLoop(&waitp, flags);
+  return waitp.cond != nullptr ||  // => cond known true from LockSlowLoop
+         cond == nullptr || EvalConditionAnnotated(cond, this, true, how);
+}
+
+// RAW_CHECK_FMT() takes a condition, a printf-style format std::string, and
+// the printf-style argument list.   The format std::string must be a literal.
+// Arguments after the first are not evaluated unless the condition is true.
+#define RAW_CHECK_FMT(cond, ...)                                   \
+  do {                                                             \
+    if (ABSL_PREDICT_FALSE(!(cond))) {                             \
+      ABSL_RAW_LOG(FATAL, "Check " #cond " failed: " __VA_ARGS__); \
+    }                                                              \
+  } while (0)
+
+static void CheckForMutexCorruption(intptr_t v, const char* label) {
+  // Test for either of two situations that should not occur in v:
+  //   kMuWriter and kMuReader
+  //   kMuWrWait and !kMuWait
+  const intptr_t w = v ^ kMuWait;
+  // By flipping that bit, we can now test for:
+  //   kMuWriter and kMuReader in w
+  //   kMuWrWait and kMuWait in w
+  // We've chosen these two pairs of values to be so that they will overlap,
+  // respectively, when the word is left shifted by three.  This allows us to
+  // save a branch in the common (correct) case of them not being coincident.
+  static_assert(kMuReader << 3 == kMuWriter, "must match");
+  static_assert(kMuWait << 3 == kMuWrWait, "must match");
+  if (ABSL_PREDICT_TRUE((w & (w << 3) & (kMuWriter | kMuWrWait)) == 0)) return;
+  RAW_CHECK_FMT((v & (kMuWriter | kMuReader)) != (kMuWriter | kMuReader),
+                "%s: Mutex corrupt: both reader and writer lock held: %p",
+                label, reinterpret_cast<void *>(v));
+  RAW_CHECK_FMT((v & (kMuWait | kMuWrWait)) != kMuWrWait,
+                "%s: Mutex corrupt: waiting writer with no waiters: %p",
+                label, reinterpret_cast<void *>(v));
+  assert(false);
+}
+
+void Mutex::LockSlowLoop(SynchWaitParams *waitp, int flags) {
+  int c = 0;
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+  if ((v & kMuEvent) != 0) {
+    PostSynchEvent(this,
+         waitp->how == kExclusive?  SYNCH_EV_LOCK: SYNCH_EV_READERLOCK);
+  }
+  ABSL_RAW_CHECK(
+      waitp->thread->waitp == nullptr || waitp->thread->suppress_fatal_errors,
+      "detected illegal recursion into Mutex code");
+  for (;;) {
+    v = mu_.load(std::memory_order_relaxed);
+    CheckForMutexCorruption(v, "Lock");
+    if ((v & waitp->how->slow_need_zero) == 0) {
+      if (mu_.compare_exchange_strong(
+              v, (waitp->how->fast_or |
+                  (v & zap_desig_waker[flags & kMuHasBlocked])) +
+                     waitp->how->fast_add,
+              std::memory_order_acquire, std::memory_order_relaxed)) {
+        if (waitp->cond == nullptr ||
+            EvalConditionAnnotated(waitp->cond, this, true, waitp->how)) {
+          break;  // we timed out, or condition true, so return
+        }
+        this->UnlockSlow(waitp);  // got lock but condition false
+        this->Block(waitp->thread);
+        flags |= kMuHasBlocked;
+        c = 0;
+      }
+    } else {                      // need to access waiter list
+      bool dowait = false;
+      if ((v & (kMuSpin|kMuWait)) == 0) {   // no waiters
+        // This thread tries to become the one and only waiter.
+        PerThreadSynch *new_h = Enqueue(nullptr, waitp, v, flags);
+        intptr_t nv = (v & zap_desig_waker[flags & kMuHasBlocked] & kMuLow) |
+                      kMuWait;
+        ABSL_RAW_CHECK(new_h != nullptr, "Enqueue to empty list failed");
+        if (waitp->how == kExclusive && (v & kMuReader) != 0) {
+          nv |= kMuWrWait;
+        }
+        if (mu_.compare_exchange_strong(
+                v, reinterpret_cast<intptr_t>(new_h) | nv,
+                std::memory_order_release, std::memory_order_relaxed)) {
+          dowait = true;
+        } else {            // attempted Enqueue() failed
+          // zero out the waitp field set by Enqueue()
+          waitp->thread->waitp = nullptr;
+        }
+      } else if ((v & waitp->how->slow_inc_need_zero &
+                  ignore_waiting_writers[flags & kMuHasBlocked]) == 0) {
+        // This is a reader that needs to increment the reader count,
+        // but the count is currently held in the last waiter.
+        if (mu_.compare_exchange_strong(
+                v, (v & zap_desig_waker[flags & kMuHasBlocked]) | kMuSpin |
+                       kMuReader,
+                std::memory_order_acquire, std::memory_order_relaxed)) {
+          PerThreadSynch *h = GetPerThreadSynch(v);
+          h->readers += kMuOne;       // inc reader count in waiter
+          do {                        // release spinlock
+            v = mu_.load(std::memory_order_relaxed);
+          } while (!mu_.compare_exchange_weak(v, (v & ~kMuSpin) | kMuReader,
+                                              std::memory_order_release,
+                                              std::memory_order_relaxed));
+          if (waitp->cond == nullptr ||
+              EvalConditionAnnotated(waitp->cond, this, true, waitp->how)) {
+            break;  // we timed out, or condition true, so return
+          }
+          this->UnlockSlow(waitp);           // got lock but condition false
+          this->Block(waitp->thread);
+          flags |= kMuHasBlocked;
+          c = 0;
+        }
+      } else if ((v & kMuSpin) == 0 &&  // attempt to queue ourselves
+                 mu_.compare_exchange_strong(
+                     v, (v & zap_desig_waker[flags & kMuHasBlocked]) | kMuSpin |
+                            kMuWait,
+                     std::memory_order_acquire, std::memory_order_relaxed)) {
+        PerThreadSynch *h = GetPerThreadSynch(v);
+        PerThreadSynch *new_h = Enqueue(h, waitp, v, flags);
+        intptr_t wr_wait = 0;
+        ABSL_RAW_CHECK(new_h != nullptr, "Enqueue to list failed");
+        if (waitp->how == kExclusive && (v & kMuReader) != 0) {
+          wr_wait = kMuWrWait;      // give priority to a waiting writer
+        }
+        do {                        // release spinlock
+          v = mu_.load(std::memory_order_relaxed);
+        } while (!mu_.compare_exchange_weak(
+            v, (v & (kMuLow & ~kMuSpin)) | kMuWait | wr_wait |
+            reinterpret_cast<intptr_t>(new_h),
+            std::memory_order_release, std::memory_order_relaxed));
+        dowait = true;
+      }
+      if (dowait) {
+        this->Block(waitp->thread);  // wait until removed from list or timeout
+        flags |= kMuHasBlocked;
+        c = 0;
+      }
+    }
+    ABSL_RAW_CHECK(
+        waitp->thread->waitp == nullptr || waitp->thread->suppress_fatal_errors,
+        "detected illegal recursion into Mutex code");
+    c = Delay(c, GENTLE);          // delay, then try again
+  }
+  ABSL_RAW_CHECK(
+      waitp->thread->waitp == nullptr || waitp->thread->suppress_fatal_errors,
+      "detected illegal recursion into Mutex code");
+  if ((v & kMuEvent) != 0) {
+    PostSynchEvent(this,
+                   waitp->how == kExclusive? SYNCH_EV_LOCK_RETURNING :
+                                      SYNCH_EV_READERLOCK_RETURNING);
+  }
+}
+
+// Unlock this mutex, which is held by the current thread.
+// If waitp is non-zero, it must be the wait parameters for the current thread
+// which holds the lock but is not runnable because its condition is false
+// or it n the process of blocking on a condition variable; it must requeue
+// itself on the mutex/condvar to wait for its condition to become true.
+void Mutex::UnlockSlow(SynchWaitParams *waitp) {
+  intptr_t v = mu_.load(std::memory_order_relaxed);
+  this->AssertReaderHeld();
+  CheckForMutexCorruption(v, "Unlock");
+  if ((v & kMuEvent) != 0) {
+    PostSynchEvent(this,
+                (v & kMuWriter) != 0? SYNCH_EV_UNLOCK: SYNCH_EV_READERUNLOCK);
+  }
+  int c = 0;
+  // the waiter under consideration to wake, or zero
+  PerThreadSynch *w = nullptr;
+  // the predecessor to w or zero
+  PerThreadSynch *pw = nullptr;
+  // head of the list searched previously, or zero
+  PerThreadSynch *old_h = nullptr;
+  // a condition that's known to be false.
+  const Condition *known_false = nullptr;
+  PerThreadSynch *wake_list = kPerThreadSynchNull;   // list of threads to wake
+  intptr_t wr_wait = 0;        // set to kMuWrWait if we wake a reader and a
+                               // later writer could have acquired the lock
+                               // (starvation avoidance)
+  ABSL_RAW_CHECK(waitp == nullptr || waitp->thread->waitp == nullptr ||
+                     waitp->thread->suppress_fatal_errors,
+                 "detected illegal recursion into Mutex code");
+  // This loop finds threads wake_list to wakeup if any, and removes them from
+  // the list of waiters.  In addition, it places waitp.thread on the queue of
+  // waiters if waitp is non-zero.
+  for (;;) {
+    v = mu_.load(std::memory_order_relaxed);
+    if ((v & kMuWriter) != 0 && (v & (kMuWait | kMuDesig)) != kMuWait &&
+        waitp == nullptr) {
+      // fast writer release (writer with no waiters or with designated waker)
+      if (mu_.compare_exchange_strong(v, v & ~(kMuWrWait | kMuWriter),
+                                      std::memory_order_release,
+                                      std::memory_order_relaxed)) {
+        return;
+      }
+    } else if ((v & (kMuReader | kMuWait)) == kMuReader && waitp == nullptr) {
+      // fast reader release (reader with no waiters)
+      intptr_t clear = ExactlyOneReader(v) ? kMuReader | kMuOne : kMuOne;
+      if (mu_.compare_exchange_strong(v, v - clear,
+                                      std::memory_order_release,
+                                      std::memory_order_relaxed)) {
+        return;
+      }
+    } else if ((v & kMuSpin) == 0 &&  // attempt to get spinlock
+               mu_.compare_exchange_strong(v, v | kMuSpin,
+                                           std::memory_order_acquire,
+                                           std::memory_order_relaxed)) {
+      if ((v & kMuWait) == 0) {       // no one to wake
+        intptr_t nv;
+        bool do_enqueue = true;  // always Enqueue() the first time
+        ABSL_RAW_CHECK(waitp != nullptr,
+                       "UnlockSlow is confused");  // about to sleep
+        do {    // must loop to release spinlock as reader count may change
+          v = mu_.load(std::memory_order_relaxed);
+          // decrement reader count if there are readers
+          intptr_t new_readers = (v >= kMuOne)?  v - kMuOne : v;
+          PerThreadSynch *new_h = nullptr;
+          if (do_enqueue) {
+            // If we are enqueuing on a CondVar (waitp->cv_word != nullptr) then
+            // we must not retry here.  The initial attempt will always have
+            // succeeded, further attempts would enqueue us against *this due to
+            // Fer() handling.
+            do_enqueue = (waitp->cv_word == nullptr);
+            new_h = Enqueue(nullptr, waitp, new_readers, kMuIsCond);
+          }
+          intptr_t clear = kMuWrWait | kMuWriter;  // by default clear write bit
+          if ((v & kMuWriter) == 0 && ExactlyOneReader(v)) {  // last reader
+            clear = kMuWrWait | kMuReader;                    // clear read bit
+          }
+          nv = (v & kMuLow & ~clear & ~kMuSpin);
+          if (new_h != nullptr) {
+            nv |= kMuWait | reinterpret_cast<intptr_t>(new_h);
+          } else {  // new_h could be nullptr if we queued ourselves on a
+                    // CondVar
+            // In that case, we must place the reader count back in the mutex
+            // word, as Enqueue() did not store it in the new waiter.
+            nv |= new_readers & kMuHigh;
+          }
+          // release spinlock & our lock; retry if reader-count changed
+          // (writer count cannot change since we hold lock)
+        } while (!mu_.compare_exchange_weak(v, nv,
+                                            std::memory_order_release,
+                                            std::memory_order_relaxed));
+        break;
+      }
+
+      // There are waiters.
+      // Set h to the head of the circular waiter list.
+      PerThreadSynch *h = GetPerThreadSynch(v);
+      if ((v & kMuReader) != 0 && (h->readers & kMuHigh) > kMuOne) {
+        // a reader but not the last
+        h->readers -= kMuOne;  // release our lock
+        intptr_t nv = v;       // normally just release spinlock
+        if (waitp != nullptr) {  // but waitp!=nullptr => must queue ourselves
+          PerThreadSynch *new_h = Enqueue(h, waitp, v, kMuIsCond);
+          ABSL_RAW_CHECK(new_h != nullptr,
+                         "waiters disappeared during Enqueue()!");
+          nv &= kMuLow;
+          nv |= kMuWait | reinterpret_cast<intptr_t>(new_h);
+        }
+        mu_.store(nv, std::memory_order_release);  // release spinlock
+        // can release with a store because there were waiters
+        break;
+      }
+
+      // Either we didn't search before, or we marked the queue
+      // as "maybe_unlocking" and no one else should have changed it.
+      ABSL_RAW_CHECK(old_h == nullptr || h->maybe_unlocking,
+                     "Mutex queue changed beneath us");
+
+      // The lock is becoming free, and there's a waiter
+      if (old_h != nullptr &&
+          !old_h->may_skip) {                  // we used old_h as a terminator
+        old_h->may_skip = true;                // allow old_h to skip once more
+        ABSL_RAW_CHECK(old_h->skip == nullptr, "illegal skip from head");
+        if (h != old_h && MuSameCondition(old_h, old_h->next)) {
+          old_h->skip = old_h->next;  // old_h not head & can skip to successor
+        }
+      }
+      if (h->next->waitp->how == kExclusive &&
+          Condition::GuaranteedEqual(h->next->waitp->cond, nullptr)) {
+        // easy case: writer with no condition; no need to search
+        pw = h;                       // wake w, the successor of h (=pw)
+        w = h->next;
+        w->wake = true;
+        // We are waking up a writer.  This writer may be racing against
+        // an already awake reader for the lock.  We want the
+        // writer to usually win this race,
+        // because if it doesn't, we can potentially keep taking a reader
+        // perpetually and writers will starve.  Worse than
+        // that, this can also starve other readers if kMuWrWait gets set
+        // later.
+        wr_wait = kMuWrWait;
+      } else if (w != nullptr && (w->waitp->how == kExclusive || h == old_h)) {
+        // we found a waiter w to wake on a previous iteration and either it's
+        // a writer, or we've searched the entire list so we have all the
+        // readers.
+        if (pw == nullptr) {  // if w's predecessor is unknown, it must be h
+          pw = h;
+        }
+      } else {
+        // At this point we don't know all the waiters to wake, and the first
+        // waiter has a condition or is a reader.  We avoid searching over
+        // waiters we've searched on previous iterations by starting at
+        // old_h if it's set.  If old_h==h, there's no one to wakeup at all.
+        if (old_h == h) {      // we've searched before, and nothing's new
+                               // so there's no one to wake.
+          intptr_t nv = (v & ~(kMuReader|kMuWriter|kMuWrWait));
+          h->readers = 0;
+          h->maybe_unlocking = false;   // finished unlocking
+          if (waitp != nullptr) {       // we must queue ourselves and sleep
+            PerThreadSynch *new_h = Enqueue(h, waitp, v, kMuIsCond);
+            nv &= kMuLow;
+            if (new_h != nullptr) {
+              nv |= kMuWait | reinterpret_cast<intptr_t>(new_h);
+            }  // else new_h could be nullptr if we queued ourselves on a
+               // CondVar
+          }
+          // release spinlock & lock
+          // can release with a store because there were waiters
+          mu_.store(nv, std::memory_order_release);
+          break;
+        }
+
+        // set up to walk the list
+        PerThreadSynch *w_walk;   // current waiter during list walk
+        PerThreadSynch *pw_walk;  // previous waiter during list walk
+        if (old_h != nullptr) {  // we've searched up to old_h before
+          pw_walk = old_h;
+          w_walk = old_h->next;
+        } else {            // no prior search, start at beginning
+          pw_walk =
+              nullptr;  // h->next's predecessor may change; don't record it
+          w_walk = h->next;
+        }
+
+        h->may_skip = false;  // ensure we never skip past h in future searches
+                              // even if other waiters are queued after it.
+        ABSL_RAW_CHECK(h->skip == nullptr, "illegal skip from head");
+
+        h->maybe_unlocking = true;  // we're about to scan the waiter list
+                                    // without the spinlock held.
+                                    // Enqueue must be conservative about
+                                    // priority queuing.
+
+        // We must release the spinlock to evaluate the conditions.
+        mu_.store(v, std::memory_order_release);  // release just spinlock
+        // can release with a store because there were waiters
+
+        // h is the last waiter queued, and w_walk the first unsearched waiter.
+        // Without the spinlock, the locations mu_ and h->next may now change
+        // underneath us, but since we hold the lock itself, the only legal
+        // change is to add waiters between h and w_walk.  Therefore, it's safe
+        // to walk the path from w_walk to h inclusive. (TryRemove() can remove
+        // a waiter anywhere, but it acquires both the spinlock and the Mutex)
+
+        old_h = h;        // remember we searched to here
+
+        // Walk the path upto and including h looking for waiters we can wake.
+        while (pw_walk != h) {
+          w_walk->wake = false;
+          if (w_walk->waitp->cond ==
+                  nullptr ||  // no condition => vacuously true OR
+              (w_walk->waitp->cond != known_false &&
+               // this thread's condition is not known false, AND
+               //  is in fact true
+               EvalConditionIgnored(this, w_walk->waitp->cond))) {
+            if (w == nullptr) {
+              w_walk->wake = true;    // can wake this waiter
+              w = w_walk;
+              pw = pw_walk;
+              if (w_walk->waitp->how == kExclusive) {
+                wr_wait = kMuWrWait;
+                break;                // bail if waking this writer
+              }
+            } else if (w_walk->waitp->how == kShared) {  // wake if a reader
+              w_walk->wake = true;
+            } else {   // writer with true condition
+              wr_wait = kMuWrWait;
+            }
+          } else {                  // can't wake; condition false
+            known_false = w_walk->waitp->cond;  // remember last false condition
+          }
+          if (w_walk->wake) {   // we're waking reader w_walk
+            pw_walk = w_walk;   // don't skip similar waiters
+          } else {              // not waking; skip as much as possible
+            pw_walk = Skip(w_walk);
+          }
+          // If pw_walk == h, then load of pw_walk->next can race with
+          // concurrent write in Enqueue(). However, at the same time
+          // we do not need to do the load, because we will bail out
+          // from the loop anyway.
+          if (pw_walk != h) {
+            w_walk = pw_walk->next;
+          }
+        }
+
+        continue;  // restart for(;;)-loop to wakeup w or to find more waiters
+      }
+      ABSL_RAW_CHECK(pw->next == w, "pw not w's predecessor");
+      // The first (and perhaps only) waiter we've chosen to wake is w, whose
+      // predecessor is pw.  If w is a reader, we must wake all the other
+      // waiters with wake==true as well.  We may also need to queue
+      // ourselves if waitp != null.  The spinlock and the lock are still
+      // held.
+
+      // This traverses the list in [ pw->next, h ], where h is the head,
+      // removing all elements with wake==true and placing them in the
+      // singly-linked list wake_list.  Returns the new head.
+      h = DequeueAllWakeable(h, pw, &wake_list);
+
+      intptr_t nv = (v & kMuEvent) | kMuDesig;
+                                             // assume no waiters left,
+                                             // set kMuDesig for INV1a
+
+      if (waitp != nullptr) {  // we must queue ourselves and sleep
+        h = Enqueue(h, waitp, v, kMuIsCond);
+        // h is new last waiter; could be null if we queued ourselves on a
+        // CondVar
+      }
+
+      ABSL_RAW_CHECK(wake_list != kPerThreadSynchNull,
+                     "unexpected empty wake list");
+
+      if (h != nullptr) {  // there are waiters left
+        h->readers = 0;
+        h->maybe_unlocking = false;     // finished unlocking
+        nv |= wr_wait | kMuWait | reinterpret_cast<intptr_t>(h);
+      }
+
+      // release both spinlock & lock
+      // can release with a store because there were waiters
+      mu_.store(nv, std::memory_order_release);
+      break;  // out of for(;;)-loop
+    }
+    c = Delay(c, AGGRESSIVE);  // aggressive here; no one can proceed till we do
+  }                            // end of for(;;)-loop
+
+  if (wake_list != kPerThreadSynchNull) {
+    int64_t enqueue_timestamp = wake_list->waitp->contention_start_cycles;
+    bool cond_waiter = wake_list->cond_waiter;
+    do {
+      wake_list = Wakeup(wake_list);              // wake waiters
+    } while (wake_list != kPerThreadSynchNull);
+    if (!cond_waiter) {
+      // Sample lock contention events only if the (first) waiter was trying to
+      // acquire the lock, not waiting on a condition variable or Condition.
+      int64_t wait_cycles = base_internal::CycleClock::Now() - enqueue_timestamp;
+      mutex_tracer("slow release", this, wait_cycles);
+      ABSL_TSAN_MUTEX_PRE_DIVERT(this, 0);
+      submit_profile_data(enqueue_timestamp);
+      ABSL_TSAN_MUTEX_POST_DIVERT(this, 0);
+    }
+  }
+}
+
+// Used by CondVar implementation to reacquire mutex after waking from
+// condition variable.  This routine is used instead of Lock() because the
+// waiting thread may have been moved from the condition variable queue to the
+// mutex queue without a wakeup, by Trans().  In that case, when the thread is
+// finally woken, the woken thread will believe it has been woken from the
+// condition variable (i.e. its PC will be in when in the CondVar code), when
+// in fact it has just been woken from the mutex.  Thus, it must enter the slow
+// path of the mutex in the same state as if it had just woken from the mutex.
+// That is, it must ensure to clear kMuDesig (INV1b).
+void Mutex::Trans(MuHow how) {
+  this->LockSlow(how, nullptr, kMuHasBlocked | kMuIsCond);
+}
+
+// Used by CondVar implementation to effectively wake thread w from the
+// condition variable.  If this mutex is free, we simply wake the thread.
+// It will later acquire the mutex with high probability.  Otherwise, we
+// enqueue thread w on this mutex.
+void Mutex::Fer(PerThreadSynch *w) {
+  int c = 0;
+  ABSL_RAW_CHECK(w->waitp->cond == nullptr,
+                 "Mutex::Fer while waiting on Condition");
+  ABSL_RAW_CHECK(!w->waitp->timeout.has_timeout(),
+                 "Mutex::Fer while in timed wait");
+  ABSL_RAW_CHECK(w->waitp->cv_word == nullptr,
+                 "Mutex::Fer with pending CondVar queueing");
+  for (;;) {
+    intptr_t v = mu_.load(std::memory_order_relaxed);
+    // Note: must not queue if the mutex is unlocked (nobody will wake it).
+    // For example, we can have only kMuWait (conditional) or maybe
+    // kMuWait|kMuWrWait.
+    // conflicting != 0 implies that the waking thread cannot currently take
+    // the mutex, which in turn implies that someone else has it and can wake
+    // us if we queue.
+    const intptr_t conflicting =
+        kMuWriter | (w->waitp->how == kShared ? 0 : kMuReader);
+    if ((v & conflicting) == 0) {
+      w->next = nullptr;
+      w->state.store(PerThreadSynch::kAvailable, std::memory_order_release);
+      IncrementSynchSem(this, w);
+      return;
+    } else {
+      if ((v & (kMuSpin|kMuWait)) == 0) {       // no waiters
+        // This thread tries to become the one and only waiter.
+        PerThreadSynch *new_h = Enqueue(nullptr, w->waitp, v, kMuIsCond);
+        ABSL_RAW_CHECK(new_h != nullptr,
+                       "Enqueue failed");  // we must queue ourselves
+        if (mu_.compare_exchange_strong(
+                v, reinterpret_cast<intptr_t>(new_h) | (v & kMuLow) | kMuWait,
+                std::memory_order_release, std::memory_order_relaxed)) {
+          return;
+        }
+      } else if ((v & kMuSpin) == 0 &&
+                 mu_.compare_exchange_strong(v, v | kMuSpin | kMuWait)) {
+        PerThreadSynch *h = GetPerThreadSynch(v);
+        PerThreadSynch *new_h = Enqueue(h, w->waitp, v, kMuIsCond);
+        ABSL_RAW_CHECK(new_h != nullptr,
+                       "Enqueue failed");  // we must queue ourselves
+        do {
+          v = mu_.load(std::memory_order_relaxed);
+        } while (!mu_.compare_exchange_weak(
+            v,
+            (v & kMuLow & ~kMuSpin) | kMuWait |
+                reinterpret_cast<intptr_t>(new_h),
+            std::memory_order_release, std::memory_order_relaxed));
+        return;
+      }
+    }
+    c = Delay(c, GENTLE);
+  }
+}
+
+void Mutex::AssertHeld() const {
+  if ((mu_.load(std::memory_order_relaxed) & kMuWriter) == 0) {
+    SynchEvent *e = GetSynchEvent(this);
+    ABSL_RAW_LOG(FATAL, "thread should hold write lock on Mutex %p %s",
+                 static_cast<const void *>(this),
+                 (e == nullptr ? "" : e->name));
+  }
+}
+
+void Mutex::AssertReaderHeld() const {
+  if ((mu_.load(std::memory_order_relaxed) & (kMuReader | kMuWriter)) == 0) {
+    SynchEvent *e = GetSynchEvent(this);
+    ABSL_RAW_LOG(
+        FATAL, "thread should hold at least a read lock on Mutex %p %s",
+        static_cast<const void *>(this), (e == nullptr ? "" : e->name));
+  }
+}
+
+// -------------------------------- condition variables
+static const intptr_t kCvSpin = 0x0001L;   // spinlock protects waiter list
+static const intptr_t kCvEvent = 0x0002L;  // record events
+
+static const intptr_t kCvLow = 0x0003L;  // low order bits of CV
+
+// Hack to make constant values available to gdb pretty printer
+enum { kGdbCvSpin = kCvSpin, kGdbCvEvent = kCvEvent, kGdbCvLow = kCvLow, };
+
+static_assert(PerThreadSynch::kAlignment > kCvLow,
+              "PerThreadSynch::kAlignment must be greater than kCvLow");
+
+void CondVar::EnableDebugLog(const char *name) {
+  SynchEvent *e = EnsureSynchEvent(&this->cv_, name, kCvEvent, kCvSpin);
+  e->log = true;
+  UnrefSynchEvent(e);
+}
+
+CondVar::~CondVar() {
+  if ((cv_.load(std::memory_order_relaxed) & kCvEvent) != 0) {
+    ForgetSynchEvent(&this->cv_, kCvEvent, kCvSpin);
+  }
+}
+
+
+// Remove thread s from the list of waiters on this condition variable.
+void CondVar::Remove(PerThreadSynch *s) {
+  intptr_t v;
+  int c = 0;
+  for (v = cv_.load(std::memory_order_relaxed);;
+       v = cv_.load(std::memory_order_relaxed)) {
+    if ((v & kCvSpin) == 0 &&  // attempt to acquire spinlock
+        cv_.compare_exchange_strong(v, v | kCvSpin,
+                                    std::memory_order_acquire,
+                                    std::memory_order_relaxed)) {
+      PerThreadSynch *h = reinterpret_cast<PerThreadSynch *>(v & ~kCvLow);
+      if (h != nullptr) {
+        PerThreadSynch *w = h;
+        while (w->next != s && w->next != h) {  // search for thread
+          w = w->next;
+        }
+        if (w->next == s) {           // found thread; remove it
+          w->next = s->next;
+          if (h == s) {
+            h = (w == s) ? nullptr : w;
+          }
+          s->next = nullptr;
+          s->state.store(PerThreadSynch::kAvailable, std::memory_order_release);
+        }
+      }
+                                      // release spinlock
+      cv_.store((v & kCvEvent) | reinterpret_cast<intptr_t>(h),
+                std::memory_order_release);
+      return;
+    } else {
+      c = Delay(c, GENTLE);            // try again after a delay
+    }
+  }
+}
+
+// Queue thread waitp->thread on condition variable word cv_word using
+// wait parameters waitp.
+// We split this into a separate routine, rather than simply doing it as part
+// of WaitCommon().  If we were to queue ourselves on the condition variable
+// before calling Mutex::UnlockSlow(), the Mutex code might be re-entered (via
+// the logging code, or via a Condition function) and might potentially attempt
+// to block this thread.  That would be a problem if the thread were already on
+// a the condition variable waiter queue.  Thus, we use the waitp->cv_word
+// to tell the unlock code to call CondVarEnqueue() to queue the thread on the
+// condition variable queue just before the mutex is to be unlocked, and (most
+// importantly) after any call to an external routine that might re-enter the
+// mutex code.
+static void CondVarEnqueue(SynchWaitParams *waitp) {
+  // This thread might be transferred to the Mutex queue by Fer() when
+  // we are woken.  To make sure that is what happens, Enqueue() doesn't
+  // call CondVarEnqueue() again but instead uses its normal code.  We
+  // must do this before we queue ourselves so that cv_word will be null
+  // when seen by the dequeuer, who may wish immediately to requeue
+  // this thread on another queue.
+  std::atomic<intptr_t> *cv_word = waitp->cv_word;
+  waitp->cv_word = nullptr;
+
+  intptr_t v = cv_word->load(std::memory_order_relaxed);
+  int c = 0;
+  while ((v & kCvSpin) != 0 ||  // acquire spinlock
+         !cv_word->compare_exchange_weak(v, v | kCvSpin,
+                                         std::memory_order_acquire,
+                                         std::memory_order_relaxed)) {
+    c = Delay(c, GENTLE);
+    v = cv_word->load(std::memory_order_relaxed);
+  }
+  ABSL_RAW_CHECK(waitp->thread->waitp == nullptr, "waiting when shouldn't be");
+  waitp->thread->waitp = waitp;      // prepare ourselves for waiting
+  PerThreadSynch *h = reinterpret_cast<PerThreadSynch *>(v & ~kCvLow);
+  if (h == nullptr) {  // add this thread to waiter list
+    waitp->thread->next = waitp->thread;
+  } else {
+    waitp->thread->next = h->next;
+    h->next = waitp->thread;
+  }
+  waitp->thread->state.store(PerThreadSynch::kQueued,
+                             std::memory_order_relaxed);
+  cv_word->store((v & kCvEvent) | reinterpret_cast<intptr_t>(waitp->thread),
+                 std::memory_order_release);
+}
+
+bool CondVar::WaitCommon(Mutex *mutex, KernelTimeout t) {
+  bool rc = false;          // return value; true iff we timed-out
+
+  intptr_t mutex_v = mutex->mu_.load(std::memory_order_relaxed);
+  Mutex::MuHow mutex_how = ((mutex_v & kMuWriter) != 0) ? kExclusive : kShared;
+  ABSL_TSAN_MUTEX_PRE_UNLOCK(mutex, TsanFlags(mutex_how));
+
+  // maybe trace this call
+  intptr_t v = cv_.load(std::memory_order_relaxed);
+  cond_var_tracer("Wait", this);
+  if ((v & kCvEvent) != 0) {
+    PostSynchEvent(this, SYNCH_EV_WAIT);
+  }
+
+  // Release mu and wait on condition variable.
+  SynchWaitParams waitp(mutex_how, nullptr, t, mutex,
+                        Synch_GetPerThreadAnnotated(mutex), &cv_);
+  // UnlockSlow() will call CondVarEnqueue() just before releasing the
+  // Mutex, thus queuing this thread on the condition variable.  See
+  // CondVarEnqueue() for the reasons.
+  mutex->UnlockSlow(&waitp);
+
+  // wait for signal
+  while (waitp.thread->state.load(std::memory_order_acquire) ==
+         PerThreadSynch::kQueued) {
+    if (!Mutex::DecrementSynchSem(mutex, waitp.thread, t)) {
+      this->Remove(waitp.thread);
+      rc = true;
+    }
+  }
+
+  ABSL_RAW_CHECK(waitp.thread->waitp != nullptr, "not waiting when should be");
+  waitp.thread->waitp = nullptr;  // cleanup
+
+  // maybe trace this call
+  cond_var_tracer("Unwait", this);
+  if ((v & kCvEvent) != 0) {
+    PostSynchEvent(this, SYNCH_EV_WAIT_RETURNING);
+  }
+
+  // From synchronization point of view Wait is unlock of the mutex followed
+  // by lock of the mutex. We've annotated start of unlock in the beginning
+  // of the function. Now, finish unlock and annotate lock of the mutex.
+  // (Trans is effectively lock).
+  ABSL_TSAN_MUTEX_POST_UNLOCK(mutex, TsanFlags(mutex_how));
+  ABSL_TSAN_MUTEX_PRE_LOCK(mutex, TsanFlags(mutex_how));
+  mutex->Trans(mutex_how);  // Reacquire mutex
+  ABSL_TSAN_MUTEX_POST_LOCK(mutex, TsanFlags(mutex_how), 0);
+  return rc;
+}
+
+bool CondVar::WaitWithTimeout(Mutex *mu, absl::Duration timeout) {
+  return WaitWithDeadline(mu, DeadlineFromTimeout(timeout));
+}
+
+bool CondVar::WaitWithDeadline(Mutex *mu, absl::Time deadline) {
+  return WaitCommon(mu, KernelTimeout(deadline));
+}
+
+void CondVar::Wait(Mutex *mu) {
+  WaitCommon(mu, KernelTimeout::Never());
+}
+
+// Wake thread w
+// If it was a timed wait, w will be waiting on w->cv
+// Otherwise, if it was not a Mutex mutex, w will be waiting on w->sem
+// Otherwise, w is transferred to the Mutex mutex via Mutex::Fer().
+void CondVar::Wakeup(PerThreadSynch *w) {
+  if (w->waitp->timeout.has_timeout() || w->waitp->cvmu == nullptr) {
+    // The waiting thread only needs to observe "w->state == kAvailable" to be
+    // released, we must cache "cvmu" before clearing "next".
+    Mutex *mu = w->waitp->cvmu;
+    w->next = nullptr;
+    w->state.store(PerThreadSynch::kAvailable, std::memory_order_release);
+    Mutex::IncrementSynchSem(mu, w);
+  } else {
+    w->waitp->cvmu->Fer(w);
+  }
+}
+
+void CondVar::Signal() {
+  ABSL_TSAN_MUTEX_PRE_SIGNAL(0, 0);
+  intptr_t v;
+  int c = 0;
+  for (v = cv_.load(std::memory_order_relaxed); v != 0;
+       v = cv_.load(std::memory_order_relaxed)) {
+    if ((v & kCvSpin) == 0 &&  // attempt to acquire spinlock
+        cv_.compare_exchange_strong(v, v | kCvSpin,
+                                    std::memory_order_acquire,
+                                    std::memory_order_relaxed)) {
+      PerThreadSynch *h = reinterpret_cast<PerThreadSynch *>(v & ~kCvLow);
+      PerThreadSynch *w = nullptr;
+      if (h != nullptr) {  // remove first waiter
+        w = h->next;
+        if (w == h) {
+          h = nullptr;
+        } else {
+          h->next = w->next;
+        }
+      }
+                                      // release spinlock
+      cv_.store((v & kCvEvent) | reinterpret_cast<intptr_t>(h),
+                std::memory_order_release);
+      if (w != nullptr) {
+        CondVar::Wakeup(w);                // wake waiter, if there was one
+        cond_var_tracer("Signal wakeup", this);
+      }
+      if ((v & kCvEvent) != 0) {
+        PostSynchEvent(this, SYNCH_EV_SIGNAL);
+      }
+      ABSL_TSAN_MUTEX_POST_SIGNAL(0, 0);
+      return;
+    } else {
+      c = Delay(c, GENTLE);
+    }
+  }
+  ABSL_TSAN_MUTEX_POST_SIGNAL(0, 0);
+}
+
+void CondVar::SignalAll () {
+  ABSL_TSAN_MUTEX_PRE_SIGNAL(0, 0);
+  intptr_t v;
+  int c = 0;
+  for (v = cv_.load(std::memory_order_relaxed); v != 0;
+       v = cv_.load(std::memory_order_relaxed)) {
+    // empty the list if spinlock free
+    // We do this by simply setting the list to empty using
+    // compare and swap.   We then have the entire list in our hands,
+    // which cannot be changing since we grabbed it while no one
+    // held the lock.
+    if ((v & kCvSpin) == 0 &&
+        cv_.compare_exchange_strong(v, v & kCvEvent, std::memory_order_acquire,
+                                    std::memory_order_relaxed)) {
+      PerThreadSynch *h = reinterpret_cast<PerThreadSynch *>(v & ~kCvLow);
+      if (h != nullptr) {
+        PerThreadSynch *w;
+        PerThreadSynch *n = h->next;
+        do {                          // for every thread, wake it up
+          w = n;
+          n = n->next;
+          CondVar::Wakeup(w);
+        } while (w != h);
+        cond_var_tracer("SignalAll wakeup", this);
+      }
+      if ((v & kCvEvent) != 0) {
+        PostSynchEvent(this, SYNCH_EV_SIGNALALL);
+      }
+      ABSL_TSAN_MUTEX_POST_SIGNAL(0, 0);
+      return;
+    } else {
+      c = Delay(c, GENTLE);           // try again after a delay
+    }
+  }
+  ABSL_TSAN_MUTEX_POST_SIGNAL(0, 0);
+}
+
+void ReleasableMutexLock::Release() {
+  ABSL_RAW_CHECK(this->mu_ != nullptr,
+                 "ReleasableMutexLock::Release may only be called once");
+  this->mu_->Unlock();
+  this->mu_ = nullptr;
+}
+
+#ifdef THREAD_SANITIZER
+extern "C" void __tsan_read1(void *addr);
+#else
+#define __tsan_read1(addr)  // do nothing if TSan not enabled
+#endif
+
+// A function that just returns its argument, dereferenced
+static bool Dereference(void *arg) {
+  // ThreadSanitizer does not instrument this file for memory accesses.
+  // This function dereferences a user variable that can participate
+  // in a data race, so we need to manually tell TSan about this memory access.
+  __tsan_read1(arg);
+  return *(static_cast<bool *>(arg));
+}
+
+Condition::Condition() {}   // null constructor, used for kTrue only
+const Condition Condition::kTrue;
+
+Condition::Condition(bool (*func)(void *), void *arg)
+    : eval_(&CallVoidPtrFunction),
+      function_(func),
+      method_(nullptr),
+      arg_(arg) {}
+
+bool Condition::CallVoidPtrFunction(const Condition *c) {
+  return (*c->function_)(c->arg_);
+}
+
+Condition::Condition(const bool *cond)
+    : eval_(CallVoidPtrFunction),
+      function_(Dereference),
+      method_(nullptr),
+      // const_cast is safe since Dereference does not modify arg
+      arg_(const_cast<bool *>(cond)) {}
+
+bool Condition::Eval() const {
+  // eval_ == null for kTrue
+  return (this->eval_ == nullptr) || (*this->eval_)(this);
+}
+
+bool Condition::GuaranteedEqual(const Condition *a, const Condition *b) {
+  if (a == nullptr) {
+    return b == nullptr || b->eval_ == nullptr;
+  }
+  if (b == nullptr || b->eval_ == nullptr) {
+    return a->eval_ == nullptr;
+  }
+  return a->eval_ == b->eval_ && a->function_ == b->function_ &&
+         a->arg_ == b->arg_ && a->method_ == b->method_;
+}
+
+}  // namespace absl
diff --git a/absl/synchronization/mutex.h b/absl/synchronization/mutex.h
new file mode 100644
index 000000000000..a417802677cc
--- /dev/null
+++ b/absl/synchronization/mutex.h
@@ -0,0 +1,1013 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// -----------------------------------------------------------------------------
+// mutex.h
+// -----------------------------------------------------------------------------
+//
+// This header file defines a `Mutex` -- a mutually exclusive lock -- and the
+// most common type of synchronization primitive for facilitating locks on
+// shared resources. A mutex is used to prevent multiple threads from accessing
+// and/or writing to a shared resource concurrently.
+//
+// Unlike a `std::mutex`, the Abseil `Mutex` provides the following additional
+// features:
+//   * Conditional predicates intrinsic to the `Mutex` object
+//   * Reader/writer locks, in addition to standard exclusive/writer locks
+//   * Deadlock detection and debug support.
+//
+// The following helper classes are also defined within this file:
+//
+//  MutexLock - An RAII wrapper to acquire and release a `Mutex` for exclusive/
+//              write access within the current scope.
+//  ReaderMutexLock
+//            - An RAII wrapper to acquire and release a `Mutex` for shared/read
+//              access within the current scope.
+//
+//  WriterMutexLock
+//            - Alias for `MutexLock` above, designed for use in distinguishing
+//              reader and writer locks within code.
+//
+// In addition to simple mutex locks, this file also defines ways to perform
+// locking under certain conditions.
+//
+//  Condition   - (Preferred) Used to wait for a particular predicate that
+//                depends on state protected by the `Mutex` to become true.
+//  CondVar     - A lower-level variant of `Condition` that relies on
+//                application code to explicitly signal the `CondVar` when
+//                a condition has been met.
+//
+// See below for more information on using `Condition` or `CondVar`.
+//
+// Mutexes and mutex behavior can be quite complicated. The information within
+// this header file is limited, as a result. Please consult the Mutex guide for
+// more complete information and examples.
+
+#ifndef ABSL_SYNCHRONIZATION_MUTEX_H_
+#define ABSL_SYNCHRONIZATION_MUTEX_H_
+
+#include <atomic>
+#include <cstdint>
+#include <string>
+
+#include "absl/base/internal/identity.h"
+#include "absl/base/internal/low_level_alloc.h"
+#include "absl/base/internal/thread_identity.h"
+#include "absl/base/port.h"
+#include "absl/base/thread_annotations.h"
+#include "absl/synchronization/internal/kernel_timeout.h"
+#include "absl/synchronization/internal/per_thread_sem.h"
+#include "absl/time/time.h"
+
+// Decide if we should use the non-production implementation because
+// the production implementation hasn't been fully ported yet.
+#ifdef ABSL_INTERNAL_USE_NONPROD_MUTEX
+#error ABSL_INTERNAL_USE_NONPROD_MUTEX cannot be directly set
+#elif defined(ABSL_LOW_LEVEL_ALLOC_MISSING)
+#define ABSL_INTERNAL_USE_NONPROD_MUTEX 1
+#include "absl/synchronization/internal/mutex_nonprod.inc"
+#endif
+
+namespace absl {
+
+struct SynchWaitParams;
+class Condition;
+
+// -----------------------------------------------------------------------------
+// Mutex
+// -----------------------------------------------------------------------------
+//
+// A `Mutex` is a non-reentrant (aka non-recursive) Mutually Exclusive lock
+// on some resource, typically a variable or data structure with associated
+// invariants. Proper usage of mutexes prevents concurrent access by different
+// threads to the same resource.
+//
+// A `Mutex` has two basic operations: `Mutex::Lock()` and `Mutex::Unlock()`.
+// The `Lock()` operation *acquires* a `Mutex` (in a state known as an
+// *exclusive* -- or write -- lock), while the `Unlock()` operation *releases* a
+// Mutex. During the span of time between the Lock() and Unlock() operations,
+// a mutex is said to be *held*. By design all mutexes support exclusive/write
+// locks, as this is the most common way to use a mutex.
+//
+// The `Mutex` state machine for basic lock/unlock operations is quite simple:
+//
+// |                | Lock()     | Unlock() |
+// |----------------+------------+----------|
+// | Free           | Exclusive  | invalid  |
+// | Exclusive      | blocks     | Free     |
+//
+// Attempts to `Unlock()` must originate from the thread that performed the
+// corresponding `Lock()` operation.
+//
+// An "invalid" operation is disallowed by the API. The `Mutex` implementation
+// is allowed to do anything on an invalid call, including but not limited to
+// crashing with a useful error message, silently succeeding, or corrupting
+// data structures. In debug mode, the implementation attempts to crash with a
+// useful error message.
+//
+// `Mutex` is not guaranteed to be "fair" in prioritizing waiting threads; it
+// is, however, approximately fair over long periods, and starvation-free for
+// threads at the same priority.
+//
+// The lock/unlock primitives are now annotated with lock annotations
+// defined in (base/thread_annotations.h). When writing multi-threaded code,
+// you should use lock annotations whenever possible to document your lock
+// synchronization policy. Besides acting as documentation, these annotations
+// also help compilers or static analysis tools to identify and warn about
+// issues that could potentially result in race conditions and deadlocks.
+//
+// For more information about the lock annotations, please see
+// [Thread Safety Analysis](http://clang.llvm.org/docs/ThreadSafetyAnalysis.html)
+// in the Clang documentation.
+//
+// See also `MutexLock`, below, for scoped `Mutex` acquisition.
+
+class LOCKABLE Mutex {
+ public:
+  Mutex();
+  ~Mutex();
+
+  // Mutex::Lock()
+  //
+  // Blocks the calling thread, if necessary, until this `Mutex` is free, and
+  // then acquires it exclusively. (This lock is also known as a "write lock.")
+  void Lock() EXCLUSIVE_LOCK_FUNCTION();
+
+  // Mutex::Unlock()
+  //
+  // Releases this `Mutex` and returns it from the exclusive/write state to the
+  // free state. Caller must hold the `Mutex` exclusively.
+  void Unlock() UNLOCK_FUNCTION();
+
+  // Mutex::TryLock()
+  //
+  // If the mutex can be acquired without blocking, does so exclusively and
+  // returns `true`. Otherwise, returns `false`. Returns `true` with high
+  // probability if the `Mutex` was free.
+  bool TryLock() EXCLUSIVE_TRYLOCK_FUNCTION(true);
+
+  // Mutex::AssertHeld()
+  //
+  // Return immediately if this thread holds the `Mutex` exclusively (in write
+  // mode). Otherwise, may report an error (typically by crashing with a
+  // diagnostic), or may return immediately.
+  void AssertHeld() const ASSERT_EXCLUSIVE_LOCK();
+
+  // ---------------------------------------------------------------------------
+  // Reader-Writer Locking
+  // ---------------------------------------------------------------------------
+
+  // A Mutex can also be used as a starvation-free reader-writer lock.
+  // Neither read-locks nor write-locks are reentrant/recursive to avoid
+  // potential client programming errors.
+  //
+  // The Mutex API provides `Writer*()` aliases for the existing `Lock()`,
+  // `Unlock()` and `TryLock()` methods for use within applications mixing
+  // reader/writer locks. Using `Reader*()` and `Writer*()` operations in this
+  // manner can make locking behavior clearer when mixing read and write modes.
+  //
+  // Introducing reader locks necessarily complicates the `Mutex` state
+  // machine somewhat. The table below illustrates the allowed state transitions
+  // of a mutex in such cases. Note that ReaderLock() may block even if the lock
+  // is held in shared mode; this occurs when another thread is blocked on a
+  // call to WriterLock().
+  //
+  // ---------------------------------------------------------------------------
+  //     Operation: WriterLock() Unlock()  ReaderLock()           ReaderUnlock()
+  // ---------------------------------------------------------------------------
+  // State
+  // ---------------------------------------------------------------------------
+  // Free           Exclusive    invalid   Shared(1)              invalid
+  // Shared(1)      blocks       invalid   Shared(2) or blocks    Free
+  // Shared(n) n>1  blocks       invalid   Shared(n+1) or blocks  Shared(n-1)
+  // Exclusive      blocks       Free      blocks                 invalid
+  // ---------------------------------------------------------------------------
+  //
+  // In comments below, "shared" refers to a state of Shared(n) for any n > 0.
+
+  // Mutex::ReaderLock()
+  //
+  // Blocks the calling thread, if necessary, until this `Mutex` is either free,
+  // or in shared mode, and then acquires a share of it. Note that
+  // `ReaderLock()` will block if some other thread has an exclusive/writer lock
+  // on the mutex.
+
+  void ReaderLock() SHARED_LOCK_FUNCTION();
+
+  // Mutex::ReaderUnlock()
+  //
+  // Releases a read share of this `Mutex`. `ReaderUnlock` may return a mutex to
+  // the free state if this thread holds the last reader lock on the mutex. Note
+  // that you cannot call `ReaderUnlock()` on a mutex held in write mode.
+  void ReaderUnlock() UNLOCK_FUNCTION();
+
+  // Mutex::ReaderTryLock()
+  //
+  // If the mutex can be acquired without blocking, acquires this mutex for
+  // shared access and returns `true`. Otherwise, returns `false`. Returns
+  // `true` with high probability if the `Mutex` was free or shared.
+  bool ReaderTryLock() SHARED_TRYLOCK_FUNCTION(true);
+
+  // Mutex::AssertReaderHeld()
+  //
+  // Returns immediately if this thread holds the `Mutex` in at least shared
+  // mode (read mode). Otherwise, may report an error (typically by
+  // crashing with a diagnostic), or may return immediately.
+  void AssertReaderHeld() const ASSERT_SHARED_LOCK();
+
+  // Mutex::WriterLock()
+  // Mutex::WriterUnlock()
+  // Mutex::WriterTryLock()
+  //
+  // Aliases for `Mutex::Lock()`, `Mutex::Unlock()`, and `Mutex::TryLock()`.
+  //
+  // Use the `Writer*()` versions of these method names when using complementary
+  // `Reader*()` methods to distingish simple exclusive `Mutex` usage (`Lock()`,
+  // etc.) from reader/writer lock usage.
+  void WriterLock() EXCLUSIVE_LOCK_FUNCTION() { this->Lock(); }
+
+  void WriterUnlock() UNLOCK_FUNCTION() { this->Unlock(); }
+
+  bool WriterTryLock() EXCLUSIVE_TRYLOCK_FUNCTION(true) {
+    return this->TryLock();
+  }
+
+  // ---------------------------------------------------------------------------
+  // Conditional Critical Regions
+  // ---------------------------------------------------------------------------
+
+  // Conditional usage of a `Mutex` can occur using two distinct paradigms:
+  //
+  //   * Use of `Mutex` member functions with `Condition` objects.
+  //   * Use of the separate `CondVar` abstraction.
+  //
+  // In general, prefer use of `Condition` and the `Mutex` member functions
+  // listed below over `CondVar`. When there are multiple threads waiting on
+  // distinctly different conditions, however, a battery of `CondVar`s may be
+  // more efficient. This section discusses use of `Condition` objects.
+  //
+  // `Mutex` contains member functions for performing lock operations only under
+  // certain conditions, of class `Condition`. For correctness, the `Condition`
+  // must return a boolean that is a pure function, only of state protected by
+  // the `Mutex`. The condition must be invariant w.r.t. environmental state
+  // such as thread, cpu id, or time, and must be `noexcept`. The condition will
+  // always be invoked with the mutex held in at least read mode, so you should
+  // not block it for long periods or sleep it on a timer.
+  //
+  // Since a condition must not depend directly on the current time, use
+  // `*WithTimeout()` member function variants to make your condition
+  // effectively true after a given duration, or `*WithDeadline()` variants to
+  // make your condition effectively true after a given time.
+  //
+  // The condition function should have no side-effects aside from debug
+  // logging; as a special exception, the function may acquire other mutexes
+  // provided it releases all those that it acquires.  (This exception was
+  // required to allow logging.)
+
+  // Mutex::Await()
+  //
+  // Unlocks this `Mutex` and blocks until simultaneously both `cond` is `true`
+  // and this `Mutex` can be reacquired, then reacquires this `Mutex` in the
+  // same mode in which it was previously held. If the condition is initially
+  // `true`, `Await()` *may* skip the release/re-acquire step.
+  //
+  // `Await()` requires that this thread holds this `Mutex` in some mode.
+  void Await(const Condition &cond);
+
+  // Mutex::LockWhen()
+  // Mutex::ReaderLockWhen()
+  // Mutex::WriterLockWhen()
+  //
+  // Blocks until simultaneously both `cond` is `true` and this` Mutex` can
+  // be acquired, then atomically acquires this `Mutex`. `LockWhen()` is
+  // logically equivalent to `*Lock(); Await();` though they may have different
+  // performance characteristics.
+  void LockWhen(const Condition &cond) EXCLUSIVE_LOCK_FUNCTION();
+
+  void ReaderLockWhen(const Condition &cond) SHARED_LOCK_FUNCTION();
+
+  void WriterLockWhen(const Condition &cond) EXCLUSIVE_LOCK_FUNCTION() {
+    this->LockWhen(cond);
+  }
+
+  // ---------------------------------------------------------------------------
+  // Mutex Variants with Timeouts/Deadlines
+  // ---------------------------------------------------------------------------
+
+  // Mutex::AwaitWithTimeout()
+  // Mutex::AwaitWithDeadline()
+  //
+  // If `cond` is initially true, do nothing, or act as though `cond` is
+  // initially false.
+  //
+  // If `cond` is initially false, unlock this `Mutex` and block until
+  // simultaneously:
+  //   - either `cond` is true or the {timeout has expired, deadline has passed}
+  //     and
+  //   - this `Mutex` can be reacquired,
+  // then reacquire this `Mutex` in the same mode in which it was previously
+  // held, returning `true` iff `cond` is `true` on return.
+  //
+  // Deadlines in the past are equivalent to an immediate deadline.
+  // Negative timeouts are equivalent to a zero timeout.
+  //
+  // This method requires that this thread holds this `Mutex` in some mode.
+  bool AwaitWithTimeout(const Condition &cond, absl::Duration timeout);
+
+  bool AwaitWithDeadline(const Condition &cond, absl::Time deadline);
+
+  // Mutex::LockWhenWithTimeout()
+  // Mutex::ReaderLockWhenWithTimeout()
+  // Mutex::WriterLockWhenWithTimeout()
+  //
+  // Blocks until simultaneously both:
+  //   - either `cond` is `true` or the timeout has expired, and
+  //   - this `Mutex` can be acquired,
+  // then atomically acquires this `Mutex`, returning `true` iff `cond` is
+  // `true` on return.
+  //
+  // Negative timeouts are equivalent to a zero timeout.
+  bool LockWhenWithTimeout(const Condition &cond, absl::Duration timeout)
+      EXCLUSIVE_LOCK_FUNCTION();
+  bool ReaderLockWhenWithTimeout(const Condition &cond, absl::Duration timeout)
+      SHARED_LOCK_FUNCTION();
+  bool WriterLockWhenWithTimeout(const Condition &cond, absl::Duration timeout)
+      EXCLUSIVE_LOCK_FUNCTION() {
+    return this->LockWhenWithTimeout(cond, timeout);
+  }
+
+  // Mutex::LockWhenWithDeadline()
+  // Mutex::ReaderLockWhenWithDeadline()
+  // Mutex::WriterLockWhenWithDeadline()
+  //
+  // Blocks until simultaneously both:
+  //   - either `cond` is `true` or the deadline has been passed, and
+  //   - this `Mutex` can be acquired,
+  // then atomically acquires this Mutex, returning `true` iff `cond` is `true`
+  // on return.
+  //
+  // Deadlines in the past are equivalent to an immediate deadline.
+  bool LockWhenWithDeadline(const Condition &cond, absl::Time deadline)
+      EXCLUSIVE_LOCK_FUNCTION();
+  bool ReaderLockWhenWithDeadline(const Condition &cond, absl::Time deadline)
+      SHARED_LOCK_FUNCTION();
+  bool WriterLockWhenWithDeadline(const Condition &cond, absl::Time deadline)
+      EXCLUSIVE_LOCK_FUNCTION() {
+    return this->LockWhenWithDeadline(cond, deadline);
+  }
+
+  // ---------------------------------------------------------------------------
+  // Debug Support: Invariant Checking, Deadlock Detection, Logging.
+  // ---------------------------------------------------------------------------
+
+  // Mutex::EnableInvariantDebugging()
+  //
+  // If `invariant`!=null and if invariant debugging has been enabled globally,
+  // cause `(*invariant)(arg)` to be called at moments when the invariant for
+  // this `Mutex` should hold (for example: just after acquire, just before
+  // release).
+  //
+  // The routine `invariant` should have no side-effects since it is not
+  // guaranteed how many times it will be called; it should check the invariant
+  // and crash if it does not hold. Enabling global invariant debugging may
+  // substantially reduce `Mutex` performance; it should be set only for
+  // non-production runs.  Optimization options may also disable invariant
+  // checks.
+  void EnableInvariantDebugging(void (*invariant)(void *), void *arg);
+
+  // Mutex::EnableDebugLog()
+  //
+  // Cause all subsequent uses of this `Mutex` to be logged via
+  // `ABSL_RAW_LOG(INFO)`. Log entries are tagged with `name` if no previous
+  // call to `EnableInvariantDebugging()` or `EnableDebugLog()` has been made.
+  //
+  // Note: This method substantially reduces `Mutex` performance.
+  void EnableDebugLog(const char *name);
+
+  // Deadlock detection
+
+  // Mutex::ForgetDeadlockInfo()
+  //
+  // Forget any deadlock-detection information previously gathered
+  // about this `Mutex`. Call this method in debug mode when the lock ordering
+  // of a `Mutex` changes.
+  void ForgetDeadlockInfo();
+
+  // Mutex::AssertNotHeld()
+  //
+  // Return immediately if this thread does not hold this `Mutex` in any
+  // mode; otherwise, may report an error (typically by crashing with a
+  // diagnostic), or may return immediately.
+  //
+  // Currently this check is performed only if all of:
+  //    - in debug mode
+  //    - SetMutexDeadlockDetectionMode() has been set to kReport or kAbort
+  //    - number of locks concurrently held by this thread is not large.
+  // are true.
+  void AssertNotHeld() const;
+
+  // Special cases.
+
+  // A `MuHow` is a constant that indicates how a lock should be acquired.
+  // Internal implementation detail.  Clients should ignore.
+  typedef const struct MuHowS *MuHow;
+
+  // Mutex::InternalAttemptToUseMutexInFatalSignalHandler()
+  //
+  // Causes the `Mutex` implementation to prepare itself for re-entry caused by
+  // future use of `Mutex` within a fatal signal handler. This method is
+  // intended for use only for last-ditch attempts to log crash information.
+  // It does not guarantee that attempts to use Mutexes within the handler will
+  // not deadlock; it merely makes other faults less likely.
+  //
+  // WARNING:  This routine must be invoked from a signal handler, and the
+  // signal handler must either loop forever or terminate the process.
+  // Attempts to return from (or `longjmp` out of) the signal handler once this
+  // call has been made may cause arbitrary program behaviour including
+  // crashes and deadlocks.
+  static void InternalAttemptToUseMutexInFatalSignalHandler();
+
+ private:
+#ifdef ABSL_INTERNAL_USE_NONPROD_MUTEX
+  friend class CondVar;
+
+  synchronization_internal::MutexImpl *impl() { return impl_.get(); }
+
+  synchronization_internal::SynchronizationStorage<
+      synchronization_internal::MutexImpl>
+      impl_;
+#else
+  std::atomic<intptr_t> mu_;  // The Mutex state.
+
+  // Post()/Wait() versus associated PerThreadSem; in class for required
+  // friendship with PerThreadSem.
+  static inline void IncrementSynchSem(Mutex *mu,
+                                       base_internal::PerThreadSynch *w);
+  static inline bool DecrementSynchSem(
+      Mutex *mu, base_internal::PerThreadSynch *w,
+      synchronization_internal::KernelTimeout t);
+
+  // slow path acquire
+  void LockSlowLoop(SynchWaitParams *waitp, int flags);
+  // wrappers around LockSlowLoop()
+  bool LockSlowWithDeadline(MuHow how, const Condition *cond,
+                            synchronization_internal::KernelTimeout t,
+                            int flags);
+  void LockSlow(MuHow how, const Condition *cond,
+                int flags) ABSL_ATTRIBUTE_COLD;
+  // slow path release
+  void UnlockSlow(SynchWaitParams *waitp) ABSL_ATTRIBUTE_COLD;
+  // Common code between Await() and AwaitWithTimeout/Deadline()
+  bool AwaitCommon(const Condition &cond,
+                   synchronization_internal::KernelTimeout t);
+  // Attempt to remove thread s from queue.
+  void TryRemove(base_internal::PerThreadSynch *s);
+  // Block a thread on mutex.
+  void Block(base_internal::PerThreadSynch *s);
+  // Wake a thread; return successor.
+  base_internal::PerThreadSynch *Wakeup(base_internal::PerThreadSynch *w);
+
+  friend class CondVar;   // for access to Trans()/Fer().
+  void Trans(MuHow how);  // used for CondVar->Mutex transfer
+  void Fer(
+      base_internal::PerThreadSynch *w);  // used for CondVar->Mutex transfer
+#endif
+
+  // Catch the error of writing Mutex when intending MutexLock.
+  Mutex(const volatile Mutex * /*ignored*/) {}  // NOLINT(runtime/explicit)
+
+  Mutex(const Mutex&) = delete;
+  Mutex& operator=(const Mutex&) = delete;
+};
+
+// -----------------------------------------------------------------------------
+// Mutex RAII Wrappers
+// -----------------------------------------------------------------------------
+
+// MutexLock
+//
+// `MutexLock` is a helper class, which acquires and releases a `Mutex` via
+// RAII.
+//
+// Example:
+//
+// Class Foo {
+//
+//   Foo::Bar* Baz() {
+//     MutexLock l(&lock_);
+//     ...
+//     return bar;
+//   }
+//
+// private:
+//   Mutex lock_;
+// };
+class SCOPED_LOCKABLE MutexLock {
+ public:
+  explicit MutexLock(Mutex *mu) EXCLUSIVE_LOCK_FUNCTION(mu) : mu_(mu) {
+    this->mu_->Lock();
+  }
+  ~MutexLock() UNLOCK_FUNCTION() { this->mu_->Unlock(); }
+ private:
+  Mutex *const mu_;
+  MutexLock(const MutexLock &) = delete;  // NOLINT(runtime/mutex)
+  MutexLock& operator=(const MutexLock&) = delete;
+};
+
+// ReaderMutexLock
+//
+// The `ReaderMutexLock` is a helper class, like `MutexLock`, which acquires and
+// releases a shared lock on a `Mutex` via RAII.
+class SCOPED_LOCKABLE ReaderMutexLock {
+ public:
+  explicit ReaderMutexLock(Mutex *mu) SHARED_LOCK_FUNCTION(mu)
+      :  mu_(mu) {
+    mu->ReaderLock();
+  }
+  ~ReaderMutexLock() UNLOCK_FUNCTION() {
+    this->mu_->ReaderUnlock();
+  }
+ private:
+  Mutex *const mu_;
+  ReaderMutexLock(const ReaderMutexLock&) = delete;
+  ReaderMutexLock& operator=(const ReaderMutexLock&) = delete;
+};
+
+// WriterMutexLock
+//
+// The `WriterMutexLock` is a helper class, like `MutexLock`, which acquires and
+// releases a write (exclusive) lock on a `Mutex` va RAII.
+class SCOPED_LOCKABLE WriterMutexLock {
+ public:
+  explicit WriterMutexLock(Mutex *mu) EXCLUSIVE_LOCK_FUNCTION(mu)
+      : mu_(mu) {
+    mu->WriterLock();
+  }
+  ~WriterMutexLock() UNLOCK_FUNCTION() {
+    this->mu_->WriterUnlock();
+  }
+ private:
+  Mutex *const mu_;
+  WriterMutexLock(const WriterMutexLock&) = delete;
+  WriterMutexLock& operator=(const WriterMutexLock&) = delete;
+};
+
+// -----------------------------------------------------------------------------
+// Condition
+// -----------------------------------------------------------------------------
+//
+// As noted above, `Mutex` contains a number of member functions which take a
+// `Condition` as a argument; clients can wait for conditions to become `true`
+// before attempting to acquire the mutex. These sections are known as
+// "condition critical" sections. To use a `Condition`, you simply need to
+// construct it, and use within an appropriate `Mutex` member function;
+// everything else in the `Condition` class is an implementation detail.
+//
+// A `Condition` is specified as a function pointer which returns a boolean.
+// `Condition` functions should be pure functions -- their results should depend
+// only on passed arguments, should not consult any external state (such as
+// clocks), and should have no side-effects, aside from debug logging. Any
+// objects that the function may access should be limited to those which are
+// constant while the mutex is blocked on the condition (e.g. a stack variable),
+// or objects of state protected explicitly by the mutex.
+//
+// No matter which construction is used for `Condition`, the underlying
+// function pointer / functor / callable must not throw any
+// exceptions. Correctness of `Mutex` / `Condition` is not guaranteed in
+// the face of a throwing `Condition`. (When Abseil is allowed to depend
+// on C++17, these function pointers will be explicitly marked
+// `noexcept`; until then this requirement cannot be enforced in the
+// type system.)
+//
+// Note: to use a `Condition`, you need only construct it and pass it within the
+// appropriate `Mutex' member function, such as `Mutex::Await()`.
+//
+// Example:
+//
+//   // assume count_ is not internal reference count
+//   int count_ GUARDED_BY(mu_);
+//
+//   mu_.LockWhen(Condition(+[](const int* count) { return *count == 0; },
+//         &count_));
+//
+// When multiple threads are waiting on exactly the same condition, make sure
+// that they are constructed with the same parameters (same pointer to function
+// + arg, or same pointer to object + method), so that the mutex implementation
+// can avoid redundantly evaluating the same condition for each thread.
+class Condition {
+ public:
+  // A Condition that returns the result of "(*func)(arg)"
+  Condition(bool (*func)(void *), void *arg);
+
+  // Templated version for people who are averse to casts.
+  //
+  // To use a lambda, prepend it with unary plus, which converts the lambda
+  // into a function pointer:
+  //     Condition(+[](T* t) { return ...; }, arg).
+  //
+  // Note: lambdas in this case must contain no bound variables.
+  //
+  // See class comment for performance advice.
+  template<typename T>
+  Condition(bool (*func)(T *), T *arg);
+
+  // Templated version for invoking a method that returns a `bool`.
+  //
+  // `Condition(object, &Class::Method)` constructs a `Condition` that evaluates
+  // `object->Method()`.
+  //
+  // Implementation Note: `absl::internal::identity` is used to allow methods to
+  // come from base classes. A simpler signature like
+  // `Condition(T*, bool (T::*)())` does not suffice.
+  template<typename T>
+  Condition(T *object, bool (absl::internal::identity<T>::type::* method)());
+
+  // Same as above, for const members
+  template<typename T>
+  Condition(const T *object,
+            bool (absl::internal::identity<T>::type::* method)() const);
+
+  // A Condition that returns the value of `*cond`
+  explicit Condition(const bool *cond);
+
+  // Templated version for invoking a functor that returns a `bool`.
+  // This approach accepts pointers to non-mutable lambdas, `std::function`,
+  // the result of` std::bind` and user-defined functors that define
+  // `bool F::operator()() const`.
+  //
+  // Example:
+  //
+  //   auto reached = [this, current]() {
+  //     mu_.AssertReaderHeld();                // For annotalysis.
+  //     return processed_ >= current;
+  //   };
+  //   mu_.Await(Condition(&reached));
+
+  // See class comment for performance advice. In particular, if there
+  // might be more than one waiter for the same condition, make sure
+  // that all waiters construct the condition with the same pointers.
+
+  // Implementation note: The second template parameter ensures that this
+  // constructor doesn't participate in overload resolution if T doesn't have
+  // `bool operator() const`.
+  template <typename T, typename E = decltype(
+      static_cast<bool (T::*)() const>(&T::operator()))>
+  explicit Condition(const T *obj)
+      : Condition(obj, static_cast<bool (T::*)() const>(&T::operator())) {}
+
+  // A Condition that always returns `true`.
+  static const Condition kTrue;
+
+  // Evaluates the condition.
+  bool Eval() const;
+
+  // Returns `true` if the two conditions are guaranteed to return the same
+  // value if evaluated at the same time, `false` if the evaluation *may* return
+  // different results.
+  //
+  // Two `Condition` values are guaranteed equal if both their `func` and `arg`
+  // components are the same. A null pointer is equivalent to a `true`
+  // condition.
+  static bool GuaranteedEqual(const Condition *a, const Condition *b);
+
+ private:
+  typedef bool (*InternalFunctionType)(void * arg);
+  typedef bool (Condition::*InternalMethodType)();
+  typedef bool (*InternalMethodCallerType)(void * arg,
+                                           InternalMethodType internal_method);
+
+  bool (*eval_)(const Condition*);  // Actual evaluator
+  InternalFunctionType function_;   // function taking pointer returning bool
+  InternalMethodType method_;       // method returning bool
+  void *arg_;                       // arg of function_ or object of method_
+
+  Condition();        // null constructor used only to create kTrue
+
+  // Various functions eval_ can point to:
+  static bool CallVoidPtrFunction(const Condition*);
+  template <typename T> static bool CastAndCallFunction(const Condition* c);
+  template <typename T> static bool CastAndCallMethod(const Condition* c);
+};
+
+// -----------------------------------------------------------------------------
+// CondVar
+// -----------------------------------------------------------------------------
+//
+// A condition variable, reflecting state evaluated separately outside of the
+// `Mutex` object, which can be signaled to wake callers.
+// This class is not normally needed; use `Mutex` member functions such as
+// `Mutex::Await()` and intrinsic `Condition` abstractions. In rare cases
+// with many threads and many conditions, `CondVar` may be faster.
+//
+// The implementation may deliver signals to any condition variable at
+// any time, even when no call to `Signal()` or `SignalAll()` is made; as a
+// result, upon being awoken, you must check the logical condition you have
+// been waiting upon. The implementation wakes waiters in the FIFO order.
+//
+// Examples:
+//
+// Usage for a thread waiting for some condition C protected by mutex mu:
+//       mu.Lock();
+//       while (!C) { cv->Wait(&mu); }        // releases and reacquires mu
+//       //  C holds; process data
+//       mu.Unlock();
+//
+// Usage to wake T is:
+//       mu.Lock();
+//      // process data, possibly establishing C
+//      if (C) { cv->Signal(); }
+//      mu.Unlock();
+//
+// If C may be useful to more than one waiter, use `SignalAll()` instead of
+// `Signal()`.
+//
+// With this implementation it is efficient to use `Signal()/SignalAll()` inside
+// the locked region; this usage can make reasoning about your program easier.
+//
+class CondVar {
+ public:
+  CondVar();
+  ~CondVar();
+
+  // CondVar::Wait()
+  //
+  // Atomically releases a `Mutex` and blocks on this condition variable. After
+  // blocking, the thread will unblock, reacquire the `Mutex`, and return if
+  // either:
+  //  - this condition variable is signalled with `SignalAll()`, or
+  //  - this condition variable is signalled in any manner and this thread
+  //    was the most recently blocked thread that has not yet woken.
+  // Requires and ensures that the current thread holds the `Mutex`.
+  void Wait(Mutex *mu);
+
+  // CondVar::WaitWithTimeout()
+  //
+  // Atomically releases a `Mutex`, blocks on this condition variable, and
+  // attempts to reacquire the mutex upon being signalled, or upon reaching the
+  // timeout.
+  //
+  // After blocking, the thread will unblock, reacquire the `Mutex`, and return
+  // for any of the following:
+  //  - this condition variable is signalled with `SignalAll()`
+  //  - the timeout has expired
+  //  - this condition variable is signalled in any manner and this thread
+  //    was the most recently blocked thread that has not yet woken.
+  //
+  // Negative timeouts are equivalent to a zero timeout.
+  //
+  // Returns true if the timeout has expired without this `CondVar`
+  // being signalled in any manner. If both the timeout has expired
+  // and this `CondVar` has been signalled, the implementation is free
+  // to return `true` or `false`.
+  //
+  // Requires and ensures that the current thread holds the `Mutex`.
+  bool WaitWithTimeout(Mutex *mu, absl::Duration timeout);
+
+  // CondVar::WaitWithDeadline()
+  //
+  // Atomically releases a `Mutex`, blocks on this condition variable, and
+  // attempts to reacquire the mutex within the provided deadline.
+  //
+  // After blocking, the thread will unblock, reacquire the `Mutex`, and return
+  // for any of the following:
+  //  - this condition variable is signalled with `SignalAll()`
+  //  - the deadline has passed
+  //  - this condition variable is signalled in any manner and this thread
+  //    was the most recently blocked thread that has not yet woken.
+  //
+  // Deadlines in the past are equivalent to an immediate deadline.
+  //
+  // Returns true if the deadline has passed without this `CondVar`
+  // being signalled in any manner. If both the deadline has passed
+  // and this `CondVar` has been signalled, the implementation is free
+  // to return `true` or `false`.
+  //
+  // Requires and ensures that the current thread holds the `Mutex`.
+  bool WaitWithDeadline(Mutex *mu, absl::Time deadline);
+
+  // CondVar::Signal()
+  //
+  // Signal this `CondVar`; wake at least one waiter if one exists.
+  void Signal();
+
+  // CondVar::SignalAll()
+  //
+  // Signal this `CondVar`; wake all waiters.
+  void SignalAll();
+
+  // CondVar::EnableDebugLog()
+  //
+  // Causes all subsequent uses of this `CondVar` to be logged via
+  // `ABSL_RAW_LOG(INFO)`. Log entries are tagged with `name` if `name != 0`.
+  // Note: this method substantially reduces `CondVar` performance.
+  void EnableDebugLog(const char *name);
+
+ private:
+#ifdef ABSL_INTERNAL_USE_NONPROD_MUTEX
+  synchronization_internal::CondVarImpl *impl() { return impl_.get(); }
+  synchronization_internal::SynchronizationStorage<
+      synchronization_internal::CondVarImpl>
+      impl_;
+#else
+  bool WaitCommon(Mutex *mutex, synchronization_internal::KernelTimeout t);
+  void Remove(base_internal::PerThreadSynch *s);
+  void Wakeup(base_internal::PerThreadSynch *w);
+  std::atomic<intptr_t> cv_;  // Condition variable state.
+#endif
+  CondVar(const CondVar&) = delete;
+  CondVar& operator=(const CondVar&) = delete;
+};
+
+
+// Variants of MutexLock.
+//
+// If you find yourself using one of these, consider instead using
+// Mutex::Unlock() and/or if-statements for clarity.
+
+// MutexLockMaybe
+//
+// MutexLockMaybe is like MutexLock, but is a no-op when mu is null.
+class SCOPED_LOCKABLE MutexLockMaybe {
+ public:
+  explicit MutexLockMaybe(Mutex *mu) EXCLUSIVE_LOCK_FUNCTION(mu)
+      : mu_(mu) { if (this->mu_ != nullptr) { this->mu_->Lock(); } }
+  ~MutexLockMaybe() UNLOCK_FUNCTION() {
+    if (this->mu_ != nullptr) { this->mu_->Unlock(); }
+  }
+ private:
+  Mutex *const mu_;
+  MutexLockMaybe(const MutexLockMaybe&) = delete;
+  MutexLockMaybe& operator=(const MutexLockMaybe&) = delete;
+};
+
+// ReleaseableMutexLock
+//
+// ReleasableMutexLock is like MutexLock, but permits `Release()` of its
+// mutex before destruction. `Release()` may be called at most once.
+class SCOPED_LOCKABLE ReleasableMutexLock {
+ public:
+  explicit ReleasableMutexLock(Mutex *mu) EXCLUSIVE_LOCK_FUNCTION(mu)
+      : mu_(mu) {
+    this->mu_->Lock();
+  }
+  ~ReleasableMutexLock() UNLOCK_FUNCTION() {
+    if (this->mu_ != nullptr) { this->mu_->Unlock(); }
+  }
+
+  void Release() UNLOCK_FUNCTION();
+
+ private:
+  Mutex *mu_;
+  ReleasableMutexLock(const ReleasableMutexLock&) = delete;
+  ReleasableMutexLock& operator=(const ReleasableMutexLock&) = delete;
+};
+
+#ifdef ABSL_INTERNAL_USE_NONPROD_MUTEX
+#else
+inline CondVar::CondVar() : cv_(0) {}
+#endif
+
+// static
+template <typename T>
+bool Condition::CastAndCallMethod(const Condition *c) {
+  typedef bool (T::*MemberType)();
+  MemberType rm = reinterpret_cast<MemberType>(c->method_);
+  T *x = static_cast<T *>(c->arg_);
+  return (x->*rm)();
+}
+
+// static
+template <typename T>
+bool Condition::CastAndCallFunction(const Condition *c) {
+  typedef bool (*FuncType)(T *);
+  FuncType fn = reinterpret_cast<FuncType>(c->function_);
+  T *x = static_cast<T *>(c->arg_);
+  return (*fn)(x);
+}
+
+template <typename T>
+inline Condition::Condition(bool (*func)(T *), T *arg)
+    : eval_(&CastAndCallFunction<T>),
+      function_(reinterpret_cast<InternalFunctionType>(func)),
+      method_(nullptr),
+      arg_(const_cast<void *>(static_cast<const void *>(arg))) {}
+
+template <typename T>
+inline Condition::Condition(T *object,
+                            bool (absl::internal::identity<T>::type::*method)())
+    : eval_(&CastAndCallMethod<T>),
+      function_(nullptr),
+      method_(reinterpret_cast<InternalMethodType>(method)),
+      arg_(object) {}
+
+template <typename T>
+inline Condition::Condition(const T *object,
+                            bool (absl::internal::identity<T>::type::*method)()
+                                const)
+    : eval_(&CastAndCallMethod<T>),
+      function_(nullptr),
+      method_(reinterpret_cast<InternalMethodType>(method)),
+      arg_(reinterpret_cast<void *>(const_cast<T *>(object))) {}
+
+// Register a hook for profiling support.
+//
+// The function pointer registered here will be called whenever a mutex is
+// contended.  The callback is given the absl/base/cycleclock.h timestamp when
+// waiting began.
+//
+// Calls to this function do not race or block, but there is no ordering
+// guaranteed between calls to this function and call to the provided hook.
+// In particular, the previously registered hook may still be called for some
+// time after this function returns.
+void RegisterMutexProfiler(void (*fn)(int64_t wait_timestamp));
+
+// Register a hook for Mutex tracing.
+//
+// The function pointer registered here will be called whenever a mutex is
+// contended.  The callback is given an opaque handle to the contended mutex,
+// an event name, and the number of wait cycles (as measured by
+// //absl/base/internal/cycleclock.h, and which may not be real
+// "cycle" counts.)
+//
+// The only event name currently sent is "slow release".
+//
+// This has the same memory ordering concerns as RegisterMutexProfiler() above.
+void RegisterMutexTracer(void (*fn)(const char *msg, const void *obj,
+                              int64_t wait_cycles));
+
+// TODO(gfalcon): Combine RegisterMutexProfiler() and RegisterMutexTracer()
+// into a single interface, since they are only ever called in pairs.
+
+// Register a hook for CondVar tracing.
+//
+// The function pointer registered here will be called here on various CondVar
+// events.  The callback is given an opaque handle to the CondVar object and
+// a std::string identifying the event.  This is thread-safe, but only a single
+// tracer can be registered.
+//
+// Events that can be sent are "Wait", "Unwait", "Signal wakeup", and
+// "SignalAll wakeup".
+//
+// This has the same memory ordering concerns as RegisterMutexProfiler() above.
+void RegisterCondVarTracer(void (*fn)(const char *msg, const void *cv));
+
+// Register a hook for symbolizing stack traces in deadlock detector reports.
+//
+// 'pc' is the program counter being symbolized, 'out' is the buffer to write
+// into, and 'out_size' is the size of the buffer.  This function can return
+// false if symbolizing failed, or true if a null-terminated symbol was written
+// to 'out.'
+//
+// This has the same memory ordering concerns as RegisterMutexProfiler() above.
+void RegisterSymbolizer(bool (*fn)(const void *pc, char *out, int out_size));
+
+// EnableMutexInvariantDebugging()
+//
+// Enable or disable global support for Mutex invariant debugging.  If enabled,
+// then invariant predicates can be registered per-Mutex for debug checking.
+// See Mutex::EnableInvariantDebugging().
+void EnableMutexInvariantDebugging(bool enabled);
+
+// When in debug mode, and when the feature has been enabled globally, the
+// implementation will keep track of lock ordering and complain (or optionally
+// crash) if a cycle is detected in the acquired-before graph.
+
+// Possible modes of operation for the deadlock detector in debug mode.
+enum class OnDeadlockCycle {
+  kIgnore,  // Neither report on nor attempt to track cycles in lock ordering
+  kReport,  // Report lock cycles to stderr when detected
+  kAbort,  // Report lock cycles to stderr when detected, then abort
+};
+
+// SetMutexDeadlockDetectionMode()
+//
+// Enable or disable global support for detection of potential deadlocks
+// due to Mutex lock ordering inversions.  When set to 'kIgnore', tracking of
+// lock ordering is disabled.  Otherwise, in debug builds, a lock ordering graph
+// will be maintained internally, and detected cycles will be reported in
+// the manner chosen here.
+void SetMutexDeadlockDetectionMode(OnDeadlockCycle mode);
+
+}  // namespace absl
+
+// In some build configurations we pass --detect-odr-violations to the
+// gold linker.  This causes it to flag weak symbol overrides as ODR
+// violations.  Because ODR only applies to C++ and not C,
+// --detect-odr-violations ignores symbols not mangled with C++ names.
+// By changing our extension points to be extern "C", we dodge this
+// check.
+extern "C" {
+void AbslInternalMutexYield();
+}  // extern "C"
+#endif  // ABSL_SYNCHRONIZATION_MUTEX_H_
diff --git a/absl/synchronization/mutex_test.cc b/absl/synchronization/mutex_test.cc
new file mode 100644
index 000000000000..9cf34fce0271
--- /dev/null
+++ b/absl/synchronization/mutex_test.cc
@@ -0,0 +1,1538 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/synchronization/mutex.h"
+
+#ifdef WIN32
+#include <windows.h>
+#endif
+
+#include <algorithm>
+#include <atomic>
+#include <cstdint>
+#include <cstdlib>
+#include <functional>
+#include <memory>
+#include <random>
+#include <string>
+#include <thread>  // NOLINT(build/c++11)
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "absl/base/internal/raw_logging.h"
+#include "absl/base/internal/sysinfo.h"
+#include "absl/base/macros.h"
+#include "absl/base/thread_annotations.h"
+#include "absl/memory/memory.h"
+#include "absl/synchronization/internal/thread_pool.h"
+#include "absl/time/clock.h"
+#include "absl/time/time.h"
+
+namespace {
+
+// TODO(dmauro): Replace with a commandline flag.
+static constexpr bool kExtendedTest = false;
+
+std::unique_ptr<absl::synchronization_internal::ThreadPool> CreatePool(
+    int threads) {
+  return absl::make_unique<absl::synchronization_internal::ThreadPool>(threads);
+}
+
+std::unique_ptr<absl::synchronization_internal::ThreadPool>
+CreateDefaultPool() {
+  return CreatePool(kExtendedTest ? 32 : 10);
+}
+
+// Hack to schedule a function to run on a thread pool thread after a
+// duration has elapsed.
+static void ScheduleAfter(absl::synchronization_internal::ThreadPool *tp,
+                          const std::function<void()> &func,
+                          absl::Duration after) {
+  tp->Schedule([func, after] {
+    absl::SleepFor(after);
+    func();
+  });
+}
+
+struct TestContext {
+  int iterations;
+  int threads;
+  int g0;  // global 0
+  int g1;  // global 1
+  absl::Mutex mu;
+  absl::CondVar cv;
+};
+
+// To test whether the invariant check call occurs
+static std::atomic<bool> invariant_checked;
+
+static bool GetInvariantChecked() {
+  return invariant_checked.load(std::memory_order_relaxed);
+}
+
+static void SetInvariantChecked(bool new_value) {
+  invariant_checked.store(new_value, std::memory_order_relaxed);
+}
+
+static void CheckSumG0G1(void *v) {
+  TestContext *cxt = static_cast<TestContext *>(v);
+  ABSL_RAW_CHECK(cxt->g0 == -cxt->g1, "Error in CheckSumG0G1");
+  SetInvariantChecked(true);
+}
+
+static void TestMu(TestContext *cxt, int c) {
+  SetInvariantChecked(false);
+  cxt->mu.EnableInvariantDebugging(CheckSumG0G1, cxt);
+  for (int i = 0; i != cxt->iterations; i++) {
+    absl::MutexLock l(&cxt->mu);
+    int a = cxt->g0 + 1;
+    cxt->g0 = a;
+    cxt->g1--;
+  }
+}
+
+static void TestTry(TestContext *cxt, int c) {
+  SetInvariantChecked(false);
+  cxt->mu.EnableInvariantDebugging(CheckSumG0G1, cxt);
+  for (int i = 0; i != cxt->iterations; i++) {
+    do {
+      std::this_thread::yield();
+    } while (!cxt->mu.TryLock());
+    int a = cxt->g0 + 1;
+    cxt->g0 = a;
+    cxt->g1--;
+    cxt->mu.Unlock();
+  }
+}
+
+static void TestR20ms(TestContext *cxt, int c) {
+  for (int i = 0; i != cxt->iterations; i++) {
+    absl::ReaderMutexLock l(&cxt->mu);
+    absl::SleepFor(absl::Milliseconds(20));
+    cxt->mu.AssertReaderHeld();
+  }
+}
+
+static void TestRW(TestContext *cxt, int c) {
+  SetInvariantChecked(false);
+  cxt->mu.EnableInvariantDebugging(CheckSumG0G1, cxt);
+  if ((c & 1) == 0) {
+    for (int i = 0; i != cxt->iterations; i++) {
+      absl::WriterMutexLock l(&cxt->mu);
+      cxt->g0++;
+      cxt->g1--;
+      cxt->mu.AssertHeld();
+      cxt->mu.AssertReaderHeld();
+    }
+  } else {
+    for (int i = 0; i != cxt->iterations; i++) {
+      absl::ReaderMutexLock l(&cxt->mu);
+      ABSL_RAW_CHECK(cxt->g0 == -cxt->g1, "Error in TestRW");
+      cxt->mu.AssertReaderHeld();
+    }
+  }
+}
+
+struct MyContext {
+  int target;
+  TestContext *cxt;
+  bool MyTurn();
+};
+
+bool MyContext::MyTurn() {
+  TestContext *cxt = this->cxt;
+  return cxt->g0 == this->target || cxt->g0 == cxt->iterations;
+}
+
+static void TestAwait(TestContext *cxt, int c) {
+  MyContext mc;
+  mc.target = c;
+  mc.cxt = cxt;
+  absl::MutexLock l(&cxt->mu);
+  cxt->mu.AssertHeld();
+  while (cxt->g0 < cxt->iterations) {
+    cxt->mu.Await(absl::Condition(&mc, &MyContext::MyTurn));
+    ABSL_RAW_CHECK(mc.MyTurn(), "Error in TestAwait");
+    cxt->mu.AssertHeld();
+    if (cxt->g0 < cxt->iterations) {
+      int a = cxt->g0 + 1;
+      cxt->g0 = a;
+      mc.target += cxt->threads;
+    }
+  }
+}
+
+static void TestSignalAll(TestContext *cxt, int c) {
+  int target = c;
+  absl::MutexLock l(&cxt->mu);
+  cxt->mu.AssertHeld();
+  while (cxt->g0 < cxt->iterations) {
+    while (cxt->g0 != target && cxt->g0 != cxt->iterations) {
+      cxt->cv.Wait(&cxt->mu);
+    }
+    if (cxt->g0 < cxt->iterations) {
+      int a = cxt->g0 + 1;
+      cxt->g0 = a;
+      cxt->cv.SignalAll();
+      target += cxt->threads;
+    }
+  }
+}
+
+static void TestSignal(TestContext *cxt, int c) {
+  ABSL_RAW_CHECK(cxt->threads == 2, "TestSignal should use 2 threads");
+  int target = c;
+  absl::MutexLock l(&cxt->mu);
+  cxt->mu.AssertHeld();
+  while (cxt->g0 < cxt->iterations) {
+    while (cxt->g0 != target && cxt->g0 != cxt->iterations) {
+      cxt->cv.Wait(&cxt->mu);
+    }
+    if (cxt->g0 < cxt->iterations) {
+      int a = cxt->g0 + 1;
+      cxt->g0 = a;
+      cxt->cv.Signal();
+      target += cxt->threads;
+    }
+  }
+}
+
+static void TestCVTimeout(TestContext *cxt, int c) {
+  int target = c;
+  absl::MutexLock l(&cxt->mu);
+  cxt->mu.AssertHeld();
+  while (cxt->g0 < cxt->iterations) {
+    while (cxt->g0 != target && cxt->g0 != cxt->iterations) {
+      cxt->cv.WaitWithTimeout(&cxt->mu, absl::Seconds(100));
+    }
+    if (cxt->g0 < cxt->iterations) {
+      int a = cxt->g0 + 1;
+      cxt->g0 = a;
+      cxt->cv.SignalAll();
+      target += cxt->threads;
+    }
+  }
+}
+
+static bool G0GE2(TestContext *cxt) { return cxt->g0 >= 2; }
+
+static void TestTime(TestContext *cxt, int c, bool use_cv) {
+  ABSL_RAW_CHECK(cxt->iterations == 1, "TestTime should only use 1 iteration");
+  ABSL_RAW_CHECK(cxt->threads > 2, "TestTime should use more than 2 threads");
+  const bool kFalse = false;
+  absl::Condition false_cond(&kFalse);
+  absl::Condition g0ge2(G0GE2, cxt);
+  if (c == 0) {
+    absl::MutexLock l(&cxt->mu);
+
+    absl::Time start = absl::Now();
+    if (use_cv) {
+      cxt->cv.WaitWithTimeout(&cxt->mu, absl::Seconds(1));
+    } else {
+      ABSL_RAW_CHECK(!cxt->mu.AwaitWithTimeout(false_cond, absl::Seconds(1)),
+                     "TestTime failed");
+    }
+    absl::Duration elapsed = absl::Now() - start;
+    ABSL_RAW_CHECK(
+        absl::Seconds(0.9) <= elapsed && elapsed <= absl::Seconds(2.0),
+        "TestTime failed");
+    ABSL_RAW_CHECK(cxt->g0 == 1, "TestTime failed");
+
+    start = absl::Now();
+    if (use_cv) {
+      cxt->cv.WaitWithTimeout(&cxt->mu, absl::Seconds(1));
+    } else {
+      ABSL_RAW_CHECK(!cxt->mu.AwaitWithTimeout(false_cond, absl::Seconds(1)),
+                     "TestTime failed");
+    }
+    elapsed = absl::Now() - start;
+    ABSL_RAW_CHECK(
+        absl::Seconds(0.9) <= elapsed && elapsed <= absl::Seconds(2.0),
+        "TestTime failed");
+    cxt->g0++;
+    if (use_cv) {
+      cxt->cv.Signal();
+    }
+
+    start = absl::Now();
+    if (use_cv) {
+      cxt->cv.WaitWithTimeout(&cxt->mu, absl::Seconds(4));
+    } else {
+      ABSL_RAW_CHECK(!cxt->mu.AwaitWithTimeout(false_cond, absl::Seconds(4)),
+                     "TestTime failed");
+    }
+    elapsed = absl::Now() - start;
+    ABSL_RAW_CHECK(
+        absl::Seconds(3.9) <= elapsed && elapsed <= absl::Seconds(6.0),
+        "TestTime failed");
+    ABSL_RAW_CHECK(cxt->g0 >= 3, "TestTime failed");
+
+    start = absl::Now();
+    if (use_cv) {
+      cxt->cv.WaitWithTimeout(&cxt->mu, absl::Seconds(1));
+    } else {
+      ABSL_RAW_CHECK(!cxt->mu.AwaitWithTimeout(false_cond, absl::Seconds(1)),
+                     "TestTime failed");
+    }
+    elapsed = absl::Now() - start;
+    ABSL_RAW_CHECK(
+        absl::Seconds(0.9) <= elapsed && elapsed <= absl::Seconds(2.0),
+        "TestTime failed");
+    if (use_cv) {
+      cxt->cv.SignalAll();
+    }
+
+    start = absl::Now();
+    if (use_cv) {
+      cxt->cv.WaitWithTimeout(&cxt->mu, absl::Seconds(1));
+    } else {
+      ABSL_RAW_CHECK(!cxt->mu.AwaitWithTimeout(false_cond, absl::Seconds(1)),
+                     "TestTime failed");
+    }
+    elapsed = absl::Now() - start;
+    ABSL_RAW_CHECK(absl::Seconds(0.9) <= elapsed &&
+                   elapsed <= absl::Seconds(2.0), "TestTime failed");
+    ABSL_RAW_CHECK(cxt->g0 == cxt->threads, "TestTime failed");
+
+  } else if (c == 1) {
+    absl::MutexLock l(&cxt->mu);
+    const absl::Time start = absl::Now();
+    if (use_cv) {
+      cxt->cv.WaitWithTimeout(&cxt->mu, absl::Milliseconds(500));
+    } else {
+      ABSL_RAW_CHECK(
+          !cxt->mu.AwaitWithTimeout(false_cond, absl::Milliseconds(500)),
+          "TestTime failed");
+    }
+    const absl::Duration elapsed = absl::Now() - start;
+    ABSL_RAW_CHECK(
+        absl::Seconds(0.4) <= elapsed && elapsed <= absl::Seconds(0.9),
+        "TestTime failed");
+    cxt->g0++;
+  } else if (c == 2) {
+    absl::MutexLock l(&cxt->mu);
+    if (use_cv) {
+      while (cxt->g0 < 2) {
+        cxt->cv.WaitWithTimeout(&cxt->mu, absl::Seconds(100));
+      }
+    } else {
+      ABSL_RAW_CHECK(cxt->mu.AwaitWithTimeout(g0ge2, absl::Seconds(100)),
+                     "TestTime failed");
+    }
+    cxt->g0++;
+  } else {
+    absl::MutexLock l(&cxt->mu);
+    if (use_cv) {
+      while (cxt->g0 < 2) {
+        cxt->cv.Wait(&cxt->mu);
+      }
+    } else {
+      cxt->mu.Await(g0ge2);
+    }
+    cxt->g0++;
+  }
+}
+
+static void TestMuTime(TestContext *cxt, int c) { TestTime(cxt, c, false); }
+
+static void TestCVTime(TestContext *cxt, int c) { TestTime(cxt, c, true); }
+
+static void EndTest(int *c0, int *c1, absl::Mutex *mu, absl::CondVar *cv,
+                    const std::function<void(int)>& cb) {
+  mu->Lock();
+  int c = (*c0)++;
+  mu->Unlock();
+  cb(c);
+  absl::MutexLock l(mu);
+  (*c1)++;
+  cv->Signal();
+}
+
+// Basis for the parameterized tests configured below.
+static int RunTest(void (*test)(TestContext *cxt, int), int threads,
+                   int iterations, int operations) {
+  TestContext cxt;
+  absl::Mutex mu2;
+  absl::CondVar cv2;
+  int c0;
+  int c1;
+
+  // run with large thread count for full test and to get timing
+
+#if !defined(ABSL_MUTEX_ENABLE_INVARIANT_DEBUGGING_NOT_IMPLEMENTED)
+  absl::EnableMutexInvariantDebugging(false);
+#endif
+  c0 = 0;
+  c1 = 0;
+  cxt.g0 = 0;
+  cxt.g1 = 0;
+  cxt.iterations = iterations;
+  cxt.threads = threads;
+  absl::synchronization_internal::ThreadPool tp(threads);
+  for (int i = 0; i != threads; i++) {
+    tp.Schedule(std::bind(&EndTest, &c0, &c1, &mu2, &cv2,
+                          std::function<void(int)>(
+                              std::bind(test, &cxt, std::placeholders::_1))));
+  }
+  mu2.Lock();
+  while (c1 != threads) {
+    cv2.Wait(&mu2);
+  }
+  mu2.Unlock();
+  int saved_g0 = cxt.g0;
+
+  // run again with small number of iterations to test invariant checking
+
+#if !defined(ABSL_MUTEX_ENABLE_INVARIANT_DEBUGGING_NOT_IMPLEMENTED)
+  absl::EnableMutexInvariantDebugging(true);
+#endif
+  SetInvariantChecked(true);
+  c0 = 0;
+  c1 = 0;
+  cxt.g0 = 0;
+  cxt.g1 = 0;
+  cxt.iterations = (iterations > 10 ? 10 : iterations);
+  cxt.threads = threads;
+  for (int i = 0; i != threads; i++) {
+    tp.Schedule(std::bind(&EndTest, &c0, &c1, &mu2, &cv2,
+                          std::function<void(int)>(
+                              std::bind(test, &cxt, std::placeholders::_1))));
+  }
+  mu2.Lock();
+  while (c1 != threads) {
+    cv2.Wait(&mu2);
+  }
+  mu2.Unlock();
+#if !defined(ABSL_MUTEX_ENABLE_INVARIANT_DEBUGGING_NOT_IMPLEMENTED)
+  ABSL_RAW_CHECK(GetInvariantChecked(), "Invariant not checked");
+#endif
+
+  return saved_g0;
+}
+
+// --------------------------------------------------------
+// Test for fix of bug in TryRemove()
+struct TimeoutBugStruct {
+  absl::Mutex mu;
+  bool a;
+  int a_waiter_count;
+};
+
+static void WaitForA(TimeoutBugStruct *x) {
+  x->mu.LockWhen(absl::Condition(&x->a));
+  x->a_waiter_count--;
+  x->mu.Unlock();
+}
+
+static bool NoAWaiters(TimeoutBugStruct *x) { return x->a_waiter_count == 0; }
+
+// Test that a CondVar.Wait(&mutex) can un-block a call to mutex.Await() in
+// another thread.
+TEST(Mutex, CondVarWaitSignalsAwait) {
+  // Use a struct so the lock annotations apply.
+  struct {
+    absl::Mutex barrier_mu;
+    bool barrier GUARDED_BY(barrier_mu) = false;
+
+    absl::Mutex release_mu;
+    bool release GUARDED_BY(release_mu) = false;
+    absl::CondVar released_cv;
+  } state;
+
+  auto pool = CreateDefaultPool();
+
+  // Thread A.  Sets barrier, waits for release using Mutex::Await, then
+  // signals released_cv.
+  pool->Schedule([&state] {
+    state.release_mu.Lock();
+
+    state.barrier_mu.Lock();
+    state.barrier = true;
+    state.barrier_mu.Unlock();
+
+    state.release_mu.Await(absl::Condition(&state.release));
+    state.released_cv.Signal();
+    state.release_mu.Unlock();
+  });
+
+  state.barrier_mu.LockWhen(absl::Condition(&state.barrier));
+  state.barrier_mu.Unlock();
+  state.release_mu.Lock();
+  // Thread A is now blocked on release by way of Mutex::Await().
+
+  // Set release.  Calling released_cv.Wait() should un-block thread A,
+  // which will signal released_cv.  If not, the test will hang.
+  state.release = true;
+  state.released_cv.Wait(&state.release_mu);
+  state.release_mu.Unlock();
+}
+
+// Test that a CondVar.WaitWithTimeout(&mutex) can un-block a call to
+// mutex.Await() in another thread.
+TEST(Mutex, CondVarWaitWithTimeoutSignalsAwait) {
+  // Use a struct so the lock annotations apply.
+  struct {
+    absl::Mutex barrier_mu;
+    bool barrier GUARDED_BY(barrier_mu) = false;
+
+    absl::Mutex release_mu;
+    bool release GUARDED_BY(release_mu) = false;
+    absl::CondVar released_cv;
+  } state;
+
+  auto pool = CreateDefaultPool();
+
+  // Thread A.  Sets barrier, waits for release using Mutex::Await, then
+  // signals released_cv.
+  pool->Schedule([&state] {
+    state.release_mu.Lock();
+
+    state.barrier_mu.Lock();
+    state.barrier = true;
+    state.barrier_mu.Unlock();
+
+    state.release_mu.Await(absl::Condition(&state.release));
+    state.released_cv.Signal();
+    state.release_mu.Unlock();
+  });
+
+  state.barrier_mu.LockWhen(absl::Condition(&state.barrier));
+  state.barrier_mu.Unlock();
+  state.release_mu.Lock();
+  // Thread A is now blocked on release by way of Mutex::Await().
+
+  // Set release.  Calling released_cv.Wait() should un-block thread A,
+  // which will signal released_cv.  If not, the test will hang.
+  state.release = true;
+  EXPECT_TRUE(
+      !state.released_cv.WaitWithTimeout(&state.release_mu, absl::Seconds(10)))
+      << "; Unrecoverable test failure: CondVar::WaitWithTimeout did not "
+         "unblock the absl::Mutex::Await call in another thread.";
+
+  state.release_mu.Unlock();
+}
+
+// Test for regression of a bug in loop of TryRemove()
+TEST(Mutex, MutexTimeoutBug) {
+  auto tp = CreateDefaultPool();
+
+  TimeoutBugStruct x;
+  x.a = false;
+  x.a_waiter_count = 2;
+  tp->Schedule(std::bind(&WaitForA, &x));
+  tp->Schedule(std::bind(&WaitForA, &x));
+  absl::SleepFor(absl::Seconds(1));  // Allow first two threads to hang.
+  // The skip field of the second will point to the first because there are
+  // only two.
+
+  // Now cause a thread waiting on an always-false to time out
+  // This would deadlock when the bug was present.
+  bool always_false = false;
+  x.mu.LockWhenWithTimeout(absl::Condition(&always_false),
+                           absl::Milliseconds(500));
+
+  // if we get here, the bug is not present.   Cleanup the state.
+
+  x.a = true;                                    // wakeup the two waiters on A
+  x.mu.Await(absl::Condition(&NoAWaiters, &x));  // wait for them to exit
+  x.mu.Unlock();
+}
+
+struct CondVarWaitDeadlock : testing::TestWithParam<int> {
+  absl::Mutex mu;
+  absl::CondVar cv;
+  bool cond1 = false;
+  bool cond2 = false;
+  bool read_lock1;
+  bool read_lock2;
+  bool signal_unlocked;
+
+  CondVarWaitDeadlock() {
+    read_lock1 = GetParam() & (1 << 0);
+    read_lock2 = GetParam() & (1 << 1);
+    signal_unlocked = GetParam() & (1 << 2);
+  }
+
+  void Waiter1() {
+    if (read_lock1) {
+      mu.ReaderLock();
+      while (!cond1) {
+        cv.Wait(&mu);
+      }
+      mu.ReaderUnlock();
+    } else {
+      mu.Lock();
+      while (!cond1) {
+        cv.Wait(&mu);
+      }
+      mu.Unlock();
+    }
+  }
+
+  void Waiter2() {
+    if (read_lock2) {
+      mu.ReaderLockWhen(absl::Condition(&cond2));
+      mu.ReaderUnlock();
+    } else {
+      mu.LockWhen(absl::Condition(&cond2));
+      mu.Unlock();
+    }
+  }
+};
+
+// Test for a deadlock bug in Mutex::Fer().
+// The sequence of events that lead to the deadlock is:
+// 1. waiter1 blocks on cv in read mode (mu bits = 0).
+// 2. waiter2 blocks on mu in either mode (mu bits = kMuWait).
+// 3. main thread locks mu, sets cond1, unlocks mu (mu bits = kMuWait).
+// 4. main thread signals on cv and this eventually calls Mutex::Fer().
+// Currently Fer wakes waiter1 since mu bits = kMuWait (mutex is unlocked).
+// Before the bug fix Fer neither woke waiter1 nor queued it on mutex,
+// which resulted in deadlock.
+TEST_P(CondVarWaitDeadlock, Test) {
+  auto waiter1 = CreatePool(1);
+  auto waiter2 = CreatePool(1);
+  waiter1->Schedule([this] { this->Waiter1(); });
+  waiter2->Schedule([this] { this->Waiter2(); });
+
+  // Wait while threads block (best-effort is fine).
+  absl::SleepFor(absl::Milliseconds(100));
+
+  // Wake condwaiter.
+  mu.Lock();
+  cond1 = true;
+  if (signal_unlocked) {
+    mu.Unlock();
+    cv.Signal();
+  } else {
+    cv.Signal();
+    mu.Unlock();
+  }
+  waiter1.reset();  // "join" waiter1
+
+  // Wake waiter.
+  mu.Lock();
+  cond2 = true;
+  mu.Unlock();
+  waiter2.reset();  // "join" waiter2
+}
+
+INSTANTIATE_TEST_CASE_P(CondVarWaitDeadlockTest, CondVarWaitDeadlock,
+                        ::testing::Range(0, 8),
+                        ::testing::PrintToStringParamName());
+
+// --------------------------------------------------------
+// Test for fix of bug in DequeueAllWakeable()
+// Bug was that if there was more than one waiting reader
+// and all should be woken, the most recently blocked one
+// would not be.
+
+struct DequeueAllWakeableBugStruct {
+  absl::Mutex mu;
+  absl::Mutex mu2;       // protects all fields below
+  int unfinished_count;  // count of unfinished readers; under mu2
+  bool done1;            // unfinished_count == 0; under mu2
+  int finished_count;    // count of finished readers, under mu2
+  bool done2;            // finished_count == 0; under mu2
+};
+
+// Test for regression of a bug in loop of DequeueAllWakeable()
+static void AcquireAsReader(DequeueAllWakeableBugStruct *x) {
+  x->mu.ReaderLock();
+  x->mu2.Lock();
+  x->unfinished_count--;
+  x->done1 = (x->unfinished_count == 0);
+  x->mu2.Unlock();
+  // make sure that both readers acquired mu before we release it.
+  absl::SleepFor(absl::Seconds(2));
+  x->mu.ReaderUnlock();
+
+  x->mu2.Lock();
+  x->finished_count--;
+  x->done2 = (x->finished_count == 0);
+  x->mu2.Unlock();
+}
+
+// Test for regression of a bug in loop of DequeueAllWakeable()
+TEST(Mutex, MutexReaderWakeupBug) {
+  auto tp = CreateDefaultPool();
+
+  DequeueAllWakeableBugStruct x;
+  x.unfinished_count = 2;
+  x.done1 = false;
+  x.finished_count = 2;
+  x.done2 = false;
+  x.mu.Lock();  // acquire mu exclusively
+  // queue two thread that will block on reader locks on x.mu
+  tp->Schedule(std::bind(&AcquireAsReader, &x));
+  tp->Schedule(std::bind(&AcquireAsReader, &x));
+  absl::SleepFor(absl::Seconds(1));  // give time for reader threads to block
+  x.mu.Unlock();                     // wake them up
+
+  // both readers should finish promptly
+  EXPECT_TRUE(
+      x.mu2.LockWhenWithTimeout(absl::Condition(&x.done1), absl::Seconds(10)));
+  x.mu2.Unlock();
+
+  EXPECT_TRUE(
+      x.mu2.LockWhenWithTimeout(absl::Condition(&x.done2), absl::Seconds(10)));
+  x.mu2.Unlock();
+}
+
+struct LockWhenTestStruct {
+  absl::Mutex mu1;
+  bool cond = false;
+
+  absl::Mutex mu2;
+  bool waiting = false;
+};
+
+static bool LockWhenTestIsCond(LockWhenTestStruct* s) {
+  s->mu2.Lock();
+  s->waiting = true;
+  s->mu2.Unlock();
+  return s->cond;
+}
+
+static void LockWhenTestWaitForIsCond(LockWhenTestStruct* s) {
+  s->mu1.LockWhen(absl::Condition(&LockWhenTestIsCond, s));
+  s->mu1.Unlock();
+}
+
+TEST(Mutex, LockWhen) {
+  LockWhenTestStruct s;
+
+  // Don't use ThreadPool for this test. See b/65107115.
+  std::thread t(LockWhenTestWaitForIsCond, &s);
+  s.mu2.LockWhen(absl::Condition(&s.waiting));
+  s.mu2.Unlock();
+
+  s.mu1.Lock();
+  s.cond = true;
+  s.mu1.Unlock();
+
+  t.join();
+}
+
+// --------------------------------------------------------
+// The following test requires Mutex::ReaderLock to be a real shared
+// lock, which is not the case in all builds.
+#if !defined(ABSL_MUTEX_READER_LOCK_IS_EXCLUSIVE)
+
+// Test for fix of bug in UnlockSlow() that incorrectly decremented the reader
+// count when putting a thread to sleep waiting for a false condition when the
+// lock was not held.
+
+// For this bug to strike, we make a thread wait on a free mutex with no
+// waiters by causing its wakeup condition to be false.   Then the
+// next two acquirers must be readers.   The bug causes the lock
+// to be released when one reader unlocks, rather than both.
+
+struct ReaderDecrementBugStruct {
+  bool cond;  // to delay first thread (under mu)
+  int done;   // reference count (under mu)
+  absl::Mutex mu;
+
+  bool waiting_on_cond;   // under mu2
+  bool have_reader_lock;  // under mu2
+  bool complete;          // under mu2
+  absl::Mutex mu2;        // > mu
+};
+
+// L >= mu, L < mu_waiting_on_cond
+static bool IsCond(void *v) {
+  ReaderDecrementBugStruct *x = reinterpret_cast<ReaderDecrementBugStruct *>(v);
+  x->mu2.Lock();
+  x->waiting_on_cond = true;
+  x->mu2.Unlock();
+  return x->cond;
+}
+
+// L >= mu
+static bool AllDone(void *v) {
+  ReaderDecrementBugStruct *x = reinterpret_cast<ReaderDecrementBugStruct *>(v);
+  return x->done == 0;
+}
+
+// L={}
+static void WaitForCond(ReaderDecrementBugStruct *x) {
+  absl::Mutex dummy;
+  absl::MutexLock l(&dummy);
+  x->mu.LockWhen(absl::Condition(&IsCond, x));
+  x->done--;
+  x->mu.Unlock();
+}
+
+// L={}
+static void GetReadLock(ReaderDecrementBugStruct *x) {
+  x->mu.ReaderLock();
+  x->mu2.Lock();
+  x->have_reader_lock = true;
+  x->mu2.Await(absl::Condition(&x->complete));
+  x->mu2.Unlock();
+  x->mu.ReaderUnlock();
+  x->mu.Lock();
+  x->done--;
+  x->mu.Unlock();
+}
+
+// Test for reader counter being decremented incorrectly by waiter
+// with false condition.
+TEST(Mutex, MutexReaderDecrementBug) NO_THREAD_SAFETY_ANALYSIS {
+  ReaderDecrementBugStruct x;
+  x.cond = false;
+  x.waiting_on_cond = false;
+  x.have_reader_lock = false;
+  x.complete = false;
+  x.done = 2;  // initial ref count
+
+  // Run WaitForCond() and wait for it to sleep
+  std::thread thread1(WaitForCond, &x);
+  x.mu2.LockWhen(absl::Condition(&x.waiting_on_cond));
+  x.mu2.Unlock();
+
+  // Run GetReadLock(), and wait for it to get the read lock
+  std::thread thread2(GetReadLock, &x);
+  x.mu2.LockWhen(absl::Condition(&x.have_reader_lock));
+  x.mu2.Unlock();
+
+  // Get the reader lock ourselves, and release it.
+  x.mu.ReaderLock();
+  x.mu.ReaderUnlock();
+
+  // The lock should be held in read mode by GetReadLock().
+  // If we have the bug, the lock will be free.
+  x.mu.AssertReaderHeld();
+
+  // Wake up all the threads.
+  x.mu2.Lock();
+  x.complete = true;
+  x.mu2.Unlock();
+
+  // TODO(delesley): turn on analysis once lock upgrading is supported.
+  // (This call upgrades the lock from shared to exclusive.)
+  x.mu.Lock();
+  x.cond = true;
+  x.mu.Await(absl::Condition(&AllDone, &x));
+  x.mu.Unlock();
+
+  thread1.join();
+  thread2.join();
+}
+#endif  // !ABSL_MUTEX_READER_LOCK_IS_EXCLUSIVE
+
+// Test that we correctly handle the situation when a lock is
+// held and then destroyed (w/o unlocking).
+TEST(Mutex, LockedMutexDestructionBug) NO_THREAD_SAFETY_ANALYSIS {
+  for (int i = 0; i != 10; i++) {
+    // Create, lock and destroy 10 locks.
+    const int kNumLocks = 10;
+    auto mu = absl::make_unique<absl::Mutex[]>(kNumLocks);
+    for (int j = 0; j != kNumLocks; j++) {
+      if ((j % 2) == 0) {
+        mu[j].WriterLock();
+      } else {
+        mu[j].ReaderLock();
+      }
+    }
+  }
+}
+
+// --------------------------------------------------------
+// Test for bug with pattern of readers using a condvar.  The bug was that if a
+// reader went to sleep on a condition variable while one or more other readers
+// held the lock, but there were no waiters, the reader count (held in the
+// mutex word) would be lost.  (This is because Enqueue() had at one time
+// always placed the thread on the Mutex queue.  Later (CL 4075610), to
+// tolerate re-entry into Mutex from a Condition predicate, Enqueue() was
+// changed so that it could also place a thread on a condition-variable.  This
+// introduced the case where Enqueue() returned with an empty queue, and this
+// case was handled incorrectly in one place.)
+
+static void ReaderForReaderOnCondVar(absl::Mutex *mu, absl::CondVar *cv,
+                                     int *running) {
+  std::random_device dev;
+  std::mt19937 gen(dev());
+  std::uniform_int_distribution<int> random_millis(0, 15);
+  mu->ReaderLock();
+  while (*running == 3) {
+    absl::SleepFor(absl::Milliseconds(random_millis(gen)));
+    cv->WaitWithTimeout(mu, absl::Milliseconds(random_millis(gen)));
+  }
+  mu->ReaderUnlock();
+  mu->Lock();
+  (*running)--;
+  mu->Unlock();
+}
+
+struct True {
+  template <class... Args>
+  bool operator()(Args...) const {
+    return true;
+  }
+};
+
+struct DerivedTrue : True {};
+
+TEST(Mutex, FunctorCondition) {
+  {  // Variadic
+    True f;
+    EXPECT_TRUE(absl::Condition(&f).Eval());
+  }
+
+  {  // Inherited
+    DerivedTrue g;
+    EXPECT_TRUE(absl::Condition(&g).Eval());
+  }
+
+  {  // lambda
+    int value = 3;
+    auto is_zero = [&value] { return value == 0; };
+    absl::Condition c(&is_zero);
+    EXPECT_FALSE(c.Eval());
+    value = 0;
+    EXPECT_TRUE(c.Eval());
+  }
+
+  {  // bind
+    int value = 0;
+    auto is_positive = std::bind(std::less<int>(), 0, std::cref(value));
+    absl::Condition c(&is_positive);
+    EXPECT_FALSE(c.Eval());
+    value = 1;
+    EXPECT_TRUE(c.Eval());
+  }
+
+  {  // std::function
+    int value = 3;
+    std::function<bool()> is_zero = [&value] { return value == 0; };
+    absl::Condition c(&is_zero);
+    EXPECT_FALSE(c.Eval());
+    value = 0;
+    EXPECT_TRUE(c.Eval());
+  }
+}
+
+static bool IntIsZero(int *x) { return *x == 0; }
+
+// Test for reader waiting condition variable when there are other readers
+// but no waiters.
+TEST(Mutex, TestReaderOnCondVar) {
+  auto tp = CreateDefaultPool();
+  absl::Mutex mu;
+  absl::CondVar cv;
+  int running = 3;
+  tp->Schedule(std::bind(&ReaderForReaderOnCondVar, &mu, &cv, &running));
+  tp->Schedule(std::bind(&ReaderForReaderOnCondVar, &mu, &cv, &running));
+  absl::SleepFor(absl::Seconds(2));
+  mu.Lock();
+  running--;
+  mu.Await(absl::Condition(&IntIsZero, &running));
+  mu.Unlock();
+}
+
+// --------------------------------------------------------
+struct AcquireFromConditionStruct {
+  absl::Mutex mu0;   // protects value, done
+  int value;         // times condition function is called; under mu0,
+  bool done;         // done with test?  under mu0
+  absl::Mutex mu1;   // used to attempt to mess up state of mu0
+  absl::CondVar cv;  // so the condition function can be invoked from
+                     // CondVar::Wait().
+};
+
+static bool ConditionWithAcquire(AcquireFromConditionStruct *x) {
+  x->value++;  // count times this function is called
+
+  if (x->value == 2 || x->value == 3) {
+    // On the second and third invocation of this function, sleep for 100ms,
+    // but with the side-effect of altering the state of a Mutex other than
+    // than one for which this is a condition.  The spec now explicitly allows
+    // this side effect; previously it did not.  it was illegal.
+    bool always_false = false;
+    x->mu1.LockWhenWithTimeout(absl::Condition(&always_false),
+                               absl::Milliseconds(100));
+    x->mu1.Unlock();
+  }
+  ABSL_RAW_CHECK(x->value < 4, "should not be invoked a fourth time");
+
+  // We arrange for the condition to return true on only the 2nd and 3rd calls.
+  return x->value == 2 || x->value == 3;
+}
+
+static void WaitForCond2(AcquireFromConditionStruct *x) {
+  // wait for cond0 to become true
+  x->mu0.LockWhen(absl::Condition(&ConditionWithAcquire, x));
+  x->done = true;
+  x->mu0.Unlock();
+}
+
+// Test for Condition whose function acquires other Mutexes
+TEST(Mutex, AcquireFromCondition) {
+  auto tp = CreateDefaultPool();
+
+  AcquireFromConditionStruct x;
+  x.value = 0;
+  x.done = false;
+  tp->Schedule(
+      std::bind(&WaitForCond2, &x));  // run WaitForCond2() in a thread T
+  // T will hang because the first invocation of ConditionWithAcquire() will
+  // return false.
+  absl::SleepFor(absl::Milliseconds(500));  // allow T time to hang
+
+  x.mu0.Lock();
+  x.cv.WaitWithTimeout(&x.mu0, absl::Milliseconds(500));  // wake T
+  // T will be woken because the Wait() will call ConditionWithAcquire()
+  // for the second time, and it will return true.
+
+  x.mu0.Unlock();
+
+  // T will then acquire the lock and recheck its own condition.
+  // It will find the condition true, as this is the third invocation,
+  // but the use of another Mutex by the calling function will
+  // cause the old mutex implementation to think that the outer
+  // LockWhen() has timed out because the inner LockWhenWithTimeout() did.
+  // T will then check the condition a fourth time because it finds a
+  // timeout occurred.  This should not happen in the new
+  // implementation that allows the Condition function to use Mutexes.
+
+  // It should also succeed, even though the Condition function
+  // is being invoked from CondVar::Wait, and thus this thread
+  // is conceptually waiting both on the condition variable, and on mu2.
+
+  x.mu0.LockWhen(absl::Condition(&x.done));
+  x.mu0.Unlock();
+}
+
+// The deadlock detector is not part of non-prod builds, so do not test it.
+#if !defined(ABSL_INTERNAL_USE_NONPROD_MUTEX)
+
+TEST(Mutex, DeadlockDetector) {
+  absl::SetMutexDeadlockDetectionMode(absl::OnDeadlockCycle::kAbort);
+
+  // check that we can call ForgetDeadlockInfo() on a lock with the lock held
+  absl::Mutex m1;
+  absl::Mutex m2;
+  absl::Mutex m3;
+  absl::Mutex m4;
+
+  m1.Lock();  // m1 gets ID1
+  m2.Lock();  // m2 gets ID2
+  m3.Lock();  // m3 gets ID3
+  m3.Unlock();
+  m2.Unlock();
+  // m1 still held
+  m1.ForgetDeadlockInfo();  // m1 loses ID
+  m2.Lock();                // m2 gets ID2
+  m3.Lock();                // m3 gets ID3
+  m4.Lock();                // m4 gets ID4
+  m3.Unlock();
+  m2.Unlock();
+  m4.Unlock();
+  m1.Unlock();
+  // Pre b/7636708 the thread local cache remembered that ID1 is assigned to m1.
+  // So, we had a cycle ID1=>ID1=>ID1.
+}
+
+// Bazel has a test "warning" file that programs can write to if the
+// test should pass with a warning.  This class disables the warning
+// file until it goes out of scope.
+class ScopedDisableBazelTestWarnings {
+ public:
+  ScopedDisableBazelTestWarnings() {
+#ifdef WIN32
+    char file[MAX_PATH];
+    if (GetEnvironmentVariable(kVarName, file, sizeof(file)) < sizeof(file)) {
+      warnings_output_file_ = file;
+      SetEnvironmentVariable(kVarName, nullptr);
+    }
+#else
+    const char *file = getenv(kVarName);
+    if (file != nullptr) {
+      warnings_output_file_ = file;
+      unsetenv(kVarName);
+    }
+#endif
+  }
+
+  ~ScopedDisableBazelTestWarnings() {
+    if (!warnings_output_file_.empty()) {
+#ifdef WIN32
+      SetEnvironmentVariable(kVarName, warnings_output_file_.c_str());
+#else
+      setenv(kVarName, warnings_output_file_.c_str(), 0);
+#endif
+    }
+  }
+
+ private:
+  static const char kVarName[];
+  std::string warnings_output_file_;
+};
+const char ScopedDisableBazelTestWarnings::kVarName[] =
+    "TEST_WARNINGS_OUTPUT_FILE";
+
+TEST(Mutex, DeadlockDetectorBazelWarning) {
+  absl::SetMutexDeadlockDetectionMode(absl::OnDeadlockCycle::kReport);
+
+  // Cause deadlock detection to detect something, if it's
+  // compiled in and enabled.  But turn off the bazel warning.
+  ScopedDisableBazelTestWarnings disable_bazel_test_warnings;
+
+  absl::Mutex mu0;
+  absl::Mutex mu1;
+  bool got_mu0 = mu0.TryLock();
+  mu1.Lock();  // acquire mu1 while holding mu0
+  if (got_mu0) {
+    mu0.Unlock();
+  }
+  if (mu0.TryLock()) {  // try lock shouldn't cause deadlock detector to fire
+    mu0.Unlock();
+  }
+  mu0.Lock();  // acquire mu0 while holding mu1; should get one deadlock
+               // report here
+  mu0.Unlock();
+  mu1.Unlock();
+
+  absl::SetMutexDeadlockDetectionMode(absl::OnDeadlockCycle::kAbort);
+}
+
+// This test is tagged with NO_THREAD_SAFETY_ANALYSIS because the
+// annotation-based static thread-safety analysis is not currently
+// predicate-aware and cannot tell if the two for-loops that acquire and
+// release the locks have the same predicates.
+TEST(Mutex, DeadlockDetectorStessTest) NO_THREAD_SAFETY_ANALYSIS {
+  // Stress test: Here we create a large number of locks and use all of them.
+  // If a deadlock detector keeps a full graph of lock acquisition order,
+  // it will likely be too slow for this test to pass.
+  const int n_locks = 1 << 17;
+  auto array_of_locks = absl::make_unique<absl::Mutex[]>(n_locks);
+  for (int i = 0; i < n_locks; i++) {
+    int end = std::min(n_locks, i + 5);
+    // acquire and then release locks i, i+1, ..., i+4
+    for (int j = i; j < end; j++) {
+      array_of_locks[j].Lock();
+    }
+    for (int j = i; j < end; j++) {
+      array_of_locks[j].Unlock();
+    }
+  }
+}
+
+TEST(Mutex, DeadlockIdBug) NO_THREAD_SAFETY_ANALYSIS {
+  // Test a scenario where a cached deadlock graph node id in the
+  // list of held locks is not invalidated when the corresponding
+  // mutex is deleted.
+  absl::SetMutexDeadlockDetectionMode(absl::OnDeadlockCycle::kAbort);
+  // Mutex that will be destroyed while being held
+  absl::Mutex *a = new absl::Mutex;
+  // Other mutexes needed by test
+  absl::Mutex b, c;
+
+  // Hold mutex.
+  a->Lock();
+
+  // Force deadlock id assignment by acquiring another lock.
+  b.Lock();
+  b.Unlock();
+
+  // Delete the mutex. The Mutex destructor tries to remove held locks,
+  // but the attempt isn't foolproof.  It can fail if:
+  //   (a) Deadlock detection is currently disabled.
+  //   (b) The destruction is from another thread.
+  // We exploit (a) by temporarily disabling deadlock detection.
+  absl::SetMutexDeadlockDetectionMode(absl::OnDeadlockCycle::kIgnore);
+  delete a;
+  absl::SetMutexDeadlockDetectionMode(absl::OnDeadlockCycle::kAbort);
+
+  // Now acquire another lock which will force a deadlock id assignment.
+  // We should end up getting assigned the same deadlock id that was
+  // freed up when "a" was deleted, which will cause a spurious deadlock
+  // report if the held lock entry for "a" was not invalidated.
+  c.Lock();
+  c.Unlock();
+}
+#endif  // !defined(ABSL_INTERNAL_USE_NONPROD_MUTEX)
+
+// --------------------------------------------------------
+// Test for timeouts/deadlines on condition waits that are specified using
+// absl::Duration and absl::Time.  For each waiting function we test with
+// a timeout/deadline that has already expired/passed, one that is infinite
+// and so never expires/passes, and one that will expire/pass in the near
+// future.
+
+// Encapsulate a Mutex-protected bool with its associated Condition/CondVar.
+class Cond {
+ public:
+  explicit Cond(bool use_deadline) : use_deadline_(use_deadline), c_(&b_) {}
+
+  void Set(bool v) {
+    absl::MutexLock lock(&mu_);
+    b_ = v;
+  }
+
+  bool AwaitWithTimeout(absl::Duration timeout) {
+    absl::MutexLock lock(&mu_);
+    return use_deadline_ ? mu_.AwaitWithDeadline(c_, absl::Now() + timeout)
+                         : mu_.AwaitWithTimeout(c_, timeout);
+  }
+
+  bool LockWhenWithTimeout(absl::Duration timeout) {
+    bool b = use_deadline_ ? mu_.LockWhenWithDeadline(c_, absl::Now() + timeout)
+                           : mu_.LockWhenWithTimeout(c_, timeout);
+    mu_.Unlock();
+    return b;
+  }
+
+  bool ReaderLockWhenWithTimeout(absl::Duration timeout) {
+    bool b = use_deadline_
+                 ? mu_.ReaderLockWhenWithDeadline(c_, absl::Now() + timeout)
+                 : mu_.ReaderLockWhenWithTimeout(c_, timeout);
+    mu_.ReaderUnlock();
+    return b;
+  }
+
+  void Await() {
+    absl::MutexLock lock(&mu_);
+    mu_.Await(c_);
+  }
+
+  void Signal(bool v) {
+    absl::MutexLock lock(&mu_);
+    b_ = v;
+    cv_.Signal();
+  }
+
+  bool WaitWithTimeout(absl::Duration timeout) {
+    absl::MutexLock lock(&mu_);
+    absl::Time deadline = absl::Now() + timeout;
+    if (use_deadline_) {
+      while (!b_ && !cv_.WaitWithDeadline(&mu_, deadline)) {
+      }
+    } else {
+      while (!b_ && !cv_.WaitWithTimeout(&mu_, timeout)) {
+        timeout = deadline - absl::Now();  // recompute timeout
+      }
+    }
+    return b_;
+  }
+
+  void Wait() {
+    absl::MutexLock lock(&mu_);
+    while (!b_) cv_.Wait(&mu_);
+  }
+
+ private:
+  const bool use_deadline_;
+
+  bool b_;
+  absl::Condition c_;
+  absl::CondVar cv_;
+  absl::Mutex mu_;
+};
+
+class OperationTimer {
+ public:
+  OperationTimer() : start_(absl::Now()) {}
+  absl::Duration Get() const { return absl::Now() - start_; }
+
+ private:
+  const absl::Time start_;
+};
+
+static void CheckResults(bool exp_result, bool act_result,
+                         absl::Duration exp_duration,
+                         absl::Duration act_duration) {
+  ABSL_RAW_CHECK(exp_result == act_result, "CheckResults failed");
+  // Allow for some worse-case scheduling delay and clock skew.
+  ABSL_RAW_CHECK(exp_duration - absl::Milliseconds(40) <= act_duration,
+                 "CheckResults failed");
+  ABSL_RAW_CHECK(exp_duration + absl::Milliseconds(150) >= act_duration,
+                 "CheckResults failed");
+}
+
+static void TestAwaitTimeout(Cond *cp, absl::Duration timeout, bool exp_result,
+                             absl::Duration exp_duration) {
+  OperationTimer t;
+  bool act_result = cp->AwaitWithTimeout(timeout);
+  CheckResults(exp_result, act_result, exp_duration, t.Get());
+}
+
+static void TestLockWhenTimeout(Cond *cp, absl::Duration timeout,
+                                bool exp_result, absl::Duration exp_duration) {
+  OperationTimer t;
+  bool act_result = cp->LockWhenWithTimeout(timeout);
+  CheckResults(exp_result, act_result, exp_duration, t.Get());
+}
+
+static void TestReaderLockWhenTimeout(Cond *cp, absl::Duration timeout,
+                                      bool exp_result,
+                                      absl::Duration exp_duration) {
+  OperationTimer t;
+  bool act_result = cp->ReaderLockWhenWithTimeout(timeout);
+  CheckResults(exp_result, act_result, exp_duration, t.Get());
+}
+
+static void TestWaitTimeout(Cond *cp, absl::Duration timeout, bool exp_result,
+                            absl::Duration exp_duration) {
+  OperationTimer t;
+  bool act_result = cp->WaitWithTimeout(timeout);
+  CheckResults(exp_result, act_result, exp_duration, t.Get());
+}
+
+// Tests with a negative timeout (deadline in the past), which should
+// immediately return the current state of the condition.
+static void TestNegativeTimeouts(absl::synchronization_internal::ThreadPool *tp,
+                                 Cond *cp) {
+  const absl::Duration negative = -absl::InfiniteDuration();
+  const absl::Duration immediate = absl::ZeroDuration();
+
+  // The condition is already true:
+  cp->Set(true);
+  TestAwaitTimeout(cp, negative, true, immediate);
+  TestLockWhenTimeout(cp, negative, true, immediate);
+  TestReaderLockWhenTimeout(cp, negative, true, immediate);
+  TestWaitTimeout(cp, negative, true, immediate);
+
+  // The condition becomes true, but the timeout has already expired:
+  const absl::Duration delay = absl::Milliseconds(200);
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Set, cp, true), 3 * delay);
+  TestAwaitTimeout(cp, negative, false, immediate);
+  TestLockWhenTimeout(cp, negative, false, immediate);
+  TestReaderLockWhenTimeout(cp, negative, false, immediate);
+  cp->Await();  // wait for the scheduled Set() to complete
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Signal, cp, true), delay);
+  TestWaitTimeout(cp, negative, false, immediate);
+  cp->Wait();  // wait for the scheduled Signal() to complete
+
+  // The condition never becomes true:
+  cp->Set(false);
+  TestAwaitTimeout(cp, negative, false, immediate);
+  TestLockWhenTimeout(cp, negative, false, immediate);
+  TestReaderLockWhenTimeout(cp, negative, false, immediate);
+  TestWaitTimeout(cp, negative, false, immediate);
+}
+
+// Tests with an infinite timeout (deadline in the infinite future), which
+// should only return when the condition becomes true.
+static void TestInfiniteTimeouts(absl::synchronization_internal::ThreadPool *tp,
+                                 Cond *cp) {
+  const absl::Duration infinite = absl::InfiniteDuration();
+  const absl::Duration immediate = absl::ZeroDuration();
+
+  // The condition is already true:
+  cp->Set(true);
+  TestAwaitTimeout(cp, infinite, true, immediate);
+  TestLockWhenTimeout(cp, infinite, true, immediate);
+  TestReaderLockWhenTimeout(cp, infinite, true, immediate);
+  TestWaitTimeout(cp, infinite, true, immediate);
+
+  // The condition becomes true before the (infinite) expiry:
+  const absl::Duration delay = absl::Milliseconds(200);
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Set, cp, true), delay);
+  TestAwaitTimeout(cp, infinite, true, delay);
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Set, cp, true), delay);
+  TestLockWhenTimeout(cp, infinite, true, delay);
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Set, cp, true), delay);
+  TestReaderLockWhenTimeout(cp, infinite, true, delay);
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Signal, cp, true), delay);
+  TestWaitTimeout(cp, infinite, true, delay);
+}
+
+// Tests with a (small) finite timeout (deadline soon), with the condition
+// becoming true both before and after its expiry.
+static void TestFiniteTimeouts(absl::synchronization_internal::ThreadPool *tp,
+                               Cond *cp) {
+  const absl::Duration finite = absl::Milliseconds(400);
+  const absl::Duration immediate = absl::ZeroDuration();
+
+  // The condition is already true:
+  cp->Set(true);
+  TestAwaitTimeout(cp, finite, true, immediate);
+  TestLockWhenTimeout(cp, finite, true, immediate);
+  TestReaderLockWhenTimeout(cp, finite, true, immediate);
+  TestWaitTimeout(cp, finite, true, immediate);
+
+  // The condition becomes true before the expiry:
+  const absl::Duration delay1 = finite / 2;
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Set, cp, true), delay1);
+  TestAwaitTimeout(cp, finite, true, delay1);
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Set, cp, true), delay1);
+  TestLockWhenTimeout(cp, finite, true, delay1);
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Set, cp, true), delay1);
+  TestReaderLockWhenTimeout(cp, finite, true, delay1);
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Signal, cp, true), delay1);
+  TestWaitTimeout(cp, finite, true, delay1);
+
+  // The condition becomes true, but the timeout has already expired:
+  const absl::Duration delay2 = finite * 2;
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Set, cp, true), 3 * delay2);
+  TestAwaitTimeout(cp, finite, false, finite);
+  TestLockWhenTimeout(cp, finite, false, finite);
+  TestReaderLockWhenTimeout(cp, finite, false, finite);
+  cp->Await();  // wait for the scheduled Set() to complete
+  cp->Set(false);
+  ScheduleAfter(tp, std::bind(&Cond::Signal, cp, true), delay2);
+  TestWaitTimeout(cp, finite, false, finite);
+  cp->Wait();  // wait for the scheduled Signal() to complete
+
+  // The condition never becomes true:
+  cp->Set(false);
+  TestAwaitTimeout(cp, finite, false, finite);
+  TestLockWhenTimeout(cp, finite, false, finite);
+  TestReaderLockWhenTimeout(cp, finite, false, finite);
+  TestWaitTimeout(cp, finite, false, finite);
+}
+
+TEST(Mutex, Timeouts) {
+  auto tp = CreateDefaultPool();
+  for (bool use_deadline : {false, true}) {
+    Cond cond(use_deadline);
+    TestNegativeTimeouts(tp.get(), &cond);
+    TestInfiniteTimeouts(tp.get(), &cond);
+    TestFiniteTimeouts(tp.get(), &cond);
+  }
+}
+
+TEST(Mutex, Logging) {
+  // Allow user to look at logging output
+  absl::Mutex logged_mutex;
+  logged_mutex.EnableDebugLog("fido_mutex");
+  absl::CondVar logged_cv;
+  logged_cv.EnableDebugLog("rover_cv");
+  logged_mutex.Lock();
+  logged_cv.WaitWithTimeout(&logged_mutex, absl::Milliseconds(20));
+  logged_mutex.Unlock();
+  logged_mutex.ReaderLock();
+  logged_mutex.ReaderUnlock();
+  logged_mutex.Lock();
+  logged_mutex.Unlock();
+  logged_cv.Signal();
+  logged_cv.SignalAll();
+}
+
+// --------------------------------------------------------
+
+// Generate the vector of thread counts for tests parameterized on thread count.
+static std::vector<int> AllThreadCountValues() {
+  if (kExtendedTest) {
+    return {2, 4, 8, 10, 16, 20, 24, 30, 32};
+  }
+  return {2, 4, 10};
+}
+
+// A test fixture parameterized by thread count.
+class MutexVariableThreadCountTest : public ::testing::TestWithParam<int> {};
+
+// Instantiate the above with AllThreadCountOptions().
+INSTANTIATE_TEST_CASE_P(ThreadCounts, MutexVariableThreadCountTest,
+                        ::testing::ValuesIn(AllThreadCountValues()),
+                        ::testing::PrintToStringParamName());
+
+// Reduces iterations by some factor for slow platforms
+// (determined empirically).
+static int ScaleIterations(int x) {
+  // ABSL_MUTEX_READER_LOCK_IS_EXCLUSIVE is set in the implementation
+  // of Mutex that uses either std::mutex or pthread_mutex_t. Use
+  // these as keys to determine the slow implementation.
+#if defined(ABSL_MUTEX_READER_LOCK_IS_EXCLUSIVE)
+  return x / 10;
+#else
+  return x;
+#endif
+}
+
+TEST_P(MutexVariableThreadCountTest, Mutex) {
+  int threads = GetParam();
+  int iterations = ScaleIterations(10000000) / threads;
+  int operations = threads * iterations;
+  EXPECT_EQ(RunTest(&TestMu, threads, iterations, operations), operations);
+}
+
+TEST_P(MutexVariableThreadCountTest, Try) {
+  int threads = GetParam();
+  int iterations = 1000000 / threads;
+  int operations = iterations * threads;
+  EXPECT_EQ(RunTest(&TestTry, threads, iterations, operations), operations);
+}
+
+TEST_P(MutexVariableThreadCountTest, R20ms) {
+  int threads = GetParam();
+  int iterations = 100;
+  int operations = iterations * threads;
+  EXPECT_EQ(RunTest(&TestR20ms, threads, iterations, operations), 0);
+}
+
+TEST_P(MutexVariableThreadCountTest, RW) {
+  int threads = GetParam();
+  int iterations = ScaleIterations(20000000) / threads;
+  int operations = iterations * threads;
+  EXPECT_EQ(RunTest(&TestRW, threads, iterations, operations), operations / 2);
+}
+
+TEST_P(MutexVariableThreadCountTest, Await) {
+  int threads = GetParam();
+  int iterations = ScaleIterations(500000);
+  int operations = iterations;
+  EXPECT_EQ(RunTest(&TestAwait, threads, iterations, operations), operations);
+}
+
+TEST_P(MutexVariableThreadCountTest, SignalAll) {
+  int threads = GetParam();
+  int iterations = 200000 / threads;
+  int operations = iterations;
+  EXPECT_EQ(RunTest(&TestSignalAll, threads, iterations, operations),
+            operations);
+}
+
+TEST(Mutex, Signal) {
+  int threads = 2;  // TestSignal must use two threads
+  int iterations = 200000;
+  int operations = iterations;
+  EXPECT_EQ(RunTest(&TestSignal, threads, iterations, operations), operations);
+}
+
+TEST(Mutex, Timed) {
+  int threads = 10;  // Use a fixed thread count of 10
+  int iterations = 1000;
+  int operations = iterations;
+  EXPECT_EQ(RunTest(&TestCVTimeout, threads, iterations, operations),
+            operations);
+}
+
+TEST(Mutex, CVTime) {
+  int threads = 10;  // Use a fixed thread count of 10
+  int iterations = 1;
+  EXPECT_EQ(RunTest(&TestCVTime, threads, iterations, 1),
+            threads * iterations);
+}
+
+TEST(Mutex, MuTime) {
+  int threads = 10;  // Use a fixed thread count of 10
+  int iterations = 1;
+  EXPECT_EQ(RunTest(&TestMuTime, threads, iterations, 1), threads * iterations);
+}
+
+}  // namespace
diff --git a/absl/synchronization/notification.cc b/absl/synchronization/notification.cc
new file mode 100644
index 000000000000..ed8cc9067740
--- /dev/null
+++ b/absl/synchronization/notification.cc
@@ -0,0 +1,84 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/synchronization/notification.h"
+
+#include <atomic>
+
+#include "absl/base/attributes.h"
+#include "absl/base/internal/raw_logging.h"
+#include "absl/synchronization/mutex.h"
+#include "absl/time/time.h"
+
+namespace absl {
+void Notification::Notify() {
+  MutexLock l(&this->mutex_);
+
+#ifndef NDEBUG
+  if (ABSL_PREDICT_FALSE(notified_yet_.load(std::memory_order_relaxed))) {
+    ABSL_RAW_LOG(
+        FATAL,
+        "Notify() method called more than once for Notification object %p",
+        static_cast<void *>(this));
+  }
+#endif
+
+  notified_yet_.store(true, std::memory_order_release);
+}
+
+Notification::~Notification() {
+  // Make sure that the thread running Notify() exits before the object is
+  // destructed.
+  MutexLock l(&this->mutex_);
+}
+
+static inline bool HasBeenNotifiedInternal(
+    const std::atomic<bool> *notified_yet) {
+  return notified_yet->load(std::memory_order_acquire);
+}
+
+bool Notification::HasBeenNotified() const {
+  return HasBeenNotifiedInternal(&this->notified_yet_);
+}
+
+void Notification::WaitForNotification() const {
+  if (!HasBeenNotifiedInternal(&this->notified_yet_)) {
+    this->mutex_.LockWhen(Condition(&HasBeenNotifiedInternal,
+                                    &this->notified_yet_));
+    this->mutex_.Unlock();
+  }
+}
+
+bool Notification::WaitForNotificationWithTimeout(
+    absl::Duration timeout) const {
+  bool notified = HasBeenNotifiedInternal(&this->notified_yet_);
+  if (!notified) {
+    notified = this->mutex_.LockWhenWithTimeout(
+        Condition(&HasBeenNotifiedInternal, &this->notified_yet_), timeout);
+    this->mutex_.Unlock();
+  }
+  return notified;
+}
+
+bool Notification::WaitForNotificationWithDeadline(absl::Time deadline) const {
+  bool notified = HasBeenNotifiedInternal(&this->notified_yet_);
+  if (!notified) {
+    notified = this->mutex_.LockWhenWithDeadline(
+        Condition(&HasBeenNotifiedInternal, &this->notified_yet_), deadline);
+    this->mutex_.Unlock();
+  }
+  return notified;
+}
+
+}  // namespace absl
diff --git a/absl/synchronization/notification.h b/absl/synchronization/notification.h
new file mode 100644
index 000000000000..107932f2909e
--- /dev/null
+++ b/absl/synchronization/notification.h
@@ -0,0 +1,112 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+//
+// -----------------------------------------------------------------------------
+// notification.h
+// -----------------------------------------------------------------------------
+//
+// This header file defines a `Notification` abstraction, which allows threads
+// to receive notification of a single occurrence of a single event.
+//
+// The `Notification` object maintains a private boolean "notified" state that
+// transitions to `true` at most once. The `Notification` class provides the
+// following primary member functions:
+//   * `HasBeenNotified() `to query its state
+//   * `WaitForNotification*()` to have threads wait until the "notified" state
+//      is `true`.
+//   * `Notify()` to set the notification's "notified" state to `true` and
+//     notify all waiting threads that the event has occurred.
+//     This method may only be called once.
+//
+// Note that while `Notify()` may only be called once, it is perfectly valid to
+// call any of the `WaitForNotification*()` methods multiple times, from
+// multiple threads -- even after the notification's "notified" state has been
+// set -- in which case those methods will immediately return.
+//
+// Note that the lifetime of a `Notification` requires careful consideration;
+// it might not be safe to destroy a notification after calling `Notify()` since
+// it is still legal for other threads to call `WaitForNotification*()` methods
+// on the notification. However, observers responding to a "notified" state of
+// `true` can safely delete the notification without interfering with the call
+// to `Notify()` in the other thread.
+//
+// Memory ordering: For any threads X and Y, if X calls `Notify()`, then any
+// action taken by X before it calls `Notify()` is visible to thread Y after:
+//  * Y returns from `WaitForNotification()`, or
+//  * Y receives a `true` return value from either `HasBeenNotified()` or
+//    `WaitForNotificationWithTimeout()`.
+
+#ifndef ABSL_SYNCHRONIZATION_NOTIFICATION_H_
+#define ABSL_SYNCHRONIZATION_NOTIFICATION_H_
+
+#include <atomic>
+
+#include "absl/synchronization/mutex.h"
+#include "absl/time/time.h"
+
+namespace absl {
+
+// -----------------------------------------------------------------------------
+// Notification
+// -----------------------------------------------------------------------------
+class Notification {
+ public:
+  // Initializes the "notified" state to unnotified.
+  Notification() : notified_yet_(false) {}
+  explicit Notification(bool prenotify) : notified_yet_(prenotify) {}
+  Notification(const Notification&) = delete;
+  Notification& operator=(const Notification&) = delete;
+  ~Notification();
+
+  // Notification::HasBeenNotified()
+  //
+  // Returns the value of the notification's internal "notified" state.
+  bool HasBeenNotified() const;
+
+  // Notification::WaitForNotification()
+  //
+  // Blocks the calling thread until the notification's "notified" state is
+  // `true`. Note that if `Notify()` has been previously called on this
+  // notification, this function will immediately return.
+  void WaitForNotification() const;
+
+  // Notification::WaitForNotificationWithTimeout()
+  //
+  // Blocks until either the notification's "notified" state is `true` (which
+  // may occur immediately) or the timeout has elapsed, returning the value of
+  // its "notified" state in either case.
+  bool WaitForNotificationWithTimeout(absl::Duration timeout) const;
+
+  // Notification::WaitForNotificationWithDeadline()
+  //
+  // Blocks until either the notification's "notified" state is `true` (which
+  // may occur immediately) or the deadline has expired, returning the value of
+  // its "notified" state in either case.
+  bool WaitForNotificationWithDeadline(absl::Time deadline) const;
+
+  // Notification::Notify()
+  //
+  // Sets the "notified" state of this notification to `true` and wakes waiting
+  // threads. Note: do not call `Notify()` multiple times on the same
+  // `Notification`; calling `Notify()` more than once on the same notification
+  // results in undefined behavior.
+  void Notify();
+
+ private:
+  mutable Mutex mutex_;
+  std::atomic<bool> notified_yet_;  // written under mutex_
+};
+
+}  // namespace absl
+#endif  // ABSL_SYNCHRONIZATION_NOTIFICATION_H_
diff --git a/absl/synchronization/notification_test.cc b/absl/synchronization/notification_test.cc
new file mode 100644
index 000000000000..9b3b6a5a9e84
--- /dev/null
+++ b/absl/synchronization/notification_test.cc
@@ -0,0 +1,124 @@
+// Copyright 2017 The Abseil Authors.
+//
+// Licensed under the Apache License, Version 2.0 (the "License");
+// you may not use this file except in compliance with the License.
+// You may obtain a copy of the License at
+//
+//      http://www.apache.org/licenses/LICENSE-2.0
+//
+// Unless required by applicable law or agreed to in writing, software
+// distributed under the License is distributed on an "AS IS" BASIS,
+// WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+// See the License for the specific language governing permissions and
+// limitations under the License.
+
+#include "absl/synchronization/notification.h"
+
+#include <thread>  // NOLINT(build/c++11)
+#include <vector>
+
+#include "gtest/gtest.h"
+#include "absl/synchronization/mutex.h"
+
+namespace absl {
+
+// A thread-safe class that holds a counter.
+class ThreadSafeCounter {
+ public:
+  ThreadSafeCounter() : count_(0) {}
+
+  void Increment() {
+    MutexLock lock(&mutex_);
+    ++count_;
+  }
+
+  int Get() const {
+    MutexLock lock(&mutex_);
+    return count_;
+  }
+
+  void WaitUntilGreaterOrEqual(int n) {
+    MutexLock lock(&mutex_);
+    auto cond = [this, n]() { return count_ >= n; };
+    mutex_.Await(Condition(&cond));
+  }
+
+ private:
+  mutable Mutex mutex_;
+  int count_;
+};
+
+// Runs the |i|'th worker thread for the tests in BasicTests().  Increments the
+// |ready_counter|, waits on the |notification|, and then increments the
+// |done_counter|.
+static void RunWorker(int i, ThreadSafeCounter* ready_counter,
+                      Notification* notification,
+                      ThreadSafeCounter* done_counter) {
+  ready_counter->Increment();
+  notification->WaitForNotification();
+  done_counter->Increment();
+}
+
+// Tests that the |notification| properly blocks and awakens threads.  Assumes
+// that the |notification| is not yet triggered.  If |notify_before_waiting| is
+// true, the |notification| is triggered before any threads are created, so the
+// threads never block in WaitForNotification().  Otherwise, the |notification|
+// is triggered at a later point when most threads are likely to be blocking in
+// WaitForNotification().
+static void BasicTests(bool notify_before_waiting, Notification* notification) {
+  EXPECT_FALSE(notification->HasBeenNotified());
+  EXPECT_FALSE(
+      notification->WaitForNotificationWithTimeout(absl::Milliseconds(0)));
+  EXPECT_FALSE(notification->WaitForNotificationWithDeadline(absl::Now()));
+
+  absl::Time start = absl::Now();
+  EXPECT_FALSE(
+      notification->WaitForNotificationWithTimeout(absl::Milliseconds(50)));
+  EXPECT_LE(start + absl::Milliseconds(50), absl::Now());
+
+  ThreadSafeCounter ready_counter;
+  ThreadSafeCounter done_counter;
+
+  if (notify_before_waiting) {
+    notification->Notify();
+  }
+
+  // Create a bunch of threads that increment the |done_counter| after being
+  // notified.
+  const int kNumThreads = 10;
+  std::vector<std::thread> workers;
+  for (int i = 0; i < kNumThreads; ++i) {
+    workers.push_back(std::thread(&RunWorker, i, &ready_counter, notification,
+                                  &done_counter));
+  }
+
+  if (!notify_before_waiting) {
+    ready_counter.WaitUntilGreaterOrEqual(kNumThreads);
+
+    // Workers have not been notified yet, so the |done_counter| should be
+    // unmodified.
+    EXPECT_EQ(0, done_counter.Get());
+
+    notification->Notify();
+  }
+
+  // After notifying and then joining the workers, both counters should be
+  // fully incremented.
+  notification->WaitForNotification();  // should exit immediately
+  EXPECT_TRUE(notification->HasBeenNotified());
+  EXPECT_TRUE(notification->WaitForNotificationWithTimeout(absl::Seconds(0)));
+  EXPECT_TRUE(notification->WaitForNotificationWithDeadline(absl::Now()));
+  for (std::thread& worker : workers) {
+    worker.join();
+  }
+  EXPECT_EQ(kNumThreads, ready_counter.Get());
+  EXPECT_EQ(kNumThreads, done_counter.Get());
+}
+
+TEST(NotificationTest, SanityTest) {
+  Notification local_notification1, local_notification2;
+  BasicTests(false, &local_notification1);
+  BasicTests(true, &local_notification2);
+}
+
+}  // namespace absl