diff --git a/libs/binder/ndk/include_ndk/android/binder_ibinder.h b/libs/binder/ndk/include_ndk/android/binder_ibinder.h
index 0ca3a07..bb70588 100644
--- a/libs/binder/ndk/include_ndk/android/binder_ibinder.h
+++ b/libs/binder/ndk/include_ndk/android/binder_ibinder.h
@@ -660,13 +660,15 @@
 /**
  * Whether AIBinder is less than another.
  *
- * This provides a per-process-unique total ordering of binders determined by
- * an underlying allocation address where a null AIBinder* is considered to be
- * ordered before all other binders.
+ * This provides a per-process-unique total ordering of binders where a null
+ * AIBinder* object is considered to be before all other binder objects.
+ * For instance, two binders refer to the same object in a local or remote
+ * process when both AIBinder_lt(a, b) and AIBinder(b, a) are false. This API
+ * might be used to insert and lookup binders in binary search trees.
  *
  * AIBinder* pointers themselves actually also create a per-process-unique total
  * ordering. However, this ordering is inconsistent with AIBinder_Weak_lt for
- * remote binders.
+ * remote binders. So, in general, this function should be preferred.
  *
  * Available since API level 31.
  *
@@ -698,14 +700,21 @@
  * the same as AIBinder_lt. Similarly, a null AIBinder_Weak* is considered to be
  * ordered before all other weak references.
  *
- * If you have many AIBinder_Weak* objects which are all references to distinct
- * binder objects which happen to have the same underlying address (as ordered
- * by AIBinder_lt), these AIBinder_Weak* objects will retain the same order with
- * respect to all other AIBinder_Weak* pointers with different underlying
- * addresses and are also guaranteed to have a per-process-unique ordering. That
- * is, even though multiple AIBinder* instances may happen to be allocated at
- * the same underlying address, this function will still correctly distinguish
- * that these are weak pointers to different binder objects.
+ * This function correctly distinguishes binders even if one is deallocated. So,
+ * for instance, an AIBinder_Weak* entry representing a deleted binder will
+ * never compare as equal to an AIBinder_Weak* entry which represents a
+ * different allocation of a binder, even if the two binders were originally
+ * allocated at the same address. That is:
+ *
+ *     AIBinder* a = ...; // imagine this has address 0x8
+ *     AIBinder_Weak* bWeak = AIBinder_Weak_new(a);
+ *     AIBinder_decStrong(a); // a may be deleted, if this is the last reference
+ *     AIBinder* b = ...; // imagine this has address 0x8 (same address as b)
+ *     AIBinder_Weak* bWeak = AIBinder_Weak_new(b);
+ *
+ * Then when a/b are compared with other binders, their order will be preserved,
+ * and it will either be the case that AIBinder_Weak_lt(aWeak, bWeak) OR
+ * AIBinder_Weak_lt(bWeak, aWeak), but not both.
  *
  * Unlike AIBinder*, the AIBinder_Weak* addresses themselves have nothing to do
  * with the underlying binder.
diff --git a/services/gpuservice/tests/unittests/Android.bp b/services/gpuservice/tests/unittests/Android.bp
index 940a26b..9606daa 100644
--- a/services/gpuservice/tests/unittests/Android.bp
+++ b/services/gpuservice/tests/unittests/Android.bp
@@ -20,6 +20,7 @@
     },
     srcs: [
         "GpuMemTest.cpp",
+        "GpuMemTracerTest.cpp",
         "GpuStatsTest.cpp",
     ],
     shared_libs: [
@@ -29,14 +30,19 @@
         "libcutils",
         "libgfxstats",
         "libgpumem",
+        "libgpumemtracer",
         "libgraphicsenv",
         "liblog",
+        "libprotobuf-cpp-lite",
+        "libprotoutil",
         "libstatslog",
         "libstatspull",
         "libutils",
     ],
     static_libs: [
         "libgmock",
+        "libperfetto_client_experimental",
+        "perfetto_trace_protos",
     ],
     require_root: true,
 }
diff --git a/services/gpuservice/tests/unittests/GpuMemTracerTest.cpp b/services/gpuservice/tests/unittests/GpuMemTracerTest.cpp
new file mode 100644
index 0000000..d76f039
--- /dev/null
+++ b/services/gpuservice/tests/unittests/GpuMemTracerTest.cpp
@@ -0,0 +1,196 @@
+/*
+ * Copyright 2020 The Android Open Source Project
+ *
+ * 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.
+ */
+
+#undef LOG_TAG
+#define LOG_TAG "gpuservice_unittest"
+
+#include <bpf/BpfMap.h>
+#include <gpumem/GpuMem.h>
+#include <gtest/gtest.h>
+#include <perfetto/trace/trace.pb.h>
+#include <tracing/GpuMemTracer.h>
+
+#include "TestableGpuMem.h"
+
+namespace android {
+
+constexpr uint32_t TEST_MAP_SIZE = 10;
+constexpr uint64_t TEST_GLOBAL_KEY = 0;
+constexpr uint32_t TEST_GLOBAL_PID = 0;
+constexpr uint64_t TEST_GLOBAL_VAL = 123;
+constexpr uint32_t TEST_GLOBAL_GPU_ID = 0;
+constexpr uint64_t TEST_PROC_KEY_1 = 1;
+constexpr uint32_t TEST_PROC_PID_1 = 1;
+constexpr uint64_t TEST_PROC_VAL_1 = 234;
+constexpr uint32_t TEST_PROC_1_GPU_ID = 0;
+constexpr uint64_t TEST_PROC_KEY_2 = 4294967298; // (1 << 32) + 2
+constexpr uint32_t TEST_PROC_PID_2 = 2;
+constexpr uint64_t TEST_PROC_VAL_2 = 345;
+constexpr uint32_t TEST_PROC_2_GPU_ID = 1;
+
+class GpuMemTracerTest : public testing::Test {
+public:
+    GpuMemTracerTest() {
+        const ::testing::TestInfo* const test_info =
+                ::testing::UnitTest::GetInstance()->current_test_info();
+        ALOGD("**** Setting up for %s.%s\n", test_info->test_case_name(), test_info->name());
+    }
+
+    ~GpuMemTracerTest() {
+        const ::testing::TestInfo* const test_info =
+                ::testing::UnitTest::GetInstance()->current_test_info();
+        ALOGD("**** Tearing down after %s.%s\n", test_info->test_case_name(), test_info->name());
+    }
+
+    void SetUp() override {
+        bpf::setrlimitForTest();
+
+        mGpuMem = std::make_shared<GpuMem>();
+        mGpuMemTracer = std::make_unique<GpuMemTracer>();
+        mGpuMemTracer->initializeForTest(mGpuMem);
+        mTestableGpuMem = TestableGpuMem(mGpuMem.get());
+
+        errno = 0;
+        mTestMap = bpf::BpfMap<uint64_t, uint64_t>(BPF_MAP_TYPE_HASH, TEST_MAP_SIZE,
+                                                   BPF_F_NO_PREALLOC);
+
+        EXPECT_EQ(0, errno);
+        EXPECT_LE(0, mTestMap.getMap().get());
+        EXPECT_TRUE(mTestMap.isValid());
+    }
+
+    int getTracerThreadCount() { return mGpuMemTracer->tracerThreadCount; }
+
+    std::vector<perfetto::protos::TracePacket> readGpuMemTotalPacketsBlocking(
+            perfetto::TracingSession* tracingSession) {
+        std::vector<char> raw_trace = tracingSession->ReadTraceBlocking();
+        perfetto::protos::Trace trace;
+        trace.ParseFromArray(raw_trace.data(), int(raw_trace.size()));
+
+        std::vector<perfetto::protos::TracePacket> packets;
+        for (const auto& packet : trace.packet()) {
+            if (!packet.has_gpu_mem_total_event()) {
+                continue;
+            }
+            packets.emplace_back(packet);
+        }
+        return packets;
+    }
+
+    std::shared_ptr<GpuMem> mGpuMem;
+    TestableGpuMem mTestableGpuMem;
+    std::unique_ptr<GpuMemTracer> mGpuMemTracer;
+    bpf::BpfMap<uint64_t, uint64_t> mTestMap;
+};
+
+static constexpr uint64_t getSizeForPid(uint32_t pid) {
+    switch (pid) {
+        case TEST_GLOBAL_PID:
+            return TEST_GLOBAL_VAL;
+        case TEST_PROC_PID_1:
+            return TEST_PROC_VAL_1;
+        case TEST_PROC_PID_2:
+            return TEST_PROC_VAL_2;
+    }
+    return 0;
+}
+
+static constexpr uint32_t getGpuIdForPid(uint32_t pid) {
+    switch (pid) {
+        case TEST_GLOBAL_PID:
+            return TEST_GLOBAL_GPU_ID;
+        case TEST_PROC_PID_1:
+            return TEST_PROC_1_GPU_ID;
+        case TEST_PROC_PID_2:
+            return TEST_PROC_2_GPU_ID;
+    }
+    return 0;
+}
+
+TEST_F(GpuMemTracerTest, traceInitialCountersAfterGpuMemInitialize) {
+    ASSERT_RESULT_OK(mTestMap.writeValue(TEST_GLOBAL_KEY, TEST_GLOBAL_VAL, BPF_ANY));
+    ASSERT_RESULT_OK(mTestMap.writeValue(TEST_PROC_KEY_1, TEST_PROC_VAL_1, BPF_ANY));
+    ASSERT_RESULT_OK(mTestMap.writeValue(TEST_PROC_KEY_2, TEST_PROC_VAL_2, BPF_ANY));
+    mTestableGpuMem.setGpuMemTotalMap(mTestMap);
+    mTestableGpuMem.setInitialized();
+
+    // Only 1 tracer thread should be existing for test.
+    EXPECT_EQ(getTracerThreadCount(), 1);
+    auto tracingSession = mGpuMemTracer->getTracingSessionForTest();
+
+    tracingSession->StartBlocking();
+    // Sleep for a short time to let the tracer thread finish its work
+    sleep(1);
+    tracingSession->StopBlocking();
+
+    // The test tracer thread should have finished its execution by now.
+    EXPECT_EQ(getTracerThreadCount(), 0);
+
+    auto packets = readGpuMemTotalPacketsBlocking(tracingSession.get());
+    EXPECT_EQ(packets.size(), 3);
+
+    const auto& packet0 = packets[0];
+    ASSERT_TRUE(packet0.has_timestamp());
+    ASSERT_TRUE(packet0.has_gpu_mem_total_event());
+    const auto& gpuMemEvent0 = packet0.gpu_mem_total_event();
+    ASSERT_TRUE(gpuMemEvent0.has_pid());
+    const auto& pid0 = gpuMemEvent0.pid();
+    ASSERT_TRUE(gpuMemEvent0.has_size());
+    EXPECT_EQ(gpuMemEvent0.size(), getSizeForPid(pid0));
+    ASSERT_TRUE(gpuMemEvent0.has_gpu_id());
+    EXPECT_EQ(gpuMemEvent0.gpu_id(), getGpuIdForPid(pid0));
+
+    const auto& packet1 = packets[1];
+    ASSERT_TRUE(packet1.has_timestamp());
+    ASSERT_TRUE(packet1.has_gpu_mem_total_event());
+    const auto& gpuMemEvent1 = packet1.gpu_mem_total_event();
+    ASSERT_TRUE(gpuMemEvent1.has_pid());
+    const auto& pid1 = gpuMemEvent1.pid();
+    ASSERT_TRUE(gpuMemEvent1.has_size());
+    EXPECT_EQ(gpuMemEvent1.size(), getSizeForPid(pid1));
+    ASSERT_TRUE(gpuMemEvent1.has_gpu_id());
+    EXPECT_EQ(gpuMemEvent1.gpu_id(), getGpuIdForPid(pid1));
+
+    const auto& packet2 = packets[2];
+    ASSERT_TRUE(packet2.has_timestamp());
+    ASSERT_TRUE(packet2.has_gpu_mem_total_event());
+    const auto& gpuMemEvent2 = packet2.gpu_mem_total_event();
+    ASSERT_TRUE(gpuMemEvent2.has_pid());
+    const auto& pid2 = gpuMemEvent2.pid();
+    ASSERT_TRUE(gpuMemEvent2.has_size());
+    EXPECT_EQ(gpuMemEvent2.size(), getSizeForPid(pid2));
+    ASSERT_TRUE(gpuMemEvent2.has_gpu_id());
+    EXPECT_EQ(gpuMemEvent2.gpu_id(), getGpuIdForPid(pid2));
+}
+
+TEST_F(GpuMemTracerTest, noTracingWithoutGpuMemInitialize) {
+    // Only 1 tracer thread should be existing for test.
+    EXPECT_EQ(getTracerThreadCount(), 1);
+
+    auto tracingSession = mGpuMemTracer->getTracingSessionForTest();
+
+    tracingSession->StartBlocking();
+    // Sleep for a short time to let the tracer thread finish its work
+    sleep(1);
+    tracingSession->StopBlocking();
+
+    // The test tracer thread should have finished its execution by now.
+    EXPECT_EQ(getTracerThreadCount(), 0);
+
+    auto packets = readGpuMemTotalPacketsBlocking(tracingSession.get());
+    EXPECT_EQ(packets.size(), 0);
+}
+} // namespace android
diff --git a/services/gpuservice/tracing/GpuMemTracer.cpp b/services/gpuservice/tracing/GpuMemTracer.cpp
index 6366e1d..44a30ea 100644
--- a/services/gpuservice/tracing/GpuMemTracer.cpp
+++ b/services/gpuservice/tracing/GpuMemTracer.cpp
@@ -44,9 +44,35 @@
     args.backends = perfetto::kSystemBackend;
     perfetto::Tracing::Initialize(args);
     registerDataSource();
-    std::thread tracerThread(&GpuMemTracer::threadLoop, this);
+    std::thread tracerThread(&GpuMemTracer::threadLoop, this, true);
     pthread_setname_np(tracerThread.native_handle(), "GpuMemTracerThread");
     tracerThread.detach();
+    tracerThreadCount++;
+}
+
+void GpuMemTracer::initializeForTest(std::shared_ptr<GpuMem> gpuMem) {
+    mGpuMem = gpuMem;
+    perfetto::TracingInitArgs args;
+    args.backends = perfetto::kInProcessBackend;
+    perfetto::Tracing::Initialize(args);
+    registerDataSource();
+    std::thread tracerThread(&GpuMemTracer::threadLoop, this, false);
+    pthread_setname_np(tracerThread.native_handle(), "GpuMemTracerThreadForTest");
+    tracerThread.detach();
+    tracerThreadCount++;
+}
+
+// Each tracing session can be used for a single block of Start -> Stop.
+std::unique_ptr<perfetto::TracingSession> GpuMemTracer::getTracingSessionForTest() {
+    perfetto::TraceConfig cfg;
+    cfg.set_duration_ms(500);
+    cfg.add_buffers()->set_size_kb(1024);
+    auto* ds_cfg = cfg.add_data_sources()->mutable_config();
+    ds_cfg->set_name(GpuMemTracer::kGpuMemDataSource);
+
+    auto tracingSession = perfetto::Tracing::NewTrace(perfetto::kInProcessBackend);
+    tracingSession->Setup(cfg);
+    return tracingSession;
 }
 
 void GpuMemTracer::registerDataSource() {
@@ -55,8 +81,8 @@
     GpuMemDataSource::Register(dsd);
 }
 
-void GpuMemTracer::threadLoop() {
-    while (true) {
+void GpuMemTracer::threadLoop(bool infiniteLoop) {
+    do {
         {
             std::unique_lock<std::mutex> lock(GpuMemTracer::sTraceMutex);
             while (!sTraceStarted) {
@@ -68,7 +94,11 @@
             std::lock_guard<std::mutex> lock(GpuMemTracer::sTraceMutex);
             sTraceStarted = false;
         }
-    }
+    } while (infiniteLoop);
+
+    // Thread loop is exiting. Reduce the tracerThreadCount to reflect the number of active threads
+    // in the wait loop.
+    tracerThreadCount--;
 }
 
 void GpuMemTracer::traceInitialCounters() {
diff --git a/services/gpuservice/tracing/include/tracing/GpuMemTracer.h b/services/gpuservice/tracing/include/tracing/GpuMemTracer.h
index 40deb4c..ae871f1 100644
--- a/services/gpuservice/tracing/include/tracing/GpuMemTracer.h
+++ b/services/gpuservice/tracing/include/tracing/GpuMemTracer.h
@@ -20,6 +20,10 @@
 
 #include <mutex>
 
+namespace perfetto::protos {
+class TracePacket;
+}
+
 namespace android {
 
 class GpuMem;
@@ -45,16 +49,37 @@
     // perfetto::kInProcessBackend in tests.
     void registerDataSource();
 
+    // TODO(b/175904796): Refactor gpuservice lib to include perfetto lib and move the test
+    // functions into the unittests.
+    // Functions only used for testing with in-process backend. These functions require the static
+    // perfetto lib to be linked. If the tests have a perfetto linked, while libgpumemtracer.so also
+    // has one linked, they will both use different static states maintained in perfetto. Since the
+    // static perfetto states are not shared, tracing sessions created in the unit test are not
+    // recognized by GpuMemTracer. As a result, we cannot use any of the perfetto functions from
+    // this class, which defeats the purpose of the unit test. To solve this, we restrict all
+    // tracing functionality to this class, while the unit test validates the data.
+    // Sets up the perfetto in-process backend and calls into registerDataSource.
+    void initializeForTest(std::shared_ptr<GpuMem>);
+    // Creates a tracing session with in process backend, for testing.
+    std::unique_ptr<perfetto::TracingSession> getTracingSessionForTest();
+    // Read and filter the gpu memory packets from the created trace.
+    std::vector<perfetto::protos::TracePacket> readGpuMemTotalPacketsForTestBlocking(
+            perfetto::TracingSession* tracingSession);
+
     static constexpr char kGpuMemDataSource[] = "android.gpu.memory";
     static std::condition_variable sCondition;
     static std::mutex sTraceMutex;
     static bool sTraceStarted;
 
 private:
-    void traceInitialCounters();
-    void threadLoop();
+    // Friend class for testing
+    friend class GpuMemTracerTest;
 
+    void threadLoop(bool infiniteLoop);
+    void traceInitialCounters();
     std::shared_ptr<GpuMem> mGpuMem;
+    // Count of how many tracer threads are currently active. Useful for testing.
+    std::atomic<int32_t> tracerThreadCount = 0;
 };
 
 } // namespace android