diff --git a/NOTICE b/NOTICE
deleted file mode 100644
index 8e8a91c..0000000
--- a/NOTICE
+++ /dev/null
@@ -1,16 +0,0 @@
-Copyright (C) 2017 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.
-
--------------------------------------------------------------------
-
diff --git a/bootstat/bootstat.cpp b/bootstat/bootstat.cpp
index ee1ae31..d6ebb0d 100644
--- a/bootstat/bootstat.cpp
+++ b/bootstat/bootstat.cpp
@@ -438,6 +438,7 @@
     {"reboot,mount_userdata_failed", 190},
     {"reboot,forcedsilent", 191},
     {"reboot,forcednonsilent", 192},
+    {"reboot,thermal,tj", 193},
 };
 
 // Converts a string value representing the reason the system booted to an
diff --git a/debuggerd/Android.bp b/debuggerd/Android.bp
index fb274ec..7b6f6c0 100644
--- a/debuggerd/Android.bp
+++ b/debuggerd/Android.bp
@@ -238,6 +238,7 @@
         "gwp_asan_crash_handler",
         "libscudo",
         "libtombstone_proto",
+        "libprocinfo",
         "libprotobuf-cpp-lite",
     ],
 
diff --git a/debuggerd/crash_dump.cpp b/debuggerd/crash_dump.cpp
index 68a43cf..04e1e4e 100644
--- a/debuggerd/crash_dump.cpp
+++ b/debuggerd/crash_dump.cpp
@@ -153,14 +153,14 @@
   }
 
   struct timeval tv = {
-      .tv_sec = 1 * android::base::TimeoutMultiplier(),
+      .tv_sec = 1 * android::base::HwTimeoutMultiplier(),
       .tv_usec = 0,
   };
   if (setsockopt(amfd.get(), SOL_SOCKET, SO_SNDTIMEO, &tv, sizeof(tv)) == -1) {
     PLOG(ERROR) << "failed to set send timeout on activity manager socket";
     return false;
   }
-  tv.tv_sec = 3 * android::base::TimeoutMultiplier();  // 3 seconds on handshake read
+  tv.tv_sec = 3 * android::base::HwTimeoutMultiplier();  // 3 seconds on handshake read
   if (setsockopt(amfd.get(), SOL_SOCKET, SO_RCVTIMEO, &tv, sizeof(tv)) == -1) {
     PLOG(ERROR) << "failed to set receive timeout on activity manager socket";
     return false;
@@ -303,6 +303,7 @@
       process_info->gwp_asan_metadata = crash_info->data.d.gwp_asan_metadata;
       process_info->scudo_stack_depot = crash_info->data.d.scudo_stack_depot;
       process_info->scudo_region_info = crash_info->data.d.scudo_region_info;
+      process_info->scudo_ring_buffer = crash_info->data.d.scudo_ring_buffer;
       FALLTHROUGH_INTENDED;
     case 1:
     case 2:
@@ -390,7 +391,7 @@
 
   // There appears to be a bug in the kernel where our death causes SIGHUP to
   // be sent to our process group if we exit while it has stopped jobs (e.g.
-  // because of wait_for_gdb). Use setsid to create a new process group to
+  // because of wait_for_debugger). Use setsid to create a new process group to
   // avoid hitting this.
   setsid();
 
@@ -447,7 +448,7 @@
   //
   // Note: processes with many threads and minidebug-info can take a bit to
   //       unwind, do not make this too small. b/62828735
-  alarm(30 * android::base::TimeoutMultiplier());
+  alarm(30 * android::base::HwTimeoutMultiplier());
 
   // Get the process name (aka cmdline).
   std::string process_name = get_process_name(g_target_thread);
@@ -547,15 +548,17 @@
   fork_exit_write.reset();
 
   // Defer the message until later, for readability.
-  bool wait_for_gdb = android::base::GetBoolProperty("debug.debuggerd.wait_for_gdb", false);
+  bool wait_for_debugger = android::base::GetBoolProperty(
+      "debug.debuggerd.wait_for_debugger",
+      android::base::GetBoolProperty("debug.debuggerd.wait_for_gdb", false));
   if (siginfo.si_signo == BIONIC_SIGNAL_DEBUGGER) {
-    wait_for_gdb = false;
+    wait_for_debugger = false;
   }
 
   // Detach from all of our attached threads before resuming.
   for (const auto& [tid, thread] : thread_info) {
     int resume_signal = thread.signo == BIONIC_SIGNAL_DEBUGGER ? 0 : thread.signo;
-    if (wait_for_gdb) {
+    if (wait_for_debugger) {
       resume_signal = 0;
       if (tgkill(target_process, tid, SIGSTOP) != 0) {
         PLOG(WARNING) << "failed to send SIGSTOP to " << tid;
@@ -640,12 +643,12 @@
     }
   }
 
-  if (wait_for_gdb) {
+  if (wait_for_debugger) {
     // Use ALOGI to line up with output from engrave_tombstone.
     ALOGI(
         "***********************************************************\n"
         "* Process %d has been suspended while crashing.\n"
-        "* To attach gdbserver and start gdb, run this on the host:\n"
+        "* To attach the debugger, run this on the host:\n"
         "*\n"
         "*     gdbclient.py -p %d\n"
         "*\n"
diff --git a/debuggerd/crasher/Android.bp b/debuggerd/crasher/Android.bp
index 7975a3a..23b106e 100644
--- a/debuggerd/crasher/Android.bp
+++ b/debuggerd/crasher/Android.bp
@@ -13,7 +13,6 @@
         "-Werror",
         "-O0",
         "-fstack-protector-all",
-        "-Wno-free-nonheap-object",
         "-Wno-date-time",
     ],
     srcs: ["crasher.cpp"],
diff --git a/debuggerd/crasher/crasher.cpp b/debuggerd/crasher/crasher.cpp
index a2b13a3..db30b8f 100644
--- a/debuggerd/crasher/crasher.cpp
+++ b/debuggerd/crasher/crasher.cpp
@@ -134,10 +134,14 @@
     return a*2;
 }
 
+#pragma clang diagnostic push
+#pragma clang diagnostic ignored "-Wfree-nonheap-object"
+
 noinline void abuse_heap() {
     char buf[16];
     free(buf); // GCC is smart enough to warn about this, but we're doing it deliberately.
 }
+#pragma clang diagnostic pop
 
 noinline void leak() {
     while (true) {
diff --git a/debuggerd/debuggerd_test.cpp b/debuggerd/debuggerd_test.cpp
index 9e9557f..144faee 100644
--- a/debuggerd/debuggerd_test.cpp
+++ b/debuggerd/debuggerd_test.cpp
@@ -34,6 +34,7 @@
 
 #include <android/fdsan.h>
 #include <android/set_abort_message.h>
+#include <bionic/malloc.h>
 #include <bionic/mte.h>
 #include <bionic/reserved_signals.h>
 
@@ -69,7 +70,7 @@
 #define ARCH_SUFFIX ""
 #endif
 
-constexpr char kWaitForGdbKey[] = "debug.debuggerd.wait_for_gdb";
+constexpr char kWaitForDebuggerKey[] = "debug.debuggerd.wait_for_debugger";
 
 #define TIMEOUT(seconds, expr)                                     \
   [&]() {                                                          \
@@ -98,6 +99,13 @@
   ASSERT_MATCH(result,                             \
                R"(#\d\d pc [0-9a-f]+\s+ \S+ (\(offset 0x[0-9a-f]+\) )?\()" frame_name R"(\+)");
 
+// Enable GWP-ASan at the start of this process. GWP-ASan is enabled using
+// process sampling, so we need to ensure we force GWP-ASan on.
+__attribute__((constructor)) static void enable_gwp_asan() {
+  bool force = true;
+  android_mallopt(M_INITIALIZE_GWP_ASAN, &force, sizeof(force));
+}
+
 static void tombstoned_intercept(pid_t target_pid, unique_fd* intercept_fd, unique_fd* output_fd,
                                  InterceptStatus* status, DebuggerdDumpType intercept_type) {
   intercept_fd->reset(socket_local_client(kTombstonedInterceptSocketName,
@@ -157,7 +165,7 @@
 class CrasherTest : public ::testing::Test {
  public:
   pid_t crasher_pid = -1;
-  bool previous_wait_for_gdb;
+  bool previous_wait_for_debugger;
   unique_fd crasher_pipe;
   unique_fd intercept_fd;
 
@@ -178,8 +186,13 @@
 };
 
 CrasherTest::CrasherTest() {
-  previous_wait_for_gdb = android::base::GetBoolProperty(kWaitForGdbKey, false);
-  android::base::SetProperty(kWaitForGdbKey, "0");
+  previous_wait_for_debugger = android::base::GetBoolProperty(kWaitForDebuggerKey, false);
+  android::base::SetProperty(kWaitForDebuggerKey, "0");
+
+  // Clear the old property too, just in case someone's been using it
+  // on this device. (We only document the new name, but we still support
+  // the old name so we don't break anyone's existing setups.)
+  android::base::SetProperty("debug.debuggerd.wait_for_gdb", "0");
 }
 
 CrasherTest::~CrasherTest() {
@@ -189,7 +202,7 @@
     TEMP_FAILURE_RETRY(waitpid(crasher_pid, &status, WUNTRACED));
   }
 
-  android::base::SetProperty(kWaitForGdbKey, previous_wait_for_gdb ? "1" : "0");
+  android::base::SetProperty(kWaitForDebuggerKey, previous_wait_for_debugger ? "1" : "0");
 }
 
 void CrasherTest::StartIntercept(unique_fd* output_fd, DebuggerdDumpType intercept_type) {
@@ -392,7 +405,77 @@
 }
 #endif
 
-TEST_F(CrasherTest, mte_uaf) {
+// Number of iterations required to reliably guarantee a GWP-ASan crash.
+// GWP-ASan's sample rate is not truly nondeterministic, it initialises a
+// thread-local counter at 2*SampleRate, and decrements on each malloc(). Once
+// the counter reaches zero, we provide a sampled allocation. Then, double that
+// figure to allow for left/right allocation alignment, as this is done randomly
+// without bias.
+#define GWP_ASAN_ITERATIONS_TO_ENSURE_CRASH (0x20000)
+
+struct GwpAsanTestParameters {
+  size_t alloc_size;
+  bool free_before_access;
+  int access_offset;
+  std::string cause_needle; // Needle to be found in the "Cause: [GWP-ASan]" line.
+};
+
+struct GwpAsanCrasherTest : CrasherTest, testing::WithParamInterface<GwpAsanTestParameters> {};
+
+GwpAsanTestParameters gwp_asan_tests[] = {
+  {/* alloc_size */ 7, /* free_before_access */ true, /* access_offset */ 0, "Use After Free, 0 bytes into a 7-byte allocation"},
+  {/* alloc_size */ 7, /* free_before_access */ true, /* access_offset */ 1, "Use After Free, 1 byte into a 7-byte allocation"},
+  {/* alloc_size */ 7, /* free_before_access */ false, /* access_offset */ 16, "Buffer Overflow, 9 bytes right of a 7-byte allocation"},
+  {/* alloc_size */ 16, /* free_before_access */ false, /* access_offset */ -1, "Buffer Underflow, 1 byte left of a 16-byte allocation"},
+};
+
+INSTANTIATE_TEST_SUITE_P(GwpAsanTests, GwpAsanCrasherTest, testing::ValuesIn(gwp_asan_tests));
+
+TEST_P(GwpAsanCrasherTest, gwp_asan_uaf) {
+  if (mte_supported()) {
+    // Skip this test on MTE hardware, as MTE will reliably catch these errors
+    // instead of GWP-ASan.
+    GTEST_SKIP() << "Skipped on MTE.";
+  }
+
+  GwpAsanTestParameters params = GetParam();
+
+  int intercept_result;
+  unique_fd output_fd;
+  StartProcess([&params]() {
+    for (unsigned i = 0; i < GWP_ASAN_ITERATIONS_TO_ENSURE_CRASH; ++i) {
+      volatile char* p = reinterpret_cast<volatile char*>(malloc(params.alloc_size));
+      if (params.free_before_access) free(static_cast<void*>(const_cast<char*>(p)));
+      p[params.access_offset] = 42;
+      if (!params.free_before_access) free(static_cast<void*>(const_cast<char*>(p)));
+    }
+  });
+
+  StartIntercept(&output_fd);
+  FinishCrasher();
+  AssertDeath(SIGSEGV);
+  FinishIntercept(&intercept_result);
+
+  ASSERT_EQ(1, intercept_result) << "tombstoned reported failure";
+
+  std::string result;
+  ConsumeFd(std::move(output_fd), &result);
+
+  ASSERT_MATCH(result, R"(signal 11 \(SIGSEGV\), code 2 \(SEGV_ACCERR\))");
+  ASSERT_MATCH(result, R"(Cause: \[GWP-ASan\]: )" + params.cause_needle);
+  if (params.free_before_access) {
+    ASSERT_MATCH(result, R"(deallocated by thread .*
+      #00 pc)");
+  }
+  ASSERT_MATCH(result, R"(allocated by thread .*
+      #00 pc)");
+}
+
+struct SizeParamCrasherTest : CrasherTest, testing::WithParamInterface<size_t> {};
+
+INSTANTIATE_TEST_SUITE_P(Sizes, SizeParamCrasherTest, testing::Values(16, 131072));
+
+TEST_P(SizeParamCrasherTest, mte_uaf) {
 #if defined(__aarch64__)
   if (!mte_supported()) {
     GTEST_SKIP() << "Requires MTE";
@@ -400,9 +483,9 @@
 
   int intercept_result;
   unique_fd output_fd;
-  StartProcess([]() {
+  StartProcess([&]() {
     SetTagCheckingLevelSync();
-    volatile int* p = (volatile int*)malloc(16);
+    volatile int* p = (volatile int*)malloc(GetParam());
     free((void *)p);
     p[0] = 42;
   });
@@ -417,19 +500,19 @@
   std::string result;
   ConsumeFd(std::move(output_fd), &result);
 
-  ASSERT_MATCH(result, R"(signal 11 \(SIGSEGV\), code 9 \(SEGV_MTESERR\))");
-  ASSERT_MATCH(result, R"(Cause: \[MTE\]: Use After Free, 0 bytes into a 16-byte allocation.*
-
-allocated by thread .*
-      #00 pc)");
+  ASSERT_MATCH(result, R"(signal 11 \(SIGSEGV\))");
+  ASSERT_MATCH(result, R"(Cause: \[MTE\]: Use After Free, 0 bytes into a )" +
+                           std::to_string(GetParam()) + R"(-byte allocation)");
   ASSERT_MATCH(result, R"(deallocated by thread .*
       #00 pc)");
+  ASSERT_MATCH(result, R"(allocated by thread .*
+      #00 pc)");
 #else
   GTEST_SKIP() << "Requires aarch64";
 #endif
 }
 
-TEST_F(CrasherTest, mte_overflow) {
+TEST_P(SizeParamCrasherTest, mte_overflow) {
 #if defined(__aarch64__)
   if (!mte_supported()) {
     GTEST_SKIP() << "Requires MTE";
@@ -437,10 +520,10 @@
 
   int intercept_result;
   unique_fd output_fd;
-  StartProcess([]() {
+  StartProcess([&]() {
     SetTagCheckingLevelSync();
-    volatile int* p = (volatile int*)malloc(16);
-    p[4] = 42;
+    volatile char* p = (volatile char*)malloc(GetParam());
+    p[GetParam()] = 42;
   });
 
   StartIntercept(&output_fd);
@@ -454,16 +537,16 @@
   ConsumeFd(std::move(output_fd), &result);
 
   ASSERT_MATCH(result, R"(signal 11 \(SIGSEGV\))");
-  ASSERT_MATCH(result, R"(Cause: \[MTE\]: Buffer Overflow, 0 bytes right of a 16-byte allocation.*
-
-allocated by thread .*
+  ASSERT_MATCH(result, R"(Cause: \[MTE\]: Buffer Overflow, 0 bytes right of a )" +
+                           std::to_string(GetParam()) + R"(-byte allocation)");
+  ASSERT_MATCH(result, R"(allocated by thread .*
       #00 pc)");
 #else
   GTEST_SKIP() << "Requires aarch64";
 #endif
 }
 
-TEST_F(CrasherTest, mte_underflow) {
+TEST_P(SizeParamCrasherTest, mte_underflow) {
 #if defined(__aarch64__)
   if (!mte_supported()) {
     GTEST_SKIP() << "Requires MTE";
@@ -471,9 +554,9 @@
 
   int intercept_result;
   unique_fd output_fd;
-  StartProcess([]() {
+  StartProcess([&]() {
     SetTagCheckingLevelSync();
-    volatile int* p = (volatile int*)malloc(16);
+    volatile int* p = (volatile int*)malloc(GetParam());
     p[-1] = 42;
   });
 
@@ -488,9 +571,9 @@
   ConsumeFd(std::move(output_fd), &result);
 
   ASSERT_MATCH(result, R"(signal 11 \(SIGSEGV\), code 9 \(SEGV_MTESERR\))");
-  ASSERT_MATCH(result, R"(Cause: \[MTE\]: Buffer Underflow, 4 bytes left of a 16-byte allocation.*
-
-allocated by thread .*
+  ASSERT_MATCH(result, R"(Cause: \[MTE\]: Buffer Underflow, 4 bytes left of a )" +
+                           std::to_string(GetParam()) + R"(-byte allocation)");
+  ASSERT_MATCH(result, R"(allocated by thread .*
       #00 pc)");
 #else
   GTEST_SKIP() << "Requires aarch64";
@@ -727,9 +810,9 @@
   AssertDeath(SIGABRT);
 }
 
-TEST_F(CrasherTest, wait_for_gdb) {
-  if (!android::base::SetProperty(kWaitForGdbKey, "1")) {
-    FAIL() << "failed to enable wait_for_gdb";
+TEST_F(CrasherTest, wait_for_debugger) {
+  if (!android::base::SetProperty(kWaitForDebuggerKey, "1")) {
+    FAIL() << "failed to enable wait_for_debugger";
   }
   sleep(1);
 
diff --git a/debuggerd/handler/debuggerd_handler.cpp b/debuggerd/handler/debuggerd_handler.cpp
index ca809e4..b607397 100644
--- a/debuggerd/handler/debuggerd_handler.cpp
+++ b/debuggerd/handler/debuggerd_handler.cpp
@@ -274,7 +274,7 @@
 
     // There appears to be a bug in the kernel where our death causes SIGHUP to
     // be sent to our process group if we exit while it has stopped jobs (e.g.
-    // because of wait_for_gdb). Use setsid to create a new process group to
+    // because of wait_for_debugger). Use setsid to create a new process group to
     // avoid hitting this.
     setsid();
 
@@ -600,7 +600,7 @@
     // starting to dump right before our death.
     pthread_mutex_unlock(&crash_mutex);
   } else {
-    // Resend the signal, so that either gdb or the parent's waitpid sees it.
+    // Resend the signal, so that either the debugger or the parent's waitpid sees it.
     resend_signal(info);
   }
 }
diff --git a/debuggerd/include/debuggerd/handler.h b/debuggerd/include/debuggerd/handler.h
index 254ed4f..bc08327 100644
--- a/debuggerd/include/debuggerd/handler.h
+++ b/debuggerd/include/debuggerd/handler.h
@@ -42,6 +42,7 @@
   const gwp_asan::AllocationMetadata* gwp_asan_metadata;
   const char* scudo_stack_depot;
   const char* scudo_region_info;
+  const char* scudo_ring_buffer;
 };
 
 // These callbacks are called in a signal handler, and thus must be async signal safe.
diff --git a/debuggerd/libdebuggerd/gwp_asan.cpp b/debuggerd/libdebuggerd/gwp_asan.cpp
index 9750fc4..3ee309f 100644
--- a/debuggerd/libdebuggerd/gwp_asan.cpp
+++ b/debuggerd/libdebuggerd/gwp_asan.cpp
@@ -15,6 +15,7 @@
  */
 
 #include "libdebuggerd/gwp_asan.h"
+#include "libdebuggerd/tombstone.h"
 #include "libdebuggerd/utility.h"
 
 #include "gwp_asan/common.h"
@@ -25,6 +26,8 @@
 #include <unwindstack/Regs.h>
 #include <unwindstack/Unwinder.h>
 
+#include "tombstone.pb.h"
+
 // Retrieve GWP-ASan state from `state_addr` inside the process at
 // `process_memory`. Place the state into `*state`.
 static bool retrieve_gwp_asan_state(unwindstack::Memory* process_memory, uintptr_t state_addr,
@@ -98,6 +101,67 @@
   return is_gwp_asan_responsible_;
 }
 
+constexpr size_t kMaxTraceLength = gwp_asan::AllocationMetadata::kMaxTraceLengthToCollect;
+
+void GwpAsanCrashData::AddCauseProtos(Tombstone* tombstone, unwindstack::Unwinder* unwinder) const {
+  if (!CrashIsMine()) {
+    ALOGE("Internal Error: AddCauseProtos() on a non-GWP-ASan crash.");
+    return;
+  }
+
+  Cause* cause = tombstone->add_causes();
+  MemoryError* memory_error = cause->mutable_memory_error();
+  HeapObject* heap_object = memory_error->mutable_heap();
+
+  memory_error->set_tool(MemoryError_Tool_GWP_ASAN);
+  switch (error_) {
+    case gwp_asan::Error::USE_AFTER_FREE:
+      memory_error->set_type(MemoryError_Type_USE_AFTER_FREE);
+      break;
+    case gwp_asan::Error::DOUBLE_FREE:
+      memory_error->set_type(MemoryError_Type_DOUBLE_FREE);
+      break;
+    case gwp_asan::Error::INVALID_FREE:
+      memory_error->set_type(MemoryError_Type_INVALID_FREE);
+      break;
+    case gwp_asan::Error::BUFFER_OVERFLOW:
+      memory_error->set_type(MemoryError_Type_BUFFER_OVERFLOW);
+      break;
+    case gwp_asan::Error::BUFFER_UNDERFLOW:
+      memory_error->set_type(MemoryError_Type_BUFFER_UNDERFLOW);
+      break;
+    default:
+      memory_error->set_type(MemoryError_Type_UNKNOWN);
+      break;
+  }
+
+  heap_object->set_address(__gwp_asan_get_allocation_address(responsible_allocation_));
+  heap_object->set_size(__gwp_asan_get_allocation_size(responsible_allocation_));
+  unwinder->SetDisplayBuildID(true);
+
+  std::unique_ptr<uintptr_t[]> frames(new uintptr_t[kMaxTraceLength]);
+
+  heap_object->set_allocation_tid(__gwp_asan_get_allocation_thread_id(responsible_allocation_));
+  size_t num_frames =
+      __gwp_asan_get_allocation_trace(responsible_allocation_, frames.get(), kMaxTraceLength);
+  for (size_t i = 0; i != num_frames; ++i) {
+    unwindstack::FrameData frame_data = unwinder->BuildFrameFromPcOnly(frames[i]);
+    BacktraceFrame* f = heap_object->add_allocation_backtrace();
+    fill_in_backtrace_frame(f, frame_data, unwinder->GetMaps());
+  }
+
+  heap_object->set_deallocation_tid(__gwp_asan_get_deallocation_thread_id(responsible_allocation_));
+  num_frames =
+      __gwp_asan_get_deallocation_trace(responsible_allocation_, frames.get(), kMaxTraceLength);
+  for (size_t i = 0; i != num_frames; ++i) {
+    unwindstack::FrameData frame_data = unwinder->BuildFrameFromPcOnly(frames[i]);
+    BacktraceFrame* f = heap_object->add_deallocation_backtrace();
+    fill_in_backtrace_frame(f, frame_data, unwinder->GetMaps());
+  }
+
+  set_human_readable_cause(cause, crash_address_);
+}
+
 void GwpAsanCrashData::DumpCause(log_t* log) const {
   if (!CrashIsMine()) {
     ALOGE("Internal Error: DumpCause() on a non-GWP-ASan crash.");
@@ -119,13 +183,6 @@
   uintptr_t alloc_address = __gwp_asan_get_allocation_address(responsible_allocation_);
   size_t alloc_size = __gwp_asan_get_allocation_size(responsible_allocation_);
 
-  if (crash_address_ == alloc_address) {
-    // Use After Free on a 41-byte allocation at 0xdeadbeef.
-    _LOG(log, logtype::HEADER, "Cause: [GWP-ASan]: %s on a %zu-byte allocation at 0x%" PRIxPTR "\n",
-         error_string_, alloc_size, alloc_address);
-    return;
-  }
-
   uintptr_t diff;
   const char* location_str;
 
@@ -157,8 +214,6 @@
        error_string_, diff, byte_suffix, location_str, alloc_size, alloc_address);
 }
 
-constexpr size_t kMaxTraceLength = gwp_asan::AllocationMetadata::kMaxTraceLengthToCollect;
-
 bool GwpAsanCrashData::HasDeallocationTrace() const {
   assert(CrashIsMine() && "HasDeallocationTrace(): Crash is not mine!");
   if (!responsible_allocation_ || !__gwp_asan_is_deallocated(responsible_allocation_)) {
@@ -171,7 +226,7 @@
   assert(HasDeallocationTrace() && "DumpDeallocationTrace(): No dealloc trace!");
   uint64_t thread_id = __gwp_asan_get_deallocation_thread_id(responsible_allocation_);
 
-  std::unique_ptr<uintptr_t> frames(new uintptr_t[kMaxTraceLength]);
+  std::unique_ptr<uintptr_t[]> frames(new uintptr_t[kMaxTraceLength]);
   size_t num_frames =
       __gwp_asan_get_deallocation_trace(responsible_allocation_, frames.get(), kMaxTraceLength);
 
@@ -183,7 +238,7 @@
 
   unwinder->SetDisplayBuildID(true);
   for (size_t i = 0; i < num_frames; ++i) {
-    unwindstack::FrameData frame_data = unwinder->BuildFrameFromPcOnly(frames.get()[i]);
+    unwindstack::FrameData frame_data = unwinder->BuildFrameFromPcOnly(frames[i]);
     frame_data.num = i;
     _LOG(log, logtype::BACKTRACE, "    %s\n", unwinder->FormatFrame(frame_data).c_str());
   }
@@ -198,7 +253,7 @@
   assert(HasAllocationTrace() && "DumpAllocationTrace(): No dealloc trace!");
   uint64_t thread_id = __gwp_asan_get_allocation_thread_id(responsible_allocation_);
 
-  std::unique_ptr<uintptr_t> frames(new uintptr_t[kMaxTraceLength]);
+  std::unique_ptr<uintptr_t[]> frames(new uintptr_t[kMaxTraceLength]);
   size_t num_frames =
       __gwp_asan_get_allocation_trace(responsible_allocation_, frames.get(), kMaxTraceLength);
 
@@ -210,7 +265,7 @@
 
   unwinder->SetDisplayBuildID(true);
   for (size_t i = 0; i < num_frames; ++i) {
-    unwindstack::FrameData frame_data = unwinder->BuildFrameFromPcOnly(frames.get()[i]);
+    unwindstack::FrameData frame_data = unwinder->BuildFrameFromPcOnly(frames[i]);
     frame_data.num = i;
     _LOG(log, logtype::BACKTRACE, "    %s\n", unwinder->FormatFrame(frame_data).c_str());
   }
diff --git a/debuggerd/libdebuggerd/include/libdebuggerd/gwp_asan.h b/debuggerd/libdebuggerd/include/libdebuggerd/gwp_asan.h
index 6c88733..f9c2481 100644
--- a/debuggerd/libdebuggerd/include/libdebuggerd/gwp_asan.h
+++ b/debuggerd/libdebuggerd/include/libdebuggerd/gwp_asan.h
@@ -26,6 +26,9 @@
 #include "types.h"
 #include "utility.h"
 
+class Cause;
+class Tombstone;
+
 class GwpAsanCrashData {
  public:
   GwpAsanCrashData() = delete;
@@ -69,6 +72,8 @@
   // HasAllocationTrace() returns true.
   void DumpAllocationTrace(log_t* log, unwindstack::Unwinder* unwinder) const;
 
+  void AddCauseProtos(Tombstone* tombstone, unwindstack::Unwinder* unwinder) const;
+
  protected:
   // Is GWP-ASan responsible for this crash.
   bool is_gwp_asan_responsible_ = false;
diff --git a/debuggerd/libdebuggerd/include/libdebuggerd/scudo.h b/debuggerd/libdebuggerd/include/libdebuggerd/scudo.h
index 4d00ece..c3b95d6 100644
--- a/debuggerd/libdebuggerd/include/libdebuggerd/scudo.h
+++ b/debuggerd/libdebuggerd/include/libdebuggerd/scudo.h
@@ -23,6 +23,9 @@
 
 #include "scudo/interface.h"
 
+class Cause;
+class Tombstone;
+
 class ScudoCrashData {
  public:
   ScudoCrashData() = delete;
@@ -32,6 +35,7 @@
   bool CrashIsMine() const;
 
   void DumpCause(log_t* log, unwindstack::Unwinder* unwinder) const;
+  void AddCauseProtos(Tombstone* tombstone, unwindstack::Unwinder* unwinder) const;
 
  private:
   scudo_error_info error_info_ = {};
@@ -39,4 +43,7 @@
 
   void DumpReport(const scudo_error_report* report, log_t* log,
                   unwindstack::Unwinder* unwinder) const;
+
+  void FillInCause(Cause* cause, const scudo_error_report* report,
+                   unwindstack::Unwinder* unwinder) const;
 };
diff --git a/debuggerd/libdebuggerd/include/libdebuggerd/tombstone.h b/debuggerd/libdebuggerd/include/libdebuggerd/tombstone.h
index bf2cbb3..2331f1e 100644
--- a/debuggerd/libdebuggerd/include/libdebuggerd/tombstone.h
+++ b/debuggerd/libdebuggerd/include/libdebuggerd/tombstone.h
@@ -31,9 +31,13 @@
 #include "types.h"
 
 // Forward declarations
+class BacktraceFrame;
+class Cause;
 class Tombstone;
 
 namespace unwindstack {
+struct FrameData;
+class Maps;
 class Unwinder;
 }
 
@@ -64,4 +68,8 @@
     const Tombstone& tombstone,
     std::function<void(const std::string& line, bool should_log)> callback);
 
+void fill_in_backtrace_frame(BacktraceFrame* f, const unwindstack::FrameData& frame,
+                             unwindstack::Maps* maps);
+void set_human_readable_cause(Cause* cause, uint64_t fault_addr);
+
 #endif  // _DEBUGGERD_TOMBSTONE_H
diff --git a/debuggerd/libdebuggerd/include/libdebuggerd/types.h b/debuggerd/libdebuggerd/include/libdebuggerd/types.h
index d5b0735..dcb52f9 100644
--- a/debuggerd/libdebuggerd/include/libdebuggerd/types.h
+++ b/debuggerd/libdebuggerd/include/libdebuggerd/types.h
@@ -46,6 +46,7 @@
   uintptr_t gwp_asan_metadata = 0;
   uintptr_t scudo_stack_depot = 0;
   uintptr_t scudo_region_info = 0;
+  uintptr_t scudo_ring_buffer = 0;
 
   bool has_fault_address = false;
   uintptr_t untagged_fault_address = 0;
diff --git a/debuggerd/libdebuggerd/include/libdebuggerd/utility.h b/debuggerd/libdebuggerd/include/libdebuggerd/utility.h
index d71b76f..c490fb1 100644
--- a/debuggerd/libdebuggerd/include/libdebuggerd/utility.h
+++ b/debuggerd/libdebuggerd/include/libdebuggerd/utility.h
@@ -81,7 +81,7 @@
 
 void log_backtrace(log_t* log, unwindstack::Unwinder* unwinder, const char* prefix);
 
-ssize_t dump_memory(void* out, size_t len, size_t* start_offset, uint64_t* addr,
+ssize_t dump_memory(void* out, size_t len, uint8_t* tags, size_t tags_len, uint64_t* addr,
                     unwindstack::Memory* memory);
 void dump_memory(log_t* log, unwindstack::Memory* backtrace, uint64_t addr, const std::string&);
 
@@ -93,4 +93,7 @@
 const char* get_signame(const siginfo_t*);
 const char* get_sigcode(const siginfo_t*);
 
+// Number of bytes per MTE granule.
+constexpr size_t kTagGranuleSize = 16;
+
 #endif // _DEBUGGERD_UTILITY_H
diff --git a/debuggerd/libdebuggerd/scudo.cpp b/debuggerd/libdebuggerd/scudo.cpp
index 141c3bd..f4690ba 100644
--- a/debuggerd/libdebuggerd/scudo.cpp
+++ b/debuggerd/libdebuggerd/scudo.cpp
@@ -15,13 +15,16 @@
  */
 
 #include "libdebuggerd/scudo.h"
-#include "libdebuggerd/gwp_asan.h"
+#include "libdebuggerd/tombstone.h"
 
 #include "unwindstack/Memory.h"
 #include "unwindstack/Unwinder.h"
 
+#include <android-base/macros.h>
 #include <bionic/macros.h>
 
+#include "tombstone.pb.h"
+
 std::unique_ptr<char[]> AllocAndReadFully(unwindstack::Memory* process_memory, uint64_t addr,
                                           size_t size) {
   auto buf = std::make_unique<char[]>(size);
@@ -31,8 +34,6 @@
   return buf;
 }
 
-static const uintptr_t kTagGranuleSize = 16;
-
 ScudoCrashData::ScudoCrashData(unwindstack::Memory* process_memory,
                                const ProcessInfo& process_info) {
   if (!process_info.has_fault_address) {
@@ -43,6 +44,8 @@
                                        __scudo_get_stack_depot_size());
   auto region_info = AllocAndReadFully(process_memory, process_info.scudo_region_info,
                                        __scudo_get_region_info_size());
+  auto ring_buffer = AllocAndReadFully(process_memory, process_info.scudo_ring_buffer,
+                                       __scudo_get_ring_buffer_size());
 
   untagged_fault_addr_ = process_info.untagged_fault_address;
   uintptr_t fault_page = untagged_fault_addr_ & ~(PAGE_SIZE - 1);
@@ -68,14 +71,66 @@
   }
 
   __scudo_get_error_info(&error_info_, process_info.maybe_tagged_fault_address, stack_depot.get(),
-                         region_info.get(), memory.get(), memory_tags.get(), memory_begin,
-                         memory_end - memory_begin);
+                         region_info.get(), ring_buffer.get(), memory.get(), memory_tags.get(),
+                         memory_begin, memory_end - memory_begin);
 }
 
 bool ScudoCrashData::CrashIsMine() const {
   return error_info_.reports[0].error_type != UNKNOWN;
 }
 
+void ScudoCrashData::FillInCause(Cause* cause, const scudo_error_report* report,
+                                 unwindstack::Unwinder* unwinder) const {
+  MemoryError* memory_error = cause->mutable_memory_error();
+  HeapObject* heap_object = memory_error->mutable_heap();
+
+  memory_error->set_tool(MemoryError_Tool_SCUDO);
+  switch (report->error_type) {
+    case USE_AFTER_FREE:
+      memory_error->set_type(MemoryError_Type_USE_AFTER_FREE);
+      break;
+    case BUFFER_OVERFLOW:
+      memory_error->set_type(MemoryError_Type_BUFFER_OVERFLOW);
+      break;
+    case BUFFER_UNDERFLOW:
+      memory_error->set_type(MemoryError_Type_BUFFER_UNDERFLOW);
+      break;
+    default:
+      memory_error->set_type(MemoryError_Type_UNKNOWN);
+      break;
+  }
+
+  heap_object->set_address(report->allocation_address);
+  heap_object->set_size(report->allocation_size);
+  unwinder->SetDisplayBuildID(true);
+
+  heap_object->set_allocation_tid(report->allocation_tid);
+  for (size_t i = 0; i < arraysize(report->allocation_trace) && report->allocation_trace[i]; ++i) {
+    unwindstack::FrameData frame_data = unwinder->BuildFrameFromPcOnly(report->allocation_trace[i]);
+    BacktraceFrame* f = heap_object->add_allocation_backtrace();
+    fill_in_backtrace_frame(f, frame_data, unwinder->GetMaps());
+  }
+
+  heap_object->set_deallocation_tid(report->deallocation_tid);
+  for (size_t i = 0; i < arraysize(report->deallocation_trace) && report->deallocation_trace[i];
+       ++i) {
+    unwindstack::FrameData frame_data =
+        unwinder->BuildFrameFromPcOnly(report->deallocation_trace[i]);
+    BacktraceFrame* f = heap_object->add_deallocation_backtrace();
+    fill_in_backtrace_frame(f, frame_data, unwinder->GetMaps());
+  }
+
+  set_human_readable_cause(cause, untagged_fault_addr_);
+}
+
+void ScudoCrashData::AddCauseProtos(Tombstone* tombstone, unwindstack::Unwinder* unwinder) const {
+  size_t report_num = 0;
+  while (report_num < sizeof(error_info_.reports) / sizeof(error_info_.reports[0]) &&
+         error_info_.reports[report_num].error_type != UNKNOWN) {
+    FillInCause(tombstone->add_causes(), &error_info_.reports[report_num++], unwinder);
+  }
+}
+
 void ScudoCrashData::DumpCause(log_t* log, unwindstack::Unwinder* unwinder) const {
   if (error_info_.reports[1].error_type != UNKNOWN) {
     _LOG(log, logtype::HEADER,
@@ -138,7 +193,8 @@
   if (report->allocation_trace[0]) {
     _LOG(log, logtype::BACKTRACE, "\nallocated by thread %u:\n", report->allocation_tid);
     unwinder->SetDisplayBuildID(true);
-    for (size_t i = 0; i < 64 && report->allocation_trace[i]; ++i) {
+    for (size_t i = 0; i < arraysize(report->allocation_trace) && report->allocation_trace[i];
+         ++i) {
       unwindstack::FrameData frame_data =
           unwinder->BuildFrameFromPcOnly(report->allocation_trace[i]);
       frame_data.num = i;
@@ -149,7 +205,8 @@
   if (report->deallocation_trace[0]) {
     _LOG(log, logtype::BACKTRACE, "\ndeallocated by thread %u:\n", report->deallocation_tid);
     unwinder->SetDisplayBuildID(true);
-    for (size_t i = 0; i < 64 && report->deallocation_trace[i]; ++i) {
+    for (size_t i = 0; i < arraysize(report->deallocation_trace) && report->deallocation_trace[i];
+         ++i) {
       unwindstack::FrameData frame_data =
           unwinder->BuildFrameFromPcOnly(report->deallocation_trace[i]);
       frame_data.num = i;
diff --git a/debuggerd/libdebuggerd/test/dump_memory_test.cpp b/debuggerd/libdebuggerd/test/dump_memory_test.cpp
index f16f578..5be145a 100644
--- a/debuggerd/libdebuggerd/test/dump_memory_test.cpp
+++ b/debuggerd/libdebuggerd/test/dump_memory_test.cpp
@@ -73,34 +73,14 @@
 "    0000000012345600 2726252423222120 2f2e2d2c2b2a2928   !\"#$%&'()*+,-./\n"
 "    0000000012345610 3736353433323130 3f3e3d3c3b3a3938  0123456789:;<=>?\n"
 "    0000000012345620 4746454443424140 4f4e4d4c4b4a4948  @ABCDEFGHIJKLMNO\n"
-"    0000000012345630 5756555453525150 5f5e5d5c5b5a5958  PQRSTUVWXYZ[\\]^_\n"
-"    0000000012345640 6766656463626160 ----------------  `abcdefg........\n"
-"    0000000012345650 ---------------- ----------------  ................\n"
-"    0000000012345660 ---------------- ----------------  ................\n"
-"    0000000012345670 ---------------- ----------------  ................\n"
-"    0000000012345680 ---------------- ----------------  ................\n"
-"    0000000012345690 ---------------- ----------------  ................\n"
-"    00000000123456a0 ---------------- ----------------  ................\n"
-"    00000000123456b0 ---------------- ----------------  ................\n"
-"    00000000123456c0 ---------------- ----------------  ................\n"
-"    00000000123456d0 ---------------- ----------------  ................\n";
+"    0000000012345630 5756555453525150 5f5e5d5c5b5a5958  PQRSTUVWXYZ[\\]^_\n";
 #else
 "    123455e0 03020100 07060504 0b0a0908 0f0e0d0c  ................\n"
 "    123455f0 13121110 17161514 1b1a1918 1f1e1d1c  ................\n"
 "    12345600 23222120 27262524 2b2a2928 2f2e2d2c   !\"#$%&'()*+,-./\n"
 "    12345610 33323130 37363534 3b3a3938 3f3e3d3c  0123456789:;<=>?\n"
 "    12345620 43424140 47464544 4b4a4948 4f4e4d4c  @ABCDEFGHIJKLMNO\n"
-"    12345630 53525150 57565554 5b5a5958 5f5e5d5c  PQRSTUVWXYZ[\\]^_\n"
-"    12345640 63626160 67666564 -------- --------  `abcdefg........\n"
-"    12345650 -------- -------- -------- --------  ................\n"
-"    12345660 -------- -------- -------- --------  ................\n"
-"    12345670 -------- -------- -------- --------  ................\n"
-"    12345680 -------- -------- -------- --------  ................\n"
-"    12345690 -------- -------- -------- --------  ................\n"
-"    123456a0 -------- -------- -------- --------  ................\n"
-"    123456b0 -------- -------- -------- --------  ................\n"
-"    123456c0 -------- -------- -------- --------  ................\n"
-"    123456d0 -------- -------- -------- --------  ................\n";
+"    12345630 53525150 57565554 5b5a5958 5f5e5d5c  PQRSTUVWXYZ[\\]^_\n";
 #endif
 
 class MemoryMock : public unwindstack::Memory {
@@ -513,15 +493,7 @@
   const char* expected_dump = \
 "\nmemory near r4:\n"
 #if defined(__LP64__)
-R"(    0000000010000f80 ---------------- ----------------  ................
-    0000000010000f90 ---------------- ----------------  ................
-    0000000010000fa0 ---------------- ----------------  ................
-    0000000010000fb0 ---------------- ----------------  ................
-    0000000010000fc0 ---------------- ----------------  ................
-    0000000010000fd0 ---------------- ----------------  ................
-    0000000010000fe0 ---------------- ----------------  ................
-    0000000010000ff0 ---------------- ----------------  ................
-    0000000010001000 8786858483828180 8f8e8d8c8b8a8988  ................
+R"(    0000000010001000 8786858483828180 8f8e8d8c8b8a8988  ................
     0000000010001010 9796959493929190 9f9e9d9c9b9a9998  ................
     0000000010001020 a7a6a5a4a3a2a1a0 afaeadacabaaa9a8  ................
     0000000010001030 b7b6b5b4b3b2b1b0 bfbebdbcbbbab9b8  ................
@@ -531,15 +503,7 @@
     0000000010001070 f7f6f5f4f3f2f1f0 fffefdfcfbfaf9f8  ................
 )";
 #else
-R"(    10000f80 -------- -------- -------- --------  ................
-    10000f90 -------- -------- -------- --------  ................
-    10000fa0 -------- -------- -------- --------  ................
-    10000fb0 -------- -------- -------- --------  ................
-    10000fc0 -------- -------- -------- --------  ................
-    10000fd0 -------- -------- -------- --------  ................
-    10000fe0 -------- -------- -------- --------  ................
-    10000ff0 -------- -------- -------- --------  ................
-    10001000 83828180 87868584 8b8a8988 8f8e8d8c  ................
+R"(    10001000 83828180 87868584 8b8a8988 8f8e8d8c  ................
     10001010 93929190 97969594 9b9a9998 9f9e9d9c  ................
     10001020 a3a2a1a0 a7a6a5a4 abaaa9a8 afaeadac  ................
     10001030 b3b2b1b0 b7b6b5b4 bbbab9b8 bfbebdbc  ................
@@ -574,39 +538,11 @@
   const char* expected_dump = \
 "\nmemory near r4:\n"
 #if defined(__LP64__)
-"    0000000010000f40 ---------------- ----------------  ................\n"
-"    0000000010000f50 ---------------- ----------------  ................\n"
-"    0000000010000f60 ---------------- ----------------  ................\n"
-"    0000000010000f70 ---------------- ----------------  ................\n"
-"    0000000010000f80 ---------------- ----------------  ................\n"
-"    0000000010000f90 ---------------- ----------------  ................\n"
-"    0000000010000fa0 ---------------- ----------------  ................\n"
-"    0000000010000fb0 ---------------- ----------------  ................\n"
-"    0000000010000fc0 ---------------- ----------------  ................\n"
-"    0000000010000fd0 ---------------- ----------------  ................\n"
-"    0000000010000fe0 ---------------- ----------------  ................\n"
-"    0000000010000ff0 ---------------- ----------------  ................\n"
 "    0000000010001000 c7c6c5c4c3c2c1c0 cfcecdcccbcac9c8  ................\n"
-"    0000000010001010 d7d6d5d4d3d2d1d0 dfdedddcdbdad9d8  ................\n"
-"    0000000010001020 ---------------- ----------------  ................\n"
-"    0000000010001030 ---------------- ----------------  ................\n";
+"    0000000010001010 d7d6d5d4d3d2d1d0 dfdedddcdbdad9d8  ................\n";
 #else
-"    10000f40 -------- -------- -------- --------  ................\n"
-"    10000f50 -------- -------- -------- --------  ................\n"
-"    10000f60 -------- -------- -------- --------  ................\n"
-"    10000f70 -------- -------- -------- --------  ................\n"
-"    10000f80 -------- -------- -------- --------  ................\n"
-"    10000f90 -------- -------- -------- --------  ................\n"
-"    10000fa0 -------- -------- -------- --------  ................\n"
-"    10000fb0 -------- -------- -------- --------  ................\n"
-"    10000fc0 -------- -------- -------- --------  ................\n"
-"    10000fd0 -------- -------- -------- --------  ................\n"
-"    10000fe0 -------- -------- -------- --------  ................\n"
-"    10000ff0 -------- -------- -------- --------  ................\n"
 "    10001000 c3c2c1c0 c7c6c5c4 cbcac9c8 cfcecdcc  ................\n"
-"    10001010 d3d2d1d0 d7d6d5d4 dbdad9d8 dfdedddc  ................\n"
-"    10001020 -------- -------- -------- --------  ................\n"
-"    10001030 -------- -------- -------- --------  ................\n";
+"    10001010 d3d2d1d0 d7d6d5d4 dbdad9d8 dfdedddc  ................\n";
 #endif
   ASSERT_STREQ(expected_dump, tombstone_contents.c_str());
 
diff --git a/debuggerd/libdebuggerd/test/sys/system_properties.h b/debuggerd/libdebuggerd/test/sys/system_properties.h
deleted file mode 100644
index 1f4f58a..0000000
--- a/debuggerd/libdebuggerd/test/sys/system_properties.h
+++ /dev/null
@@ -1,37 +0,0 @@
-/*
- * Copyright (C) 2015 The Android Open Source Project
- * All rights reserved.
- *
- * Redistribution and use in source and binary forms, with or without
- * modification, are permitted provided that the following conditions
- * are met:
- *  * Redistributions of source code must retain the above copyright
- *    notice, this list of conditions and the following disclaimer.
- *  * Redistributions in binary form must reproduce the above copyright
- *    notice, this list of conditions and the following disclaimer in
- *    the documentation and/or other materials provided with the
- *    distribution.
- *
- * THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
- * "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
- * LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
- * FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
- * COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
- * INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
- * BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
- * OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
- * AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
- * OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
- * OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
- * SUCH DAMAGE.
- */
-
-#ifndef _DEBUGGERD_TEST_SYS_SYSTEM_PROPERTIES_H
-#define _DEBUGGERD_TEST_SYS_SYSTEM_PROPERTIES_H
-
-// This is just enough to get the property code to compile on
-// the host.
-
-#define PROP_VALUE_MAX  92
-
-#endif // _DEBUGGERD_TEST_SYS_SYSTEM_PROPERTIES_H
diff --git a/debuggerd/libdebuggerd/test/tombstone_test.cpp b/debuggerd/libdebuggerd/test/tombstone_test.cpp
index 7fe8f82..79ac122 100644
--- a/debuggerd/libdebuggerd/test/tombstone_test.cpp
+++ b/debuggerd/libdebuggerd/test/tombstone_test.cpp
@@ -388,9 +388,8 @@
   ASSERT_TRUE(lseek(log_.tfd, 0, SEEK_SET) == 0);
   std::string tombstone_contents;
   ASSERT_TRUE(android::base::ReadFdToString(log_.tfd, &tombstone_contents));
-  ASSERT_THAT(tombstone_contents,
-              MatchesRegex("Cause: \\[GWP-ASan\\]: Use After Free on a 32-byte "
-                           "allocation at 0x[a-fA-F0-9]+\n"));
+  ASSERT_THAT(tombstone_contents, MatchesRegex("Cause: \\[GWP-ASan\\]: Use After Free, 0 bytes "
+                                               "into a 32-byte allocation at 0x[a-fA-F0-9]+\n"));
 }
 
 TEST_F(TombstoneTest, gwp_asan_cause_double_free) {
@@ -405,9 +404,8 @@
   ASSERT_TRUE(lseek(log_.tfd, 0, SEEK_SET) == 0);
   std::string tombstone_contents;
   ASSERT_TRUE(android::base::ReadFdToString(log_.tfd, &tombstone_contents));
-  ASSERT_THAT(tombstone_contents,
-              MatchesRegex("Cause: \\[GWP-ASan\\]: Double Free on a 32-byte "
-                           "allocation at 0x[a-fA-F0-9]+\n"));
+  ASSERT_THAT(tombstone_contents, MatchesRegex("Cause: \\[GWP-ASan\\]: Double Free, 0 bytes into a "
+                                               "32-byte allocation at 0x[a-fA-F0-9]+\n"));
 }
 
 TEST_F(TombstoneTest, gwp_asan_cause_overflow) {
diff --git a/debuggerd/libdebuggerd/tombstone_proto.cpp b/debuggerd/libdebuggerd/tombstone_proto.cpp
index 23ca070..3444e29 100644
--- a/debuggerd/libdebuggerd/tombstone_proto.cpp
+++ b/debuggerd/libdebuggerd/tombstone_proto.cpp
@@ -17,6 +17,8 @@
 #define LOG_TAG "DEBUG"
 
 #include "libdebuggerd/tombstone.h"
+#include "libdebuggerd/gwp_asan.h"
+#include "libdebuggerd/scudo.h"
 
 #include <errno.h>
 #include <fcntl.h>
@@ -29,10 +31,12 @@
 #include <time.h>
 
 #include <memory>
+#include <optional>
 #include <string>
 
 #include <async_safe/log.h>
 
+#include <android-base/file.h>
 #include <android-base/properties.h>
 #include <android-base/stringprintf.h>
 #include <android-base/strings.h>
@@ -45,6 +49,7 @@
 #include <log/logprint.h>
 #include <private/android_filesystem_config.h>
 
+#include <procinfo/process.h>
 #include <unwindstack/Maps.h>
 #include <unwindstack/Memory.h>
 #include <unwindstack/Regs.h>
@@ -106,32 +111,120 @@
   return {};
 }
 
-static void dump_probable_cause(Tombstone* tombstone, const siginfo_t* si, unwindstack::Maps* maps,
-                                unwindstack::Regs* regs) {
+void set_human_readable_cause(Cause* cause, uint64_t fault_addr) {
+  if (!cause->has_memory_error() || !cause->memory_error().has_heap()) {
+    return;
+  }
+
+  const MemoryError& memory_error = cause->memory_error();
+  const HeapObject& heap_object = memory_error.heap();
+
+  const char *tool_str;
+  switch (memory_error.tool()) {
+    case MemoryError_Tool_GWP_ASAN:
+      tool_str = "GWP-ASan";
+      break;
+    case MemoryError_Tool_SCUDO:
+      tool_str = "MTE";
+      break;
+    default:
+      tool_str = "Unknown";
+      break;
+  }
+
+  const char *error_type_str;
+  switch (memory_error.type()) {
+    case MemoryError_Type_USE_AFTER_FREE:
+      error_type_str = "Use After Free";
+      break;
+    case MemoryError_Type_DOUBLE_FREE:
+      error_type_str = "Double Free";
+      break;
+    case MemoryError_Type_INVALID_FREE:
+      error_type_str = "Invalid (Wild) Free";
+      break;
+    case MemoryError_Type_BUFFER_OVERFLOW:
+      error_type_str = "Buffer Overflow";
+      break;
+    case MemoryError_Type_BUFFER_UNDERFLOW:
+      error_type_str = "Buffer Underflow";
+      break;
+    default:
+      cause->set_human_readable(
+          StringPrintf("[%s]: Unknown error occurred at 0x%" PRIx64 ".", tool_str, fault_addr));
+      return;
+  }
+
+  uint64_t diff;
+  const char* location_str;
+
+  if (fault_addr < heap_object.address()) {
+    // Buffer Underflow, 6 bytes left of a 41-byte allocation at 0xdeadbeef.
+    location_str = "left of";
+    diff = heap_object.address() - fault_addr;
+  } else if (fault_addr - heap_object.address() < heap_object.size()) {
+    // Use After Free, 40 bytes into a 41-byte allocation at 0xdeadbeef.
+    location_str = "into";
+    diff = fault_addr - heap_object.address();
+  } else {
+    // Buffer Overflow, 6 bytes right of a 41-byte allocation at 0xdeadbeef.
+    location_str = "right of";
+    diff = fault_addr - heap_object.address() - heap_object.size();
+  }
+
+  // Suffix of 'bytes', i.e. 4 bytes' vs. '1 byte'.
+  const char* byte_suffix = "s";
+  if (diff == 1) {
+    byte_suffix = "";
+  }
+
+  cause->set_human_readable(StringPrintf(
+      "[%s]: %s, %" PRIu64 " byte%s %s a %" PRIu64 "-byte allocation at 0x%" PRIx64, tool_str,
+      error_type_str, diff, byte_suffix, location_str, heap_object.size(), heap_object.address()));
+}
+
+static void dump_probable_cause(Tombstone* tombstone, unwindstack::Unwinder* unwinder,
+                                const ProcessInfo& process_info, const ThreadInfo& main_thread) {
+  ScudoCrashData scudo_crash_data(unwinder->GetProcessMemory().get(), process_info);
+  if (scudo_crash_data.CrashIsMine()) {
+    scudo_crash_data.AddCauseProtos(tombstone, unwinder);
+    return;
+  }
+
+  GwpAsanCrashData gwp_asan_crash_data(unwinder->GetProcessMemory().get(), process_info,
+                                       main_thread);
+  if (gwp_asan_crash_data.CrashIsMine()) {
+    gwp_asan_crash_data.AddCauseProtos(tombstone, unwinder);
+    return;
+  }
+
+  const siginfo *si = main_thread.siginfo;
+  auto fault_addr = reinterpret_cast<uint64_t>(si->si_addr);
+  unwindstack::Maps* maps = unwinder->GetMaps();
+
   std::optional<std::string> cause;
   if (si->si_signo == SIGSEGV && si->si_code == SEGV_MAPERR) {
-    if (si->si_addr < reinterpret_cast<void*>(4096)) {
+    if (fault_addr < 4096) {
       cause = "null pointer dereference";
-    } else if (si->si_addr == reinterpret_cast<void*>(0xffff0ffc)) {
+    } else if (fault_addr == 0xffff0ffc) {
       cause = "call to kuser_helper_version";
-    } else if (si->si_addr == reinterpret_cast<void*>(0xffff0fe0)) {
+    } else if (fault_addr == 0xffff0fe0) {
       cause = "call to kuser_get_tls";
-    } else if (si->si_addr == reinterpret_cast<void*>(0xffff0fc0)) {
+    } else if (fault_addr == 0xffff0fc0) {
       cause = "call to kuser_cmpxchg";
-    } else if (si->si_addr == reinterpret_cast<void*>(0xffff0fa0)) {
+    } else if (fault_addr == 0xffff0fa0) {
       cause = "call to kuser_memory_barrier";
-    } else if (si->si_addr == reinterpret_cast<void*>(0xffff0f60)) {
+    } else if (fault_addr == 0xffff0f60) {
       cause = "call to kuser_cmpxchg64";
     } else {
-      cause = get_stack_overflow_cause(reinterpret_cast<uint64_t>(si->si_addr), regs->sp(), maps);
+      cause = get_stack_overflow_cause(fault_addr, main_thread.registers->sp(), maps);
     }
   } else if (si->si_signo == SIGSEGV && si->si_code == SEGV_ACCERR) {
-    uint64_t fault_addr = reinterpret_cast<uint64_t>(si->si_addr);
     unwindstack::MapInfo* map_info = maps->Find(fault_addr);
     if (map_info != nullptr && map_info->flags == PROT_EXEC) {
       cause = "execute-only (no-read) memory access error; likely due to data in .text.";
     } else {
-      cause = get_stack_overflow_cause(fault_addr, regs->sp(), maps);
+      cause = get_stack_overflow_cause(fault_addr, main_thread.registers->sp(), maps);
     }
   } else if (si->si_signo == SIGSYS && si->si_code == SYS_SECCOMP) {
     cause = StringPrintf("seccomp prevented call to disallowed %s system call %d", ABI_STRING,
@@ -139,7 +232,8 @@
   }
 
   if (cause) {
-    tombstone->mutable_cause()->set_human_readable(*cause);
+    Cause *cause_proto = tombstone->add_causes();
+    cause_proto->set_human_readable(*cause);
   }
 }
 
@@ -205,12 +299,49 @@
   }
 }
 
+void fill_in_backtrace_frame(BacktraceFrame* f, const unwindstack::FrameData& frame,
+                             unwindstack::Maps* maps) {
+  f->set_rel_pc(frame.rel_pc);
+  f->set_pc(frame.pc);
+  f->set_sp(frame.sp);
+
+  if (!frame.function_name.empty()) {
+    // TODO: Should this happen here, or on the display side?
+    char* demangled_name = __cxa_demangle(frame.function_name.c_str(), nullptr, nullptr, nullptr);
+    if (demangled_name) {
+      f->set_function_name(demangled_name);
+      free(demangled_name);
+    } else {
+      f->set_function_name(frame.function_name);
+    }
+  }
+
+  f->set_function_offset(frame.function_offset);
+
+  if (frame.map_start == frame.map_end) {
+    // No valid map associated with this frame.
+    f->set_file_name("<unknown>");
+  } else if (!frame.map_name.empty()) {
+    f->set_file_name(frame.map_name);
+  } else {
+    f->set_file_name(StringPrintf("<anonymous:%" PRIx64 ">", frame.map_start));
+  }
+
+  f->set_file_map_offset(frame.map_elf_start_offset);
+
+  unwindstack::MapInfo* map_info = maps->Find(frame.map_start);
+  if (map_info) {
+    f->set_build_id(map_info->GetPrintableBuildID());
+  }
+}
+
 static void dump_thread(Tombstone* tombstone, unwindstack::Unwinder* unwinder,
                         const ThreadInfo& thread_info, bool memory_dump = false) {
   Thread thread;
 
   thread.set_id(thread_info.tid);
   thread.set_name(thread_info.thread_name);
+  thread.set_tagged_addr_ctrl(thread_info.tagged_addr_ctrl);
 
   unwindstack::Maps* maps = unwinder->GetMaps();
   unwindstack::Memory* memory = unwinder->GetProcessMemory().get();
@@ -225,20 +356,19 @@
         if (memory_dump) {
           MemoryDump dump;
 
-          char buf[256];
-          size_t start_offset = 0;
-          ssize_t bytes = dump_memory(buf, sizeof(buf), &start_offset, &value, memory);
-          if (bytes == -1) {
-            return;
-          }
-
           dump.set_register_name(name);
-
           unwindstack::MapInfo* map_info = maps->Find(untag_address(value));
           if (map_info) {
             dump.set_mapping_name(map_info->name);
           }
 
+          char buf[256];
+          uint8_t tags[256 / kTagGranuleSize];
+          size_t start_offset = 0;
+          ssize_t bytes = dump_memory(buf, sizeof(buf), tags, sizeof(tags), &value, memory);
+          if (bytes == -1) {
+            return;
+          }
           dump.set_begin_address(value);
 
           if (start_offset + bytes > sizeof(buf)) {
@@ -246,7 +376,8 @@
                              start_offset, bytes);
           }
 
-          dump.set_memory(buf, start_offset + bytes);
+          dump.set_memory(buf, bytes);
+          dump.set_tags(tags, bytes / kTagGranuleSize);
 
           *thread.add_memory_dump() = std::move(dump);
         }
@@ -267,39 +398,7 @@
     unwinder->SetDisplayBuildID(true);
     for (const auto& frame : unwinder->frames()) {
       BacktraceFrame* f = thread.add_current_backtrace();
-      f->set_rel_pc(frame.rel_pc);
-      f->set_pc(frame.pc);
-      f->set_sp(frame.sp);
-
-      if (!frame.function_name.empty()) {
-        // TODO: Should this happen here, or on the display side?
-        char* demangled_name =
-            __cxa_demangle(frame.function_name.c_str(), nullptr, nullptr, nullptr);
-        if (demangled_name) {
-          f->set_function_name(demangled_name);
-          free(demangled_name);
-        } else {
-          f->set_function_name(frame.function_name);
-        }
-      }
-
-      f->set_function_offset(frame.function_offset);
-
-      if (frame.map_start == frame.map_end) {
-        // No valid map associated with this frame.
-        f->set_file_name("<unknown>");
-      } else if (!frame.map_name.empty()) {
-        f->set_file_name(frame.map_name);
-      } else {
-        f->set_file_name(StringPrintf("<anonymous:%" PRIx64 ">", frame.map_start));
-      }
-
-      f->set_file_map_offset(frame.map_elf_start_offset);
-
-      unwindstack::MapInfo* map_info = maps->Find(frame.map_start);
-      if (map_info) {
-        f->set_build_id(map_info->GetPrintableBuildID());
-      }
+      fill_in_backtrace_frame(f, frame, maps);
     }
   }
 
@@ -423,6 +522,14 @@
   dump_log_file(tombstone, "main", pid);
 }
 
+static std::optional<uint64_t> read_uptime_secs() {
+  std::string uptime;
+  if (!android::base::ReadFileToString("/proc/uptime", &uptime)) {
+    return {};
+  }
+  return strtoll(uptime.c_str(), nullptr, 10);
+}
+
 void engrave_tombstone_proto(Tombstone* tombstone, unwindstack::Unwinder* unwinder,
                              const std::map<pid_t, ThreadInfo>& threads, pid_t target_thread,
                              const ProcessInfo& process_info, const OpenFilesList* open_files) {
@@ -433,6 +540,22 @@
   result.set_revision(android::base::GetProperty("ro.revision", "unknown"));
   result.set_timestamp(get_timestamp());
 
+  std::optional<uint64_t> system_uptime = read_uptime_secs();
+  if (system_uptime) {
+    android::procinfo::ProcessInfo proc_info;
+    std::string error;
+    if (android::procinfo::GetProcessInfo(target_thread, &proc_info, &error)) {
+      uint64_t starttime = proc_info.starttime / sysconf(_SC_CLK_TCK);
+      result.set_process_uptime(*system_uptime - starttime);
+    } else {
+      async_safe_format_log(ANDROID_LOG_ERROR, LOG_TAG, "failed to read process info: %s",
+                            error.c_str());
+    }
+  } else {
+    async_safe_format_log(ANDROID_LOG_ERROR, LOG_TAG, "failed to read /proc/uptime: %s",
+                          strerror(errno));
+  }
+
   const ThreadInfo& main_thread = threads.at(target_thread);
   result.set_pid(main_thread.pid);
   result.set_tid(main_thread.tid);
@@ -458,7 +581,7 @@
 
   if (process_info.has_fault_address) {
     sig.set_has_fault_address(true);
-    sig.set_fault_address(process_info.untagged_fault_address);
+    sig.set_fault_address(process_info.maybe_tagged_fault_address);
   }
 
   *result.mutable_signal_info() = sig;
@@ -473,8 +596,7 @@
     }
   }
 
-  dump_probable_cause(&result, main_thread.siginfo, unwinder->GetMaps(),
-                      main_thread.registers.get());
+  dump_probable_cause(&result, unwinder, process_info, main_thread);
 
   dump_mappings(&result, unwinder);
 
diff --git a/debuggerd/libdebuggerd/tombstone_proto_to_text.cpp b/debuggerd/libdebuggerd/tombstone_proto_to_text.cpp
index 187379d..00ca7c1 100644
--- a/debuggerd/libdebuggerd/tombstone_proto_to_text.cpp
+++ b/debuggerd/libdebuggerd/tombstone_proto_to_text.cpp
@@ -74,6 +74,9 @@
   CB(should_log, "pid: %d, tid: %d, name: %s  >>> %s <<<", tombstone.pid(), thread.id(),
      thread.name().c_str(), tombstone.process_name().c_str());
   CB(should_log, "uid: %d", tombstone.uid());
+  if (thread.tagged_addr_ctrl() != -1) {
+    CB(should_log, "tagged_addr_ctrl: %016" PRIx64, thread.tagged_addr_ctrl());
+  }
 }
 
 static void print_register_row(CallbackType callback, int word_size,
@@ -136,12 +139,11 @@
   print_register_row(callback, word_size, special_row, should_log);
 }
 
-static void print_thread_backtrace(CallbackType callback, const Tombstone& tombstone,
-                                   const Thread& thread, bool should_log) {
-  CBS("");
-  CB(should_log, "backtrace:");
+static void print_backtrace(CallbackType callback, const Tombstone& tombstone,
+                            const google::protobuf::RepeatedPtrField<BacktraceFrame>& backtrace,
+                            bool should_log) {
   int index = 0;
-  for (const auto& frame : thread.current_backtrace()) {
+  for (const auto& frame : backtrace) {
     std::string function;
 
     if (!frame.function_name().empty()) {
@@ -159,16 +161,32 @@
   }
 }
 
+static void print_thread_backtrace(CallbackType callback, const Tombstone& tombstone,
+                                   const Thread& thread, bool should_log) {
+  CBS("");
+  CB(should_log, "backtrace:");
+  print_backtrace(callback, tombstone, thread.current_backtrace(), should_log);
+}
+
 static void print_thread_memory_dump(CallbackType callback, const Tombstone& tombstone,
                                      const Thread& thread) {
   static constexpr size_t bytes_per_line = 16;
+  static_assert(bytes_per_line == kTagGranuleSize);
   int word_size = pointer_width(tombstone);
   for (const auto& mem : thread.memory_dump()) {
     CBS("");
-    CBS("memory near %s (%s):", mem.register_name().c_str(), mem.mapping_name().c_str());
+    if (mem.mapping_name().empty()) {
+      CBS("memory near %s:", mem.register_name().c_str());
+    } else {
+      CBS("memory near %s (%s):", mem.register_name().c_str(), mem.mapping_name().c_str());
+    }
     uint64_t addr = mem.begin_address();
     for (size_t offset = 0; offset < mem.memory().size(); offset += bytes_per_line) {
-      std::string line = StringPrintf("    %0*" PRIx64, word_size * 2, addr + offset);
+      uint64_t tagged_addr = addr;
+      if (mem.tags().size() > offset / kTagGranuleSize) {
+        tagged_addr |= static_cast<uint64_t>(mem.tags()[offset / kTagGranuleSize]) << 56;
+      }
+      std::string line = StringPrintf("    %0*" PRIx64, word_size * 2, tagged_addr + offset);
 
       size_t bytes = std::min(bytes_per_line, mem.memory().size() - offset);
       for (size_t i = 0; i < bytes; i += word_size) {
@@ -231,9 +249,8 @@
         sender_desc.c_str(), fault_addr_desc.c_str());
   }
 
-  if (tombstone.has_cause()) {
-    const Cause& cause = tombstone.cause();
-    CBL("Cause: %s", cause.human_readable().c_str());
+  if (tombstone.causes_size() == 1) {
+    CBL("Cause: %s", tombstone.causes(0).human_readable().c_str());
   }
 
   if (!tombstone.abort_message().empty()) {
@@ -242,6 +259,36 @@
 
   print_thread_registers(callback, tombstone, thread, true);
   print_thread_backtrace(callback, tombstone, thread, true);
+
+  if (tombstone.causes_size() > 1) {
+    CBS("");
+    CBS("Note: multiple potential causes for this crash were detected, listing them in decreasing "
+        "order of probability.");
+  }
+
+  for (const Cause& cause : tombstone.causes()) {
+    if (tombstone.causes_size() > 1) {
+      CBS("");
+      CBS("Cause: %s", cause.human_readable().c_str());
+    }
+
+    if (cause.has_memory_error() && cause.memory_error().has_heap()) {
+      const HeapObject& heap_object = cause.memory_error().heap();
+
+      if (heap_object.deallocation_backtrace_size() != 0) {
+        CBS("");
+        CBS("deallocated by thread %" PRIu64 ":", heap_object.deallocation_tid());
+        print_backtrace(callback, tombstone, heap_object.deallocation_backtrace(), false);
+      }
+
+      if (heap_object.allocation_backtrace_size() != 0) {
+        CBS("");
+        CBS("allocated by thread %" PRIu64 ":", heap_object.allocation_tid());
+        print_backtrace(callback, tombstone, heap_object.allocation_backtrace(), false);
+      }
+    }
+  }
+
   print_thread_memory_dump(callback, tombstone, thread);
 
   CBS("");
@@ -313,6 +360,7 @@
   CBL("Revision: '%s'", tombstone.revision().c_str());
   CBL("ABI: '%s'", abi_string(tombstone));
   CBL("Timestamp: %s", tombstone.timestamp().c_str());
+  CBL("Process uptime: %ds", tombstone.process_uptime());
 
   // Process header
   const auto& threads = tombstone.threads();
diff --git a/debuggerd/libdebuggerd/utility.cpp b/debuggerd/libdebuggerd/utility.cpp
index 6f13ed4..2c645b5 100644
--- a/debuggerd/libdebuggerd/utility.cpp
+++ b/debuggerd/libdebuggerd/utility.cpp
@@ -125,8 +125,9 @@
 
 #define MEMORY_BYTES_TO_DUMP 256
 #define MEMORY_BYTES_PER_LINE 16
+static_assert(MEMORY_BYTES_PER_LINE == kTagGranuleSize);
 
-ssize_t dump_memory(void* out, size_t len, size_t* start_offset, uint64_t* addr,
+ssize_t dump_memory(void* out, size_t len, uint8_t* tags, size_t tags_len, uint64_t* addr,
                     unwindstack::Memory* memory) {
   // Align the address to the number of bytes per line to avoid confusing memory tag output if
   // memory is tagged and we start from a misaligned address. Start 32 bytes before the address.
@@ -154,17 +155,17 @@
     bytes &= ~(sizeof(uintptr_t) - 1);
   }
 
-  *start_offset = 0;
   bool skip_2nd_read = false;
   if (bytes == 0) {
     // In this case, we might want to try another read at the beginning of
     // the next page only if it's within the amount of memory we would have
     // read.
     size_t page_size = sysconf(_SC_PAGE_SIZE);
-    *start_offset = ((*addr + (page_size - 1)) & ~(page_size - 1)) - *addr;
-    if (*start_offset == 0 || *start_offset >= len) {
+    uint64_t next_page = (*addr + (page_size - 1)) & ~(page_size - 1);
+    if (next_page == *addr || next_page >= *addr + len) {
       skip_2nd_read = true;
     }
+    *addr = next_page;
   }
 
   if (bytes < len && !skip_2nd_read) {
@@ -174,8 +175,7 @@
     // into a readable map. Only requires one extra read because a map has
     // to contain at least one page, and the total number of bytes to dump
     // is smaller than a page.
-    size_t bytes2 = memory->Read(*addr + *start_offset + bytes, static_cast<uint8_t*>(out) + bytes,
-                                 len - bytes - *start_offset);
+    size_t bytes2 = memory->Read(*addr + bytes, static_cast<uint8_t*>(out) + bytes, len - bytes);
     bytes += bytes2;
     if (bytes2 > 0 && bytes % sizeof(uintptr_t) != 0) {
       // This should never happen, but we'll try and continue any way.
@@ -190,15 +190,24 @@
     return -1;
   }
 
+  for (uint64_t tag_granule = 0; tag_granule < bytes / kTagGranuleSize; ++tag_granule) {
+    long tag = memory->ReadTag(*addr + kTagGranuleSize * tag_granule);
+    if (tag_granule < tags_len) {
+      tags[tag_granule] = tag >= 0 ? tag : 0;
+    } else {
+      ALOGE("Insufficient space for tags");
+    }
+  }
+
   return bytes;
 }
 
 void dump_memory(log_t* log, unwindstack::Memory* memory, uint64_t addr, const std::string& label) {
   // Dump 256 bytes
   uintptr_t data[MEMORY_BYTES_TO_DUMP / sizeof(uintptr_t)];
-  size_t start_offset = 0;
+  uint8_t tags[MEMORY_BYTES_TO_DUMP / kTagGranuleSize];
 
-  ssize_t bytes = dump_memory(data, sizeof(data), &start_offset, &addr, memory);
+  ssize_t bytes = dump_memory(data, sizeof(data), tags, sizeof(tags), &addr, memory);
   if (bytes == -1) {
     return;
   }
@@ -212,38 +221,27 @@
   // On 32-bit machines, there are still 16 bytes per line but addresses and
   // words are of course presented differently.
   uintptr_t* data_ptr = data;
-  size_t current = 0;
-  size_t total_bytes = start_offset + bytes;
-  for (size_t line = 0; line < MEMORY_BYTES_TO_DUMP / MEMORY_BYTES_PER_LINE; line++) {
-    uint64_t tagged_addr = addr;
-    long tag = memory->ReadTag(addr);
-    if (tag >= 0) {
-      tagged_addr |= static_cast<uint64_t>(tag) << 56;
-    }
+  uint8_t* tags_ptr = tags;
+  for (size_t line = 0; line < static_cast<size_t>(bytes) / MEMORY_BYTES_PER_LINE; line++) {
+    uint64_t tagged_addr = addr | static_cast<uint64_t>(*tags_ptr++) << 56;
     std::string logline;
     android::base::StringAppendF(&logline, "    %" PRIPTR, tagged_addr);
 
     addr += MEMORY_BYTES_PER_LINE;
     std::string ascii;
     for (size_t i = 0; i < MEMORY_BYTES_PER_LINE / sizeof(uintptr_t); i++) {
-      if (current >= start_offset && current + sizeof(uintptr_t) <= total_bytes) {
-        android::base::StringAppendF(&logline, " %" PRIPTR, static_cast<uint64_t>(*data_ptr));
+      android::base::StringAppendF(&logline, " %" PRIPTR, static_cast<uint64_t>(*data_ptr));
 
-        // Fill out the ascii string from the data.
-        uint8_t* ptr = reinterpret_cast<uint8_t*>(data_ptr);
-        for (size_t val = 0; val < sizeof(uintptr_t); val++, ptr++) {
-          if (*ptr >= 0x20 && *ptr < 0x7f) {
-            ascii += *ptr;
-          } else {
-            ascii += '.';
-          }
+      // Fill out the ascii string from the data.
+      uint8_t* ptr = reinterpret_cast<uint8_t*>(data_ptr);
+      for (size_t val = 0; val < sizeof(uintptr_t); val++, ptr++) {
+        if (*ptr >= 0x20 && *ptr < 0x7f) {
+          ascii += *ptr;
+        } else {
+          ascii += '.';
         }
-        data_ptr++;
-      } else {
-        logline += ' ' + std::string(sizeof(uintptr_t) * 2, '-');
-        ascii += std::string(sizeof(uintptr_t), '.');
       }
-      current += sizeof(uintptr_t);
+      data_ptr++;
     }
     _LOG(log, logtype::MEMORY, "%s  %s\n", logline.c_str(), ascii.c_str());
   }
diff --git a/debuggerd/proto/tombstone.proto b/debuggerd/proto/tombstone.proto
index 2c7156b..dd15ff6 100644
--- a/debuggerd/proto/tombstone.proto
+++ b/debuggerd/proto/tombstone.proto
@@ -19,16 +19,19 @@
 
   string process_name = 9;
 
+  // Process uptime in seconds.
+  uint32 process_uptime = 20;
+
   Signal signal_info = 10;
   string abort_message = 14;
-  Cause cause = 15;
+  repeated Cause causes = 15;
 
   map<uint32, Thread> threads = 16;
   repeated MemoryMapping memory_mappings = 17;
   repeated LogBuffer log_buffers = 18;
   repeated FD open_fds = 19;
 
-  reserved 20 to 999;
+  reserved 21 to 999;
 }
 
 enum Architecture {
@@ -57,10 +60,52 @@
   reserved 10 to 999;
 }
 
+message HeapObject {
+  uint64 address = 1;
+  uint64 size = 2;
+
+  uint64 allocation_tid = 3;
+  repeated BacktraceFrame allocation_backtrace = 4;
+
+  uint64 deallocation_tid = 5;
+  repeated BacktraceFrame deallocation_backtrace = 6;
+}
+
+message MemoryError {
+  enum Tool {
+    GWP_ASAN = 0;
+    SCUDO = 1;
+
+    reserved 2 to 999;
+  }
+  Tool tool = 1;
+
+  enum Type {
+    UNKNOWN = 0;
+    USE_AFTER_FREE = 1;
+    DOUBLE_FREE = 2;
+    INVALID_FREE = 3;
+    BUFFER_OVERFLOW = 4;
+    BUFFER_UNDERFLOW = 5;
+
+    reserved 6 to 999;
+  }
+  Type type = 2;
+
+  oneof location {
+    HeapObject heap = 3;
+  }
+
+  reserved 4 to 999;
+}
+
 message Cause {
   string human_readable = 1;
+  oneof details {
+    MemoryError memory_error = 2;
+  }
 
-  reserved 2 to 999;
+  reserved 3 to 999;
 }
 
 message Register {
@@ -76,8 +121,9 @@
   repeated Register registers = 3;
   repeated BacktraceFrame current_backtrace = 4;
   repeated MemoryDump memory_dump = 5;
+  int64 tagged_addr_ctrl = 6;
 
-  reserved 6 to 999;
+  reserved 7 to 999;
 }
 
 message BacktraceFrame {
@@ -100,8 +146,9 @@
   string mapping_name = 2;
   uint64 begin_address = 3;
   bytes memory = 4;
+  bytes tags = 5;
 
-  reserved 5 to 999;
+  reserved 6 to 999;
 }
 
 message MemoryMapping {
diff --git a/debuggerd/protocol.h b/debuggerd/protocol.h
index 53a76ea..f33b2f0 100644
--- a/debuggerd/protocol.h
+++ b/debuggerd/protocol.h
@@ -97,6 +97,7 @@
   uintptr_t gwp_asan_metadata;
   uintptr_t scudo_stack_depot;
   uintptr_t scudo_region_info;
+  uintptr_t scudo_ring_buffer;
 };
 
 struct __attribute__((__packed__)) CrashInfo {
diff --git a/debuggerd/tombstoned/intercept_manager.cpp b/debuggerd/tombstoned/intercept_manager.cpp
index 4d4646a..613e6f5 100644
--- a/debuggerd/tombstoned/intercept_manager.cpp
+++ b/debuggerd/tombstoned/intercept_manager.cpp
@@ -163,7 +163,7 @@
     event_assign(intercept->intercept_event, intercept_manager->base, sockfd, EV_READ | EV_TIMEOUT,
                  intercept_close_cb, arg);
 
-    struct timeval timeout = {.tv_sec = 10 * android::base::TimeoutMultiplier(), .tv_usec = 0};
+    struct timeval timeout = {.tv_sec = 10 * android::base::HwTimeoutMultiplier(), .tv_usec = 0};
     event_add(intercept->intercept_event, &timeout);
   }
 
@@ -179,7 +179,7 @@
   intercept->intercept_manager = static_cast<InterceptManager*>(arg);
   intercept->sockfd.reset(sockfd);
 
-  struct timeval timeout = {1 * android::base::TimeoutMultiplier(), 0};
+  struct timeval timeout = {1 * android::base::HwTimeoutMultiplier(), 0};
   event_base* base = evconnlistener_get_base(listener);
   event* intercept_event =
     event_new(base, sockfd, EV_TIMEOUT | EV_READ, intercept_request_cb, intercept);
diff --git a/debuggerd/tombstoned/tombstoned.cpp b/debuggerd/tombstoned/tombstoned.cpp
index bc2d33d..0b87b7a 100644
--- a/debuggerd/tombstoned/tombstoned.cpp
+++ b/debuggerd/tombstoned/tombstoned.cpp
@@ -320,7 +320,7 @@
   }
 
   // TODO: Make this configurable by the interceptor?
-  struct timeval timeout = {10 * android::base::TimeoutMultiplier(), 0};
+  struct timeval timeout = {10 * android::base::HwTimeoutMultiplier(), 0};
 
   event_base* base = event_get_base(crash->crash_event);
 
@@ -340,7 +340,7 @@
 
   // TODO: Make sure that only java crashes come in on the java socket
   // and only native crashes on the native socket.
-  struct timeval timeout = {1 * android::base::TimeoutMultiplier(), 0};
+  struct timeval timeout = {1 * android::base::HwTimeoutMultiplier(), 0};
   event* crash_event = event_new(base, sockfd, EV_TIMEOUT | EV_READ, crash_request_cb, crash);
   crash->crash_socket_fd.reset(sockfd);
   crash->crash_event = crash_event;
diff --git a/fastboot/Android.bp b/fastboot/Android.bp
index a1f1c17..bf9ec90 100644
--- a/fastboot/Android.bp
+++ b/fastboot/Android.bp
@@ -55,6 +55,7 @@
         "tcp.cpp",
         "udp.cpp",
         "util.cpp",
+        "vendor_boot_img_utils.cpp",
         "fastboot_driver.cpp",
     ],
 
@@ -75,7 +76,9 @@
     ],
 
     header_libs: [
+        "avb_headers",
         "bootimg_headers",
+        "libstorage_literals_headers",
     ],
 
     export_header_lib_headers: [
@@ -124,6 +127,7 @@
         "-Wextra",
         "-Werror",
         "-Wvla",
+        "-DANDROID_BASE_UNIQUE_FD_DISABLE_IMPLICIT_CONVERSION",
     ],
     rtti: true,
 
@@ -138,6 +142,12 @@
 
     recovery: true,
 
+    product_variables: {
+        debuggable: {
+            cppflags: ["-DFB_ENABLE_FETCH"],
+        },
+    },
+
     srcs: [
         "device/commands.cpp",
         "device/fastboot_device.cpp",
@@ -183,7 +193,8 @@
     header_libs: [
         "avb_headers",
         "libsnapshot_headers",
-    ]
+        "libstorage_literals_headers",
+    ],
 }
 
 cc_defaults {
@@ -196,6 +207,7 @@
         "-Wextra",
         "-Werror",
         "-Wunreachable-code",
+        "-DANDROID_BASE_UNIQUE_FD_DISABLE_IMPLICIT_CONVERSION",
     ],
 
     target: {
@@ -262,12 +274,17 @@
         "tcp.cpp",
         "udp.cpp",
         "util.cpp",
+        "vendor_boot_img_utils.cpp",
         "fastboot_driver.cpp",
     ],
 
     // Only version the final binaries
     use_version_lib: false,
     static_libs: ["libbuildversion"],
+    header_libs: [
+        "avb_headers",
+        "libstorage_literals_headers",
+    ],
 
     generated_headers: ["platform_tools_version"],
 
@@ -359,3 +376,33 @@
         },
     },
 }
+
+cc_test_host {
+    name: "fastboot_vendor_boot_img_utils_test",
+    srcs: ["vendor_boot_img_utils_test.cpp"],
+    static_libs: [
+        "libbase",
+        "libc++fs",
+        "libfastboot",
+        "libgmock",
+        "liblog",
+    ],
+    header_libs: [
+        "avb_headers",
+        "bootimg_headers",
+    ],
+    cflags: [
+        "-Wall",
+        "-Werror",
+    ],
+    data: [
+        ":fastboot_test_dtb",
+        ":fastboot_test_bootconfig",
+        ":fastboot_test_vendor_ramdisk_none",
+        ":fastboot_test_vendor_ramdisk_platform",
+        ":fastboot_test_vendor_ramdisk_replace",
+        ":fastboot_test_vendor_boot_v3",
+        ":fastboot_test_vendor_boot_v4_without_frag",
+        ":fastboot_test_vendor_boot_v4_with_frag"
+    ],
+}
diff --git a/fastboot/constants.h b/fastboot/constants.h
index ba43ca5..4ea68da 100644
--- a/fastboot/constants.h
+++ b/fastboot/constants.h
@@ -35,6 +35,7 @@
 #define FB_CMD_OEM "oem"
 #define FB_CMD_GSI "gsi"
 #define FB_CMD_SNAPSHOT_UPDATE "snapshot-update"
+#define FB_CMD_FETCH "fetch"
 
 #define RESPONSE_OKAY "OKAY"
 #define RESPONSE_FAIL "FAIL"
@@ -77,3 +78,4 @@
 #define FB_VAR_FIRST_API_LEVEL "first-api-level"
 #define FB_VAR_SECURITY_PATCH_LEVEL "security-patch-level"
 #define FB_VAR_TREBLE_ENABLED "treble-enabled"
+#define FB_VAR_MAX_FETCH_SIZE "max-fetch-size"
diff --git a/fastboot/device/commands.cpp b/fastboot/device/commands.cpp
index b2b6a9e..0a72812 100644
--- a/fastboot/device/commands.cpp
+++ b/fastboot/device/commands.cpp
@@ -16,6 +16,7 @@
 
 #include "commands.h"
 
+#include <inttypes.h>
 #include <sys/socket.h>
 #include <sys/un.h>
 
@@ -36,6 +37,7 @@
 #include <liblp/builder.h>
 #include <liblp/liblp.h>
 #include <libsnapshot/snapshot.h>
+#include <storage_literals/storage_literals.h>
 #include <uuid/uuid.h>
 
 #include "constants.h"
@@ -43,6 +45,12 @@
 #include "flashing.h"
 #include "utility.h"
 
+#ifdef FB_ENABLE_FETCH
+static constexpr bool kEnableFetch = true;
+#else
+static constexpr bool kEnableFetch = false;
+#endif
+
 using android::fs_mgr::MetadataBuilder;
 using ::android::hardware::hidl_string;
 using ::android::hardware::boot::V1_0::BoolResult;
@@ -54,6 +62,8 @@
 using android::snapshot::SnapshotManager;
 using IBootControl1_1 = ::android::hardware::boot::V1_1::IBootControl;
 
+using namespace android::storage_literals;
+
 struct VariableHandlers {
     // Callback to retrieve the value of a single variable.
     std::function<bool(FastbootDevice*, const std::vector<std::string>&, std::string*)> get;
@@ -136,7 +146,9 @@
             {FB_VAR_DYNAMIC_PARTITION, {GetDynamicPartition, nullptr}},
             {FB_VAR_FIRST_API_LEVEL, {GetFirstApiLevel, nullptr}},
             {FB_VAR_SECURITY_PATCH_LEVEL, {GetSecurityPatchLevel, nullptr}},
-            {FB_VAR_TREBLE_ENABLED, {GetTrebleEnabled, nullptr}}};
+            {FB_VAR_TREBLE_ENABLED, {GetTrebleEnabled, nullptr}},
+            {FB_VAR_MAX_FETCH_SIZE, {GetMaxFetchSize, nullptr}},
+    };
 
     if (args.size() < 2) {
         return device->WriteFail("Missing argument");
@@ -380,13 +392,13 @@
 
     struct sockaddr_un addr = {.sun_family = AF_UNIX};
     strncpy(addr.sun_path, "/dev/socket/recovery", sizeof(addr.sun_path) - 1);
-    if (connect(sock, (struct sockaddr*)&addr, sizeof(addr)) < 0) {
+    if (connect(sock.get(), (struct sockaddr*)&addr, sizeof(addr)) < 0) {
         PLOG(ERROR) << "Couldn't connect to recovery";
         return false;
     }
     // Switch to recovery will not update the boot reason since it does not
     // require a reboot.
-    auto ret = write(sock, &msg_switch_to_recovery, sizeof(msg_switch_to_recovery));
+    auto ret = write(sock.get(), &msg_switch_to_recovery, sizeof(msg_switch_to_recovery));
     if (ret != sizeof(msg_switch_to_recovery)) {
         PLOG(ERROR) << "Couldn't write message to switch to recovery";
         return false;
@@ -671,3 +683,175 @@
     }
     return device->WriteStatus(FastbootResult::OKAY, "Success");
 }
+
+namespace {
+// Helper of FetchHandler.
+class PartitionFetcher {
+  public:
+    static bool Fetch(FastbootDevice* device, const std::vector<std::string>& args) {
+        if constexpr (!kEnableFetch) {
+            return device->WriteFail("Fetch is not allowed on user build");
+        }
+
+        if (GetDeviceLockStatus()) {
+            return device->WriteFail("Fetch is not allowed on locked devices");
+        }
+
+        PartitionFetcher fetcher(device, args);
+        if (fetcher.Open()) {
+            fetcher.Fetch();
+        }
+        CHECK(fetcher.ret_.has_value());
+        return *fetcher.ret_;
+    }
+
+  private:
+    PartitionFetcher(FastbootDevice* device, const std::vector<std::string>& args)
+        : device_(device), args_(&args) {}
+    // Return whether the partition is successfully opened.
+    // If successfully opened, ret_ is left untouched. Otherwise, ret_ is set to the value
+    // that FetchHandler should return.
+    bool Open() {
+        if (args_->size() < 2) {
+            ret_ = device_->WriteFail("Missing partition arg");
+            return false;
+        }
+
+        partition_name_ = args_->at(1);
+        if (std::find(kAllowedPartitions.begin(), kAllowedPartitions.end(), partition_name_) ==
+            kAllowedPartitions.end()) {
+            ret_ = device_->WriteFail("Fetch is only allowed on [" +
+                                      android::base::Join(kAllowedPartitions, ", ") + "]");
+            return false;
+        }
+
+        if (!OpenPartition(device_, partition_name_, &handle_, true /* read */)) {
+            ret_ = device_->WriteFail(
+                    android::base::StringPrintf("Cannot open %s", partition_name_.c_str()));
+            return false;
+        }
+
+        partition_size_ = get_block_device_size(handle_.fd());
+        if (partition_size_ == 0) {
+            ret_ = device_->WriteOkay(android::base::StringPrintf("Partition %s has size 0",
+                                                                  partition_name_.c_str()));
+            return false;
+        }
+
+        start_offset_ = 0;
+        if (args_->size() >= 3) {
+            if (!android::base::ParseUint(args_->at(2), &start_offset_)) {
+                ret_ = device_->WriteFail("Invalid offset, must be integer");
+                return false;
+            }
+            if (start_offset_ > std::numeric_limits<off64_t>::max()) {
+                ret_ = device_->WriteFail(
+                        android::base::StringPrintf("Offset overflows: %" PRIx64, start_offset_));
+                return false;
+            }
+        }
+        if (start_offset_ > partition_size_) {
+            ret_ = device_->WriteFail(android::base::StringPrintf(
+                    "Invalid offset 0x%" PRIx64 ", partition %s has size 0x%" PRIx64, start_offset_,
+                    partition_name_.c_str(), partition_size_));
+            return false;
+        }
+        uint64_t maximum_total_size_to_read = partition_size_ - start_offset_;
+        total_size_to_read_ = maximum_total_size_to_read;
+        if (args_->size() >= 4) {
+            if (!android::base::ParseUint(args_->at(3), &total_size_to_read_)) {
+                ret_ = device_->WriteStatus(FastbootResult::FAIL, "Invalid size, must be integer");
+                return false;
+            }
+        }
+        if (total_size_to_read_ == 0) {
+            ret_ = device_->WriteOkay("Read 0 bytes");
+            return false;
+        }
+        if (total_size_to_read_ > maximum_total_size_to_read) {
+            ret_ = device_->WriteFail(android::base::StringPrintf(
+                    "Invalid size to read 0x%" PRIx64 ", partition %s has size 0x%" PRIx64
+                    " and fetching from offset 0x%" PRIx64,
+                    total_size_to_read_, partition_name_.c_str(), partition_size_, start_offset_));
+            return false;
+        }
+
+        if (total_size_to_read_ > kMaxFetchSizeDefault) {
+            ret_ = device_->WriteFail(android::base::StringPrintf(
+                    "Cannot fetch 0x%" PRIx64
+                    " bytes because it exceeds maximum transport size 0x%x",
+                    partition_size_, kMaxDownloadSizeDefault));
+            return false;
+        }
+
+        return true;
+    }
+
+    // Assume Open() returns true.
+    // After execution, ret_ is set to the value that FetchHandler should return.
+    void Fetch() {
+        CHECK(start_offset_ <= std::numeric_limits<off64_t>::max());
+        if (lseek64(handle_.fd(), start_offset_, SEEK_SET) != static_cast<off64_t>(start_offset_)) {
+            ret_ = device_->WriteFail(android::base::StringPrintf(
+                    "On partition %s, unable to lseek(0x%" PRIx64 ": %s", partition_name_.c_str(),
+                    start_offset_, strerror(errno)));
+            return;
+        }
+
+        if (!device_->WriteStatus(FastbootResult::DATA,
+                                  android::base::StringPrintf(
+                                          "%08x", static_cast<uint32_t>(total_size_to_read_)))) {
+            ret_ = false;
+            return;
+        }
+        uint64_t end_offset = start_offset_ + total_size_to_read_;
+        std::vector<char> buf(1_MiB);
+        uint64_t current_offset = start_offset_;
+        while (current_offset < end_offset) {
+            // On any error, exit. We can't return a status message to the driver because
+            // we are in the middle of writing data, so just let the driver guess what's wrong
+            // by ending the data stream prematurely.
+            uint64_t remaining = end_offset - current_offset;
+            uint64_t chunk_size = std::min<uint64_t>(buf.size(), remaining);
+            if (!android::base::ReadFully(handle_.fd(), buf.data(), chunk_size)) {
+                PLOG(ERROR) << std::hex << "Unable to read 0x" << chunk_size << " bytes from "
+                            << partition_name_ << " @ offset 0x" << current_offset;
+                ret_ = false;
+                return;
+            }
+            if (!device_->HandleData(false /* is read */, buf.data(), chunk_size)) {
+                PLOG(ERROR) << std::hex << "Unable to send 0x" << chunk_size << " bytes of "
+                            << partition_name_ << " @ offset 0x" << current_offset;
+                ret_ = false;
+                return;
+            }
+            current_offset += chunk_size;
+        }
+
+        ret_ = device_->WriteOkay(android::base::StringPrintf(
+                "Fetched %s (offset=0x%" PRIx64 ", size=0x%" PRIx64, partition_name_.c_str(),
+                start_offset_, total_size_to_read_));
+    }
+
+    static constexpr std::array<const char*, 3> kAllowedPartitions{
+            "vendor_boot",
+            "vendor_boot_a",
+            "vendor_boot_b",
+    };
+
+    FastbootDevice* device_;
+    const std::vector<std::string>* args_ = nullptr;
+    std::string partition_name_;
+    PartitionHandle handle_;
+    uint64_t partition_size_ = 0;
+    uint64_t start_offset_ = 0;
+    uint64_t total_size_to_read_ = 0;
+
+    // What FetchHandler should return.
+    std::optional<bool> ret_ = std::nullopt;
+};
+}  // namespace
+
+bool FetchHandler(FastbootDevice* device, const std::vector<std::string>& args) {
+    return PartitionFetcher::Fetch(device, args);
+}
diff --git a/fastboot/device/commands.h b/fastboot/device/commands.h
index c1324bc..345ae1a 100644
--- a/fastboot/device/commands.h
+++ b/fastboot/device/commands.h
@@ -20,6 +20,7 @@
 #include <vector>
 
 constexpr unsigned int kMaxDownloadSizeDefault = 0x10000000;
+constexpr unsigned int kMaxFetchSizeDefault = 0x10000000;
 
 class FastbootDevice;
 
@@ -50,3 +51,4 @@
 bool OemCmdHandler(FastbootDevice* device, const std::vector<std::string>& args);
 bool GsiHandler(FastbootDevice* device, const std::vector<std::string>& args);
 bool SnapshotUpdateHandler(FastbootDevice* device, const std::vector<std::string>& args);
+bool FetchHandler(FastbootDevice* device, const std::vector<std::string>& args);
diff --git a/fastboot/device/fastboot_device.cpp b/fastboot/device/fastboot_device.cpp
index 35f3de0..64a934d 100644
--- a/fastboot/device/fastboot_device.cpp
+++ b/fastboot/device/fastboot_device.cpp
@@ -61,6 +61,7 @@
               {FB_CMD_OEM, OemCmdHandler},
               {FB_CMD_GSI, GsiHandler},
               {FB_CMD_SNAPSHOT_UPDATE, SnapshotUpdateHandler},
+              {FB_CMD_FETCH, FetchHandler},
       }),
       boot_control_hal_(IBootControl::getService()),
       health_hal_(get_health_service()),
@@ -137,14 +138,18 @@
 }
 
 bool FastbootDevice::HandleData(bool read, std::vector<char>* data) {
-    auto read_write_data_size = read ? this->get_transport()->Read(data->data(), data->size())
-                                     : this->get_transport()->Write(data->data(), data->size());
+    return HandleData(read, data->data(), data->size());
+}
+
+bool FastbootDevice::HandleData(bool read, char* data, uint64_t size) {
+    auto read_write_data_size = read ? this->get_transport()->Read(data, size)
+                                     : this->get_transport()->Write(data, size);
     if (read_write_data_size == -1) {
         LOG(ERROR) << (read ? "read from" : "write to") << " transport failed";
         return false;
     }
-    if (static_cast<size_t>(read_write_data_size) != data->size()) {
-        LOG(ERROR) << (read ? "read" : "write") << " expected " << data->size() << " bytes, got "
+    if (static_cast<size_t>(read_write_data_size) != size) {
+        LOG(ERROR) << (read ? "read" : "write") << " expected " << size << " bytes, got "
                    << read_write_data_size;
         return false;
     }
diff --git a/fastboot/device/fastboot_device.h b/fastboot/device/fastboot_device.h
index 23be721..3536136 100644
--- a/fastboot/device/fastboot_device.h
+++ b/fastboot/device/fastboot_device.h
@@ -40,6 +40,7 @@
     void ExecuteCommands();
     bool WriteStatus(FastbootResult result, const std::string& message);
     bool HandleData(bool read, std::vector<char>* data);
+    bool HandleData(bool read, char* data, uint64_t size);
     std::string GetCurrentSlot();
 
     // Shortcuts for writing status results.
diff --git a/fastboot/device/usb.cpp b/fastboot/device/usb.cpp
index 4bee7b2..4115a6d 100644
--- a/fastboot/device/usb.cpp
+++ b/fastboot/device/usb.cpp
@@ -82,7 +82,7 @@
     int orig_len = len;
     while (len > 0) {
         int write_len = std::min(USB_FFS_BULK_SIZE, len);
-        int n = write(h->bulk_in, buf, write_len);
+        int n = write(h->bulk_in.get(), buf, write_len);
         if (n < 0) {
             D("ERROR: fd = %d, n = %d: %s", h->bulk_in.get(), n, strerror(errno));
             return -1;
@@ -103,7 +103,7 @@
     unsigned count = 0;
     while (len > 0) {
         int read_len = std::min(USB_FFS_BULK_SIZE, len);
-        int n = read(h->bulk_out, buf, read_len);
+        int n = read(h->bulk_out.get(), buf, read_len);
         if (n < 0) {
             D("ERROR: fd = %d, n = %d: %s", h->bulk_out.get(), n, strerror(errno));
             return -1;
diff --git a/fastboot/device/utility.cpp b/fastboot/device/utility.cpp
index 7c6ac89..f9267e0 100644
--- a/fastboot/device/utility.cpp
+++ b/fastboot/device/utility.cpp
@@ -77,7 +77,8 @@
 
 }  // namespace
 
-bool OpenPartition(FastbootDevice* device, const std::string& name, PartitionHandle* handle) {
+bool OpenPartition(FastbootDevice* device, const std::string& name, PartitionHandle* handle,
+                   bool read) {
     // We prioritize logical partitions over physical ones, and do this
     // consistently for other partition operations (like getvar:partition-size).
     if (LogicalPartitionExists(device, name)) {
@@ -89,7 +90,9 @@
         return false;
     }
 
-    unique_fd fd(TEMP_FAILURE_RETRY(open(handle->path().c_str(), O_WRONLY | O_EXCL)));
+    int flags = (read ? O_RDONLY : O_WRONLY);
+    flags |= (O_EXCL | O_CLOEXEC | O_BINARY);
+    unique_fd fd(TEMP_FAILURE_RETRY(open(handle->path().c_str(), flags)));
     if (fd < 0) {
         PLOG(ERROR) << "Failed to open block device: " << handle->path();
         return false;
diff --git a/fastboot/device/utility.h b/fastboot/device/utility.h
index 3b71ef0..c2646d7 100644
--- a/fastboot/device/utility.h
+++ b/fastboot/device/utility.h
@@ -75,7 +75,11 @@
 std::optional<std::string> FindPhysicalPartition(const std::string& name);
 bool LogicalPartitionExists(FastbootDevice* device, const std::string& name,
                             bool* is_zero_length = nullptr);
-bool OpenPartition(FastbootDevice* device, const std::string& name, PartitionHandle* handle);
+
+// If read, partition is readonly. Else it is write only.
+bool OpenPartition(FastbootDevice* device, const std::string& name, PartitionHandle* handle,
+                   bool read = false);
+
 bool GetSlotNumber(const std::string& slot, android::hardware::boot::V1_0::Slot* number);
 std::vector<std::string> ListPartitions(FastbootDevice* device);
 bool GetDeviceLockStatus();
diff --git a/fastboot/device/variables.cpp b/fastboot/device/variables.cpp
index e7d8bc3..ee1eed8 100644
--- a/fastboot/device/variables.cpp
+++ b/fastboot/device/variables.cpp
@@ -33,6 +33,12 @@
 #include "flashing.h"
 #include "utility.h"
 
+#ifdef FB_ENABLE_FETCH
+static constexpr bool kEnableFetch = true;
+#else
+static constexpr bool kEnableFetch = false;
+#endif
+
 using ::android::hardware::boot::V1_0::BoolResult;
 using ::android::hardware::boot::V1_0::Slot;
 using ::android::hardware::boot::V1_1::MergeStatus;
@@ -509,3 +515,13 @@
     *message = android::base::GetProperty("ro.treble.enabled", "");
     return true;
 }
+
+bool GetMaxFetchSize(FastbootDevice* /* device */, const std::vector<std::string>& /* args */,
+                     std::string* message) {
+    if (!kEnableFetch) {
+        *message = "fetch not supported on user builds";
+        return false;
+    }
+    *message = android::base::StringPrintf("0x%X", kMaxFetchSizeDefault);
+    return true;
+}
diff --git a/fastboot/device/variables.h b/fastboot/device/variables.h
index c11e472..f40a025 100644
--- a/fastboot/device/variables.h
+++ b/fastboot/device/variables.h
@@ -80,6 +80,8 @@
                            std::string* message);
 bool GetTrebleEnabled(FastbootDevice* device, const std::vector<std::string>& args,
                       std::string* message);
+bool GetMaxFetchSize(FastbootDevice* /* device */, const std::vector<std::string>& /* args */,
+                     std::string* message);
 
 // Helpers for getvar all.
 std::vector<std::vector<std::string>> GetAllPartitionArgsWithSlot(FastbootDevice* device);
diff --git a/fastboot/fastboot.cpp b/fastboot/fastboot.cpp
index 829958a..f720dda 100644
--- a/fastboot/fastboot.cpp
+++ b/fastboot/fastboot.cpp
@@ -76,12 +76,15 @@
 #include "udp.h"
 #include "usb.h"
 #include "util.h"
+#include "vendor_boot_img_utils.h"
 
+using android::base::borrowed_fd;
 using android::base::ReadFully;
 using android::base::Split;
 using android::base::Trim;
 using android::base::unique_fd;
 using namespace std::string_literals;
+using namespace std::placeholders;
 
 static const char* serial = nullptr;
 
@@ -114,7 +117,7 @@
     enum fb_buffer_type type;
     void* data;
     int64_t sz;
-    int fd;
+    unique_fd fd;
     int64_t image_size;
 };
 
@@ -227,9 +230,9 @@
     fprintf(stderr, "(bootloader) %s\n", info.c_str());
 }
 
-static int64_t get_file_size(int fd) {
+static int64_t get_file_size(borrowed_fd fd) {
     struct stat sb;
-    if (fstat(fd, &sb) == -1) {
+    if (fstat(fd.get(), &sb) == -1) {
         die("could not get file size");
     }
     return sb.st_size;
@@ -414,6 +417,7 @@
             " snapshot-update merge      On devices that support snapshot-based updates, finish\n"
             "                            an in-progress update if it is in the \"merging\"\n"
             "                            phase.\n"
+            " fetch PARTITION            Fetch a partition image from the device."
             "\n"
             "boot image:\n"
             " boot KERNEL [RAMDISK [SECOND]]\n"
@@ -466,7 +470,7 @@
             " --version                  Display version.\n"
             " --help, -h                 Show this message.\n"
         );
-    // clang-format off
+    // clang-format on
     return 0;
 }
 
@@ -640,31 +644,31 @@
     }
 }
 
-static int unzip_to_file(ZipArchiveHandle zip, const char* entry_name) {
+static unique_fd unzip_to_file(ZipArchiveHandle zip, const char* entry_name) {
     unique_fd fd(make_temporary_fd(entry_name));
 
     ZipEntry64 zip_entry;
     if (FindEntry(zip, entry_name, &zip_entry) != 0) {
         fprintf(stderr, "archive does not contain '%s'\n", entry_name);
         errno = ENOENT;
-        return -1;
+        return unique_fd();
     }
 
     fprintf(stderr, "extracting %s (%" PRIu64 " MB) to disk...", entry_name,
             zip_entry.uncompressed_length / 1024 / 1024);
     double start = now();
-    int error = ExtractEntryToFile(zip, &zip_entry, fd);
+    int error = ExtractEntryToFile(zip, &zip_entry, fd.get());
     if (error != 0) {
         die("\nfailed to extract '%s': %s", entry_name, ErrorCodeString(error));
     }
 
-    if (lseek(fd, 0, SEEK_SET) != 0) {
+    if (lseek(fd.get(), 0, SEEK_SET) != 0) {
         die("\nlseek on extracted file '%s' failed: %s", entry_name, strerror(errno));
     }
 
     fprintf(stderr, " took %.3fs\n", now() - start);
 
-    return fd.release();
+    return fd;
 }
 
 static void CheckRequirement(const std::string& cur_product, const std::string& var,
@@ -851,24 +855,23 @@
     return out_s;
 }
 
-static int64_t get_target_sparse_limit() {
-    std::string max_download_size;
-    if (fb->GetVar("max-download-size", &max_download_size) != fastboot::SUCCESS ||
-        max_download_size.empty()) {
-        verbose("target didn't report max-download-size");
+static uint64_t get_uint_var(const char* var_name) {
+    std::string value_str;
+    if (fb->GetVar(var_name, &value_str) != fastboot::SUCCESS || value_str.empty()) {
+        verbose("target didn't report %s", var_name);
         return 0;
     }
 
     // Some bootloaders (angler, for example) send spurious whitespace too.
-    max_download_size = android::base::Trim(max_download_size);
+    value_str = android::base::Trim(value_str);
 
-    uint64_t limit;
-    if (!android::base::ParseUint(max_download_size, &limit)) {
-        fprintf(stderr, "couldn't parse max-download-size '%s'\n", max_download_size.c_str());
+    uint64_t value;
+    if (!android::base::ParseUint(value_str, &value)) {
+        fprintf(stderr, "couldn't parse %s '%s'\n", var_name, value_str.c_str());
         return 0;
     }
-    if (limit > 0) verbose("target reported max download size of %" PRId64 " bytes", limit);
-    return limit;
+    if (value > 0) verbose("target reported %s of %" PRId64 " bytes", var_name, value);
+    return value;
 }
 
 static int64_t get_sparse_limit(int64_t size) {
@@ -877,7 +880,7 @@
         // Unlimited, so see what the target device's limit is.
         // TODO: shouldn't we apply this limit even if you've used -S?
         if (target_sparse_limit == -1) {
-            target_sparse_limit = get_target_sparse_limit();
+            target_sparse_limit = static_cast<int64_t>(get_uint_var("max-download-size"));
         }
         if (target_sparse_limit > 0) {
             limit = target_sparse_limit;
@@ -893,23 +896,24 @@
     return 0;
 }
 
-static bool load_buf_fd(int fd, struct fastboot_buffer* buf) {
+static bool load_buf_fd(unique_fd fd, struct fastboot_buffer* buf) {
     int64_t sz = get_file_size(fd);
     if (sz == -1) {
         return false;
     }
 
-    if (sparse_file* s = sparse_file_import(fd, false, false)) {
+    if (sparse_file* s = sparse_file_import(fd.get(), false, false)) {
         buf->image_size = sparse_file_len(s, false, false);
         sparse_file_destroy(s);
     } else {
         buf->image_size = sz;
     }
 
-    lseek(fd, 0, SEEK_SET);
+    lseek(fd.get(), 0, SEEK_SET);
     int64_t limit = get_sparse_limit(sz);
+    buf->fd = std::move(fd);
     if (limit) {
-        sparse_file** s = load_sparse_files(fd, limit);
+        sparse_file** s = load_sparse_files(buf->fd.get(), limit);
         if (s == nullptr) {
             return false;
         }
@@ -918,7 +922,6 @@
     } else {
         buf->type = FB_BUFFER_FD;
         buf->data = nullptr;
-        buf->fd = fd;
         buf->sz = sz;
     }
 
@@ -933,7 +936,7 @@
     }
 
     struct stat s;
-    if (fstat(fd, &s)) {
+    if (fstat(fd.get(), &s)) {
         return false;
     }
     if (!S_ISREG(s.st_mode)) {
@@ -941,7 +944,7 @@
         return false;
     }
 
-    return load_buf_fd(fd.release(), buf);
+    return load_buf_fd(std::move(fd), buf);
 }
 
 static void rewrite_vbmeta_buffer(struct fastboot_buffer* buf, bool vbmeta_in_boot) {
@@ -987,13 +990,12 @@
         data[flags_offset] |= 0x02;
     }
 
-    int fd = make_temporary_fd("vbmeta rewriting");
+    unique_fd fd(make_temporary_fd("vbmeta rewriting"));
     if (!android::base::WriteStringToFd(data, fd)) {
         die("Failed writing to modified vbmeta");
     }
-    close(buf->fd);
-    buf->fd = fd;
-    lseek(fd, 0, SEEK_SET);
+    buf->fd = std::move(fd);
+    lseek(buf->fd.get(), 0, SEEK_SET);
 }
 
 static bool has_vbmeta_partition() {
@@ -1012,21 +1014,28 @@
     return var;
 }
 
+static uint64_t get_partition_size(const std::string& partition) {
+    std::string partition_size_str;
+    if (fb->GetVar("partition-size:" + partition, &partition_size_str) != fastboot::SUCCESS) {
+        die("cannot get partition size for %s", partition.c_str());
+    }
+
+    partition_size_str = fb_fix_numeric_var(partition_size_str);
+    uint64_t partition_size;
+    if (!android::base::ParseUint(partition_size_str, &partition_size)) {
+        die("Couldn't parse partition size '%s'.", partition_size_str.c_str());
+    }
+    return partition_size;
+}
+
 static void copy_boot_avb_footer(const std::string& partition, struct fastboot_buffer* buf) {
     if (buf->sz < AVB_FOOTER_SIZE) {
         return;
     }
 
-    std::string partition_size_str;
-    if (fb->GetVar("partition-size:" + partition, &partition_size_str) != fastboot::SUCCESS) {
-        die("cannot get boot partition size");
-    }
+    // If overflows and negative, it should be < buf->sz.
+    int64_t partition_size = static_cast<int64_t>(get_partition_size(partition));
 
-    partition_size_str = fb_fix_numeric_var(partition_size_str);
-    int64_t partition_size;
-    if (!android::base::ParseInt(partition_size_str, &partition_size)) {
-        die("Couldn't parse partition size '%s'.", partition_size_str.c_str());
-    }
     if (partition_size == buf->sz) {
         return;
     }
@@ -1044,18 +1053,17 @@
         return;
     }
 
-    int fd = make_temporary_fd("boot rewriting");
+    unique_fd fd(make_temporary_fd("boot rewriting"));
     if (!android::base::WriteStringToFd(data, fd)) {
         die("Failed writing to modified boot");
     }
-    lseek(fd, partition_size - AVB_FOOTER_SIZE, SEEK_SET);
+    lseek(fd.get(), partition_size - AVB_FOOTER_SIZE, SEEK_SET);
     if (!android::base::WriteStringToFd(data.substr(footer_offset), fd)) {
         die("Failed copying AVB footer in boot");
     }
-    close(buf->fd);
-    buf->fd = fd;
+    buf->fd = std::move(fd);
     buf->sz = partition_size;
-    lseek(fd, 0, SEEK_SET);
+    lseek(buf->fd.get(), 0, SEEK_SET);
 }
 
 static void flash_buf(const std::string& partition, struct fastboot_buffer *buf)
@@ -1185,8 +1193,10 @@
                              const std::function<void(const std::string&)>& func, bool force_slot) {
     std::string has_slot;
     std::string current_slot;
+    // |part| can be vendor_boot:default. Append slot to the first token.
+    auto part_tokens = android::base::Split(part, ":");
 
-    if (fb->GetVar("has-slot:" + part, &has_slot) != fastboot::SUCCESS) {
+    if (fb->GetVar("has-slot:" + part_tokens[0], &has_slot) != fastboot::SUCCESS) {
         /* If has-slot is not supported, the answer is no. */
         has_slot = "no";
     }
@@ -1196,14 +1206,15 @@
             if (current_slot == "") {
                 die("Failed to identify current slot");
             }
-            func(part + "_" + current_slot);
+            part_tokens[0] += "_" + current_slot;
         } else {
-            func(part + '_' + slot);
+            part_tokens[0] += "_" + slot;
         }
+        func(android::base::Join(part_tokens, ":"));
     } else {
         if (force_slot && slot != "") {
-             fprintf(stderr, "Warning: %s does not support slots, and slot %s was requested.\n",
-                     part.c_str(), slot.c_str());
+            fprintf(stderr, "Warning: %s does not support slots, and slot %s was requested.\n",
+                    part_tokens[0].c_str(), slot.c_str());
         }
         func(part);
     }
@@ -1217,10 +1228,13 @@
 static void do_for_partitions(const std::string& part, const std::string& slot,
                               const std::function<void(const std::string&)>& func, bool force_slot) {
     std::string has_slot;
+    // |part| can be vendor_boot:default. Query has-slot on the first token only.
+    auto part_tokens = android::base::Split(part, ":");
 
     if (slot == "all") {
-        if (fb->GetVar("has-slot:" + part, &has_slot) != fastboot::SUCCESS) {
-            die("Could not check if partition %s has slot %s", part.c_str(), slot.c_str());
+        if (fb->GetVar("has-slot:" + part_tokens[0], &has_slot) != fastboot::SUCCESS) {
+            die("Could not check if partition %s has slot %s", part_tokens[0].c_str(),
+                slot.c_str());
         }
         if (has_slot == "yes") {
             for (int i=0; i < get_slot_count(); i++) {
@@ -1247,7 +1261,74 @@
     return android::base::StartsWith(value, "system_");
 }
 
+// Fetch a partition from the device to a given fd. This is a wrapper over FetchToFd to fetch
+// the full image.
+static uint64_t fetch_partition(const std::string& partition, borrowed_fd fd) {
+    uint64_t fetch_size = get_uint_var(FB_VAR_MAX_FETCH_SIZE);
+    if (fetch_size == 0) {
+        die("Unable to get %s. Device does not support fetch command.", FB_VAR_MAX_FETCH_SIZE);
+    }
+    uint64_t partition_size = get_partition_size(partition);
+    if (partition_size <= 0) {
+        die("Invalid partition size for partition %s: %" PRId64, partition.c_str(), partition_size);
+    }
+
+    uint64_t offset = 0;
+    while (offset < partition_size) {
+        uint64_t chunk_size = std::min(fetch_size, partition_size - offset);
+        if (fb->FetchToFd(partition, fd, offset, chunk_size) != fastboot::RetCode::SUCCESS) {
+            die("Unable to fetch %s (offset=%" PRIx64 ", size=%" PRIx64 ")", partition.c_str(),
+                offset, chunk_size);
+        }
+        offset += chunk_size;
+    }
+    return partition_size;
+}
+
+static void do_fetch(const std::string& partition, const std::string& slot_override,
+                     const std::string& outfile) {
+    unique_fd fd(TEMP_FAILURE_RETRY(
+            open(outfile.c_str(), O_WRONLY | O_CREAT | O_TRUNC | O_CLOEXEC | O_BINARY, 0644)));
+    auto fetch = std::bind(fetch_partition, _1, borrowed_fd(fd));
+    do_for_partitions(partition, slot_override, fetch, false /* force slot */);
+}
+
+// Return immediately if not flashing a vendor boot image. If flashing a vendor boot image,
+// repack vendor_boot image with an updated ramdisk. After execution, buf is set
+// to the new image to flash, and return value is the real partition name to flash.
+static std::string repack_ramdisk(const char* pname, struct fastboot_buffer* buf) {
+    std::string_view pname_sv{pname};
+
+    if (!android::base::StartsWith(pname_sv, "vendor_boot:") &&
+        !android::base::StartsWith(pname_sv, "vendor_boot_a:") &&
+        !android::base::StartsWith(pname_sv, "vendor_boot_b:")) {
+        return std::string(pname_sv);
+    }
+    if (buf->type != FB_BUFFER_FD) {
+        die("Flashing sparse vendor ramdisk image is not supported.");
+    }
+    if (buf->sz <= 0) {
+        die("repack_ramdisk() sees negative size: %" PRId64, buf->sz);
+    }
+    std::string partition(pname_sv.substr(0, pname_sv.find(':')));
+    std::string ramdisk(pname_sv.substr(pname_sv.find(':') + 1));
+
+    unique_fd vendor_boot(make_temporary_fd("vendor boot repack"));
+    uint64_t vendor_boot_size = fetch_partition(partition, vendor_boot);
+    auto repack_res = replace_vendor_ramdisk(vendor_boot, vendor_boot_size, ramdisk, buf->fd,
+                                             static_cast<uint64_t>(buf->sz));
+    if (!repack_res.ok()) {
+        die("%s", repack_res.error().message().c_str());
+    }
+
+    buf->fd = std::move(vendor_boot);
+    buf->sz = vendor_boot_size;
+    buf->image_size = vendor_boot_size;
+    return partition;
+}
+
 static void do_flash(const char* pname, const char* fname) {
+    verbose("Do flash %s %s", pname, fname);
     struct fastboot_buffer buf;
 
     if (!load_buf(fname, &buf)) {
@@ -1256,7 +1337,8 @@
     if (is_logical(pname)) {
         fb->ResizePartition(pname, std::to_string(buf.image_size));
     }
-    flash_buf(pname, &buf);
+    std::string flash_pname = repack_ramdisk(pname, &buf);
+    flash_buf(flash_pname, &buf);
 }
 
 // Sets slot_override as the active slot. If slot_override is blank,
@@ -1311,7 +1393,7 @@
   public:
     virtual ~ImageSource() {};
     virtual bool ReadFile(const std::string& name, std::vector<char>* out) const = 0;
-    virtual int OpenFile(const std::string& name) const = 0;
+    virtual unique_fd OpenFile(const std::string& name) const = 0;
 };
 
 class FlashAllTool {
@@ -1429,8 +1511,8 @@
 void FlashAllTool::FlashImages(const std::vector<std::pair<const Image*, std::string>>& images) {
     for (const auto& [image, slot] : images) {
         fastboot_buffer buf;
-        int fd = source_.OpenFile(image->img_name);
-        if (fd < 0 || !load_buf_fd(fd, &buf)) {
+        unique_fd fd = source_.OpenFile(image->img_name);
+        if (fd < 0 || !load_buf_fd(std::move(fd), &buf)) {
             if (image->optional_if_no_image) {
                 continue;
             }
@@ -1457,7 +1539,7 @@
 }
 
 void FlashAllTool::UpdateSuperPartition() {
-    int fd = source_.OpenFile("super_empty.img");
+    unique_fd fd = source_.OpenFile("super_empty.img");
     if (fd < 0) {
         return;
     }
@@ -1495,7 +1577,7 @@
   public:
     explicit ZipImageSource(ZipArchiveHandle zip) : zip_(zip) {}
     bool ReadFile(const std::string& name, std::vector<char>* out) const override;
-    int OpenFile(const std::string& name) const override;
+    unique_fd OpenFile(const std::string& name) const override;
 
   private:
     ZipArchiveHandle zip_;
@@ -1505,7 +1587,7 @@
     return UnzipToMemory(zip_, name, out);
 }
 
-int ZipImageSource::OpenFile(const std::string& name) const {
+unique_fd ZipImageSource::OpenFile(const std::string& name) const {
     return unzip_to_file(zip_, name.c_str());
 }
 
@@ -1525,7 +1607,7 @@
 class LocalImageSource final : public ImageSource {
   public:
     bool ReadFile(const std::string& name, std::vector<char>* out) const override;
-    int OpenFile(const std::string& name) const override;
+    unique_fd OpenFile(const std::string& name) const override;
 };
 
 bool LocalImageSource::ReadFile(const std::string& name, std::vector<char>* out) const {
@@ -1536,9 +1618,9 @@
     return ReadFileToVector(path, out);
 }
 
-int LocalImageSource::OpenFile(const std::string& name) const {
+unique_fd LocalImageSource::OpenFile(const std::string& name) const {
     auto path = find_item_given_name(name);
-    return open(path.c_str(), O_RDONLY | O_BINARY);
+    return unique_fd(TEMP_FAILURE_RETRY(open(path.c_str(), O_RDONLY | O_BINARY)));
 }
 
 static void do_flashall(const std::string& slot_override, bool skip_secondary, bool wipe) {
@@ -1657,7 +1739,7 @@
     if (fd == -1) {
         die("Cannot open generated image: %s", strerror(errno));
     }
-    if (!load_buf_fd(fd.release(), &buf)) {
+    if (!load_buf_fd(std::move(fd), &buf)) {
         die("Cannot read image: %s", strerror(errno));
     }
     flash_buf(partition, &buf);
@@ -2116,7 +2198,7 @@
             if (!load_buf(filename.c_str(), &buf) || buf.type != FB_BUFFER_FD) {
                 die("cannot load '%s'", filename.c_str());
             }
-            fb->Download(filename, buf.fd, buf.sz);
+            fb->Download(filename, buf.fd.get(), buf.sz);
         } else if (command == "get_staged") {
             std::string filename = next_arg(&args);
             fb->Upload(filename);
@@ -2170,6 +2252,10 @@
                 syntax_error("expected: snapshot-update [cancel|merge]");
             }
             fb->SnapshotUpdateCommand(arg);
+        } else if (command == FB_CMD_FETCH) {
+            std::string partition = next_arg(&args);
+            std::string outfile = next_arg(&args);
+            do_fetch(partition, slot_override, outfile);
         } else {
             syntax_error("unknown command %s", command.c_str());
         }
diff --git a/fastboot/fastboot_driver.cpp b/fastboot/fastboot_driver.cpp
index 8d534ea..99a4873 100644
--- a/fastboot/fastboot_driver.cpp
+++ b/fastboot/fastboot_driver.cpp
@@ -30,6 +30,7 @@
 
 #include <errno.h>
 #include <fcntl.h>
+#include <inttypes.h>
 #include <stdio.h>
 #include <stdlib.h>
 #include <string.h>
@@ -42,14 +43,17 @@
 
 #include <android-base/file.h>
 #include <android-base/mapped_file.h>
+#include <android-base/parseint.h>
 #include <android-base/stringprintf.h>
 #include <android-base/strings.h>
 #include <android-base/unique_fd.h>
+#include <storage_literals/storage_literals.h>
 
 #include "constants.h"
 #include "transport.h"
 
 using android::base::StringPrintf;
+using namespace android::storage_literals;
 
 namespace fastboot {
 
@@ -140,7 +144,8 @@
     return Flash(partition);
 }
 
-RetCode FastBootDriver::FlashPartition(const std::string& partition, int fd, uint32_t size) {
+RetCode FastBootDriver::FlashPartition(const std::string& partition, android::base::borrowed_fd fd,
+                                       uint32_t size) {
     RetCode ret;
     if ((ret = Download(partition, fd, size))) {
         return ret;
@@ -178,15 +183,16 @@
     return SUCCESS;
 }
 
-RetCode FastBootDriver::Download(const std::string& name, int fd, size_t size,
-                                 std::string* response, std::vector<std::string>* info) {
+RetCode FastBootDriver::Download(const std::string& name, android::base::borrowed_fd fd,
+                                 size_t size, std::string* response,
+                                 std::vector<std::string>* info) {
     prolog_(StringPrintf("Sending '%s' (%zu KB)", name.c_str(), size / 1024));
     auto result = Download(fd, size, response, info);
     epilog_(result);
     return result;
 }
 
-RetCode FastBootDriver::Download(int fd, size_t size, std::string* response,
+RetCode FastBootDriver::Download(android::base::borrowed_fd fd, size_t size, std::string* response,
                                  std::vector<std::string>* info) {
     RetCode ret;
 
@@ -297,41 +303,85 @@
     return result;
 }
 
-RetCode FastBootDriver::UploadInner(const std::string& outfile, std::string* response,
-                                    std::vector<std::string>* info) {
+// This function executes cmd, then expect a "DATA" response with a number N, followed
+// by N bytes, and another response.
+// This is the common way for the device to send data to the driver used by upload and fetch.
+RetCode FastBootDriver::RunAndReadBuffer(
+        const std::string& cmd, std::string* response, std::vector<std::string>* info,
+        const std::function<RetCode(const char* data, uint64_t size)>& write_fn) {
     RetCode ret;
     int dsize = 0;
-    if ((ret = RawCommand(FB_CMD_UPLOAD, response, info, &dsize))) {
-        error_ = "Upload request failed: " + error_;
+    if ((ret = RawCommand(cmd, response, info, &dsize))) {
+        error_ = android::base::StringPrintf("%s request failed: %s", cmd.c_str(), error_.c_str());
         return ret;
     }
 
-    if (!dsize) {
-        error_ = "Upload request failed, device reports 0 bytes available";
+    if (dsize <= 0) {
+        error_ = android::base::StringPrintf("%s request failed, device reports %d bytes available",
+                                             cmd.c_str(), dsize);
         return BAD_DEV_RESP;
     }
 
-    std::vector<char> data;
-    data.resize(dsize);
-
-    if ((ret = ReadBuffer(data))) {
-        return ret;
+    const uint64_t total_size = dsize;
+    const uint64_t buf_size = std::min<uint64_t>(total_size, 1_MiB);
+    std::vector<char> data(buf_size);
+    uint64_t current_offset = 0;
+    while (current_offset < total_size) {
+        uint64_t remaining = total_size - current_offset;
+        uint64_t chunk_size = std::min(buf_size, remaining);
+        if ((ret = ReadBuffer(data.data(), chunk_size)) != SUCCESS) {
+            return ret;
+        }
+        if ((ret = write_fn(data.data(), chunk_size)) != SUCCESS) {
+            return ret;
+        }
+        current_offset += chunk_size;
     }
+    return HandleResponse(response, info);
+}
 
+RetCode FastBootDriver::UploadInner(const std::string& outfile, std::string* response,
+                                    std::vector<std::string>* info) {
     std::ofstream ofs;
     ofs.open(outfile, std::ofstream::out | std::ofstream::binary);
     if (ofs.fail()) {
         error_ = android::base::StringPrintf("Failed to open '%s'", outfile.c_str());
         return IO_ERROR;
     }
-    ofs.write(data.data(), data.size());
-    if (ofs.fail() || ofs.bad()) {
-        error_ = android::base::StringPrintf("Writing to '%s' failed", outfile.c_str());
-        return IO_ERROR;
-    }
+    auto write_fn = [&](const char* data, uint64_t size) {
+        ofs.write(data, size);
+        if (ofs.fail() || ofs.bad()) {
+            error_ = android::base::StringPrintf("Writing to '%s' failed", outfile.c_str());
+            return IO_ERROR;
+        }
+        return SUCCESS;
+    };
+    RetCode ret = RunAndReadBuffer(FB_CMD_UPLOAD, response, info, write_fn);
     ofs.close();
+    return ret;
+}
 
-    return HandleResponse(response, info);
+RetCode FastBootDriver::FetchToFd(const std::string& partition, android::base::borrowed_fd fd,
+                                  int64_t offset, int64_t size, std::string* response,
+                                  std::vector<std::string>* info) {
+    prolog_(android::base::StringPrintf("Fetching %s (offset=%" PRIx64 ", size=%" PRIx64 ")",
+                                        partition.c_str(), offset, size));
+    std::string cmd = FB_CMD_FETCH ":" + partition;
+    if (offset >= 0) {
+        cmd += android::base::StringPrintf(":0x%08" PRIx64, offset);
+        if (size >= 0) {
+            cmd += android::base::StringPrintf(":0x%08" PRIx64, size);
+        }
+    }
+    RetCode ret = RunAndReadBuffer(cmd, response, info, [&](const char* data, uint64_t size) {
+        if (!android::base::WriteFully(fd, data, size)) {
+            error_ = android::base::StringPrintf("Cannot write: %s", strerror(errno));
+            return IO_ERROR;
+        }
+        return SUCCESS;
+    });
+    epilog_(ret);
+    return ret;
 }
 
 // Helpers
@@ -473,7 +523,7 @@
 }
 
 /******************************* PRIVATE **************************************/
-RetCode FastBootDriver::SendBuffer(int fd, size_t size) {
+RetCode FastBootDriver::SendBuffer(android::base::borrowed_fd fd, size_t size) {
     static constexpr uint32_t MAX_MAP_SIZE = 512 * 1024 * 1024;
     off64_t offset = 0;
     uint32_t remaining = size;
@@ -524,11 +574,6 @@
     return SUCCESS;
 }
 
-RetCode FastBootDriver::ReadBuffer(std::vector<char>& buf) {
-    // Read the buffer
-    return ReadBuffer(buf.data(), buf.size());
-}
-
 RetCode FastBootDriver::ReadBuffer(void* buf, size_t size) {
     // Read the buffer
     ssize_t tmp = transport_->Read(buf, size);
diff --git a/fastboot/fastboot_driver.h b/fastboot/fastboot_driver.h
index 7265632..bccd668 100644
--- a/fastboot/fastboot_driver.h
+++ b/fastboot/fastboot_driver.h
@@ -34,6 +34,7 @@
 
 #include <android-base/logging.h>
 #include <android-base/stringprintf.h>
+#include <android-base/unique_fd.h>
 #include <bootimg.h>
 #include <inttypes.h>
 #include <sparse/sparse.h>
@@ -76,9 +77,9 @@
     RetCode Continue(std::string* response = nullptr, std::vector<std::string>* info = nullptr);
     RetCode CreatePartition(const std::string& partition, const std::string& size);
     RetCode DeletePartition(const std::string& partition);
-    RetCode Download(const std::string& name, int fd, size_t size, std::string* response = nullptr,
-                     std::vector<std::string>* info = nullptr);
-    RetCode Download(int fd, size_t size, std::string* response = nullptr,
+    RetCode Download(const std::string& name, android::base::borrowed_fd fd, size_t size,
+                     std::string* response = nullptr, std::vector<std::string>* info = nullptr);
+    RetCode Download(android::base::borrowed_fd fd, size_t size, std::string* response = nullptr,
                      std::vector<std::string>* info = nullptr);
     RetCode Download(const std::string& name, const std::vector<char>& buf,
                      std::string* response = nullptr, std::vector<std::string>* info = nullptr);
@@ -106,10 +107,14 @@
                    std::vector<std::string>* info = nullptr);
     RetCode SnapshotUpdateCommand(const std::string& command, std::string* response = nullptr,
                                   std::vector<std::string>* info = nullptr);
+    RetCode FetchToFd(const std::string& partition, android::base::borrowed_fd fd,
+                      int64_t offset = -1, int64_t size = -1, std::string* response = nullptr,
+                      std::vector<std::string>* info = nullptr);
 
     /* HIGHER LEVEL COMMANDS -- Composed of the commands above */
     RetCode FlashPartition(const std::string& partition, const std::vector<char>& data);
-    RetCode FlashPartition(const std::string& partition, int fd, uint32_t sz);
+    RetCode FlashPartition(const std::string& partition, android::base::borrowed_fd fd,
+                           uint32_t sz);
     RetCode FlashPartition(const std::string& partition, sparse_file* s, uint32_t sz,
                            size_t current, size_t total);
 
@@ -145,15 +150,17 @@
     Transport* transport_;
 
   private:
-    RetCode SendBuffer(int fd, size_t size);
+    RetCode SendBuffer(android::base::borrowed_fd fd, size_t size);
     RetCode SendBuffer(const std::vector<char>& buf);
     RetCode SendBuffer(const void* buf, size_t size);
 
-    RetCode ReadBuffer(std::vector<char>& buf);
     RetCode ReadBuffer(void* buf, size_t size);
 
     RetCode UploadInner(const std::string& outfile, std::string* response = nullptr,
                         std::vector<std::string>* info = nullptr);
+    RetCode RunAndReadBuffer(const std::string& cmd, std::string* response,
+                             std::vector<std::string>* info,
+                             const std::function<RetCode(const char*, uint64_t)>& write_fn);
 
     int SparseWriteCallback(std::vector<char>& tpbuf, const char* data, size_t len);
 
diff --git a/fastboot/fuzzy_fastboot/main.cpp b/fastboot/fuzzy_fastboot/main.cpp
index 34ab32c..b6beaf9 100644
--- a/fastboot/fuzzy_fastboot/main.cpp
+++ b/fastboot/fuzzy_fastboot/main.cpp
@@ -43,8 +43,10 @@
 #include <thread>
 #include <vector>
 
+#include <android-base/file.h>
 #include <android-base/parseint.h>
 #include <android-base/stringprintf.h>
+#include <android-base/strings.h>
 #include <gtest/gtest.h>
 #include <sparse/sparse.h>
 
@@ -349,22 +351,35 @@
     EXPECT_TRUE(var == "yes" || var == "no") << "getvar:battery-soc-ok must be 'yes' or 'no'";
 }
 
+void AssertHexUint32(const std::string& name, const std::string& var) {
+    ASSERT_NE(var, "") << "getvar:" << name << " responded with empty string";
+    // This must start with 0x
+    ASSERT_FALSE(isspace(var.front()))
+            << "getvar:" << name << " responded with a string with leading whitespace";
+    ASSERT_FALSE(var.compare(0, 2, "0x"))
+            << "getvar:" << name << " responded with a string that does not start with 0x...";
+    int64_t size = strtoll(var.c_str(), nullptr, 16);
+    ASSERT_GT(size, 0) << "'" + var + "' is not a valid response from getvar:" << name;
+    // At most 32-bits
+    ASSERT_LE(size, std::numeric_limits<uint32_t>::max())
+            << "getvar:" << name << " must fit in a uint32_t";
+    ASSERT_LE(var.size(), FB_RESPONSE_SZ - 4)
+            << "getvar:" << name << " responded with too large of string: " + var;
+}
+
 TEST_F(Conformance, GetVarDownloadSize) {
     std::string var;
     EXPECT_EQ(fb->GetVar("max-download-size", &var), SUCCESS) << "getvar:max-download-size failed";
-    EXPECT_NE(var, "") << "getvar:max-download-size responded with empty string";
-    // This must start with 0x
-    EXPECT_FALSE(isspace(var.front()))
-            << "getvar:max-download-size responded with a string with leading whitespace";
-    EXPECT_FALSE(var.compare(0, 2, "0x"))
-            << "getvar:max-download-size responded with a string that does not start with 0x...";
-    int64_t size = strtoll(var.c_str(), nullptr, 16);
-    EXPECT_GT(size, 0) << "'" + var + "' is not a valid response from getvar:max-download-size";
-    // At most 32-bits
-    EXPECT_LE(size, std::numeric_limits<uint32_t>::max())
-            << "getvar:max-download-size must fit in a uint32_t";
-    EXPECT_LE(var.size(), FB_RESPONSE_SZ - 4)
-            << "getvar:max-download-size responded with too large of string: " + var;
+    AssertHexUint32("max-download-size", var);
+}
+
+// If fetch is supported, getvar:max-fetch-size must return a hex string.
+TEST_F(Conformance, GetVarFetchSize) {
+    std::string var;
+    if (SUCCESS != fb->GetVar("max-fetch-size", &var)) {
+        GTEST_SKIP() << "getvar:max-fetch-size failed";
+    }
+    AssertHexUint32("max-fetch-size", var);
 }
 
 TEST_F(Conformance, GetVarAll) {
@@ -656,6 +671,33 @@
     EXPECT_EQ(fb->Partitions(&parts), SUCCESS) << "getvar:all failed in unlocked mode";
 }
 
+// If the implementation supports getvar:max-fetch-size, it must also support fetch:vendor_boot*.
+TEST_F(UnlockPermissions, FetchVendorBoot) {
+    std::string var;
+    uint64_t fetch_size;
+    if (fb->GetVar("max-fetch-size", &var) != SUCCESS) {
+        GTEST_SKIP() << "This test is skipped because fetch is not supported.";
+    }
+    ASSERT_FALSE(var.empty());
+    ASSERT_TRUE(android::base::ParseUint(var, &fetch_size)) << var << " is not an integer";
+    std::vector<std::tuple<std::string, uint64_t>> parts;
+    EXPECT_EQ(fb->Partitions(&parts), SUCCESS) << "getvar:all failed";
+    for (const auto& [partition, partition_size] : parts) {
+        if (!android::base::StartsWith(partition, "vendor_boot")) continue;
+        TemporaryFile fetched;
+
+        uint64_t offset = 0;
+        while (offset < partition_size) {
+            uint64_t chunk_size = std::min(fetch_size, partition_size - offset);
+            auto ret = fb->FetchToFd(partition, fetched.fd, offset, chunk_size);
+            ASSERT_EQ(fastboot::RetCode::SUCCESS, ret)
+                    << "Unable to fetch " << partition << " (offset=" << offset
+                    << ", size=" << chunk_size << ")";
+            offset += chunk_size;
+        }
+    }
+}
+
 TEST_F(LockPermissions, DownloadFlash) {
     std::vector<char> buf{'a', 'o', 's', 'p'};
     EXPECT_EQ(fb->Download(buf), SUCCESS) << "Download failed in locked mode";
@@ -717,6 +759,16 @@
     EXPECT_GT(resp.size(), 0) << "No error message was returned by device after FAIL";
 }
 
+TEST_F(LockPermissions, FetchVendorBoot) {
+    std::vector<std::tuple<std::string, uint64_t>> parts;
+    EXPECT_EQ(fb->Partitions(&parts), SUCCESS) << "getvar:all failed";
+    for (const auto& [partition, _] : parts) {
+        TemporaryFile fetched;
+        ASSERT_EQ(fb->FetchToFd(partition, fetched.fd, 0, 0), DEVICE_FAIL)
+                << "fetch:" << partition << ":0:0 did not fail in locked mode";
+    }
+}
+
 TEST_F(Fuzz, DownloadSize) {
     std::string var;
     EXPECT_EQ(fb->GetVar("max-download-size", &var), SUCCESS) << "getvar:max-download-size failed";
diff --git a/fastboot/testdata/Android.bp b/fastboot/testdata/Android.bp
new file mode 100644
index 0000000..a490fe2
--- /dev/null
+++ b/fastboot/testdata/Android.bp
@@ -0,0 +1,140 @@
+// Copyright (C) 2021 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.
+
+package {
+    default_applicable_licenses: ["Android-Apache-2.0"],
+}
+
+python_binary_host {
+    name: "fastboot_gen_rand",
+    visibility: [":__subpackages__"],
+    srcs: ["fastboot_gen_rand.py"],
+}
+
+genrule_defaults {
+    name: "fastboot_test_data_gen_defaults",
+    visibility: ["//system/core/fastboot"],
+    tools: [
+        "fastboot_gen_rand",
+    ],
+}
+
+// Genrules for components of test vendor boot image.
+
+// Fake dtb image.
+genrule {
+    name: "fastboot_test_dtb",
+    defaults: ["fastboot_test_data_gen_defaults"],
+    out: ["test_dtb.img"],
+    cmd: "$(location fastboot_gen_rand) --seed dtb --length 1024 > $(out)",
+}
+
+// Fake bootconfig image.
+genrule {
+    name: "fastboot_test_bootconfig",
+    defaults: ["fastboot_test_data_gen_defaults"],
+    out: ["test_bootconfig.img"],
+    cmd: "$(location fastboot_gen_rand) --seed bootconfig --length 1024 > $(out)",
+}
+
+// Fake vendor ramdisk with type "none".
+genrule {
+    name: "fastboot_test_vendor_ramdisk_none",
+    defaults: ["fastboot_test_data_gen_defaults"],
+    out: ["test_vendor_ramdisk_none.img"],
+    cmd: "$(location fastboot_gen_rand) --seed vendor_ramdisk_none --length 1024 > $(out)",
+}
+
+// Fake vendor ramdisk with type "platform".
+genrule {
+    name: "fastboot_test_vendor_ramdisk_platform",
+    defaults: ["fastboot_test_data_gen_defaults"],
+    out: ["test_vendor_ramdisk_platform.img"],
+    cmd: "$(location fastboot_gen_rand) --seed vendor_ramdisk_platform --length 1024 > $(out)",
+}
+
+// Fake replacement ramdisk.
+genrule {
+    name: "fastboot_test_vendor_ramdisk_replace",
+    defaults: ["fastboot_test_data_gen_defaults"],
+    out: ["test_vendor_ramdisk_replace.img"],
+    cmd: "$(location fastboot_gen_rand) --seed replace --length 3072 > $(out)",
+}
+
+// Genrules for test vendor boot images.
+
+fastboot_sign_test_image = "$(location avbtool) add_hash_footer --salt 00 --image $(out) " +
+    "--partition_name vendor_boot --partition_size $$(( 1 * 1024 * 1024 ))"
+
+genrule_defaults {
+    name: "fastboot_test_vendor_boot_gen_defaults",
+    defaults: ["fastboot_test_data_gen_defaults"],
+    tools: [
+        "avbtool",
+        "mkbootimg",
+    ],
+}
+
+genrule {
+    name: "fastboot_test_vendor_boot_v3",
+    defaults: ["fastboot_test_vendor_boot_gen_defaults"],
+    out: ["vendor_boot_v3.img"],
+    srcs: [
+        ":fastboot_test_dtb",
+        ":fastboot_test_vendor_ramdisk_none",
+    ],
+    cmd: "$(location mkbootimg) --header_version 3 " +
+        "--vendor_ramdisk $(location :fastboot_test_vendor_ramdisk_none) " +
+        "--dtb $(location :fastboot_test_dtb) " +
+        "--vendor_boot $(out) && " +
+        fastboot_sign_test_image,
+}
+
+genrule {
+    name: "fastboot_test_vendor_boot_v4_without_frag",
+    defaults: ["fastboot_test_vendor_boot_gen_defaults"],
+    out: ["vendor_boot_v4_without_frag.img"],
+    srcs: [
+        ":fastboot_test_dtb",
+        ":fastboot_test_vendor_ramdisk_none",
+        ":fastboot_test_bootconfig",
+    ],
+    cmd: "$(location mkbootimg) --header_version 4 " +
+        "--vendor_ramdisk $(location :fastboot_test_vendor_ramdisk_none) " +
+        "--dtb $(location :fastboot_test_dtb) " +
+        "--vendor_bootconfig $(location :fastboot_test_bootconfig) " +
+        "--vendor_boot $(out) && " +
+        fastboot_sign_test_image,
+}
+
+genrule {
+    name: "fastboot_test_vendor_boot_v4_with_frag",
+    defaults: ["fastboot_test_vendor_boot_gen_defaults"],
+    out: ["vendor_boot_v4_with_frag.img"],
+    srcs: [
+        ":fastboot_test_dtb",
+        ":fastboot_test_vendor_ramdisk_none",
+        ":fastboot_test_vendor_ramdisk_platform",
+        ":fastboot_test_bootconfig",
+    ],
+    cmd: "$(location mkbootimg) --header_version 4 " +
+        "--dtb $(location :fastboot_test_dtb) " +
+        "--vendor_bootconfig $(location :fastboot_test_bootconfig) " +
+        "--ramdisk_type none --ramdisk_name none_ramdisk " +
+        "--vendor_ramdisk_fragment $(location :fastboot_test_vendor_ramdisk_none) " +
+        "--ramdisk_type platform --ramdisk_name platform_ramdisk " +
+        "--vendor_ramdisk_fragment $(location :fastboot_test_vendor_ramdisk_platform) " +
+        "--vendor_boot $(out) && " +
+        fastboot_sign_test_image,
+}
diff --git a/fastboot/testdata/fastboot_gen_rand.py b/fastboot/testdata/fastboot_gen_rand.py
new file mode 100644
index 0000000..a87467b
--- /dev/null
+++ b/fastboot/testdata/fastboot_gen_rand.py
@@ -0,0 +1,32 @@
+#!/usr/bin/env python3
+
+# Copyright (C) 2021 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.
+
+"""
+Write given number of random bytes, generated with optional seed.
+"""
+
+import random, argparse
+
+if __name__ == '__main__':
+  parser = argparse.ArgumentParser(description=__doc__)
+  parser.add_argument('--seed', help='Seed to random generator')
+  parser.add_argument('--length', type=int, required=True, help='Length of output')
+  args = parser.parse_args()
+
+  if args.seed:
+    random.seed(args.seed)
+
+  print(''.join(chr(random.randrange(0,0xff)) for _ in range(args.length)))
diff --git a/fastboot/vendor_boot_img_utils.cpp b/fastboot/vendor_boot_img_utils.cpp
new file mode 100644
index 0000000..2db20cd
--- /dev/null
+++ b/fastboot/vendor_boot_img_utils.cpp
@@ -0,0 +1,422 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#include "vendor_boot_img_utils.h"
+
+#include <string.h>
+
+#include <android-base/file.h>
+#include <android-base/result.h>
+#include <bootimg.h>
+#include <libavb/libavb.h>
+
+namespace {
+
+using android::base::Result;
+
+// Updates a given buffer by creating a new one.
+class DataUpdater {
+  public:
+    DataUpdater(const std::string& old_data) : old_data_(&old_data) {
+        old_data_ptr_ = old_data_->data();
+        new_data_.resize(old_data_->size(), '\0');
+        new_data_ptr_ = new_data_.data();
+    }
+    // Copy |num_bytes| from src to dst.
+    [[nodiscard]] Result<void> Copy(uint32_t num_bytes) {
+        if (num_bytes == 0) return {};
+        if (auto res = CheckAdvance(old_data_ptr_, old_end(), num_bytes, __FUNCTION__); !res.ok())
+            return res;
+        if (auto res = CheckAdvance(new_data_ptr_, new_end(), num_bytes, __FUNCTION__); !res.ok())
+            return res;
+        memcpy(new_data_ptr_, old_data_ptr_, num_bytes);
+        old_data_ptr_ += num_bytes;
+        new_data_ptr_ += num_bytes;
+        return {};
+    }
+    // Replace |old_num_bytes| from src with new data.
+    [[nodiscard]] Result<void> Replace(uint32_t old_num_bytes, const std::string& new_data) {
+        return Replace(old_num_bytes, new_data.data(), new_data.size());
+    }
+    [[nodiscard]] Result<void> Replace(uint32_t old_num_bytes, const void* new_data,
+                                       uint32_t new_data_size) {
+        if (auto res = CheckAdvance(old_data_ptr_, old_end(), old_num_bytes, __FUNCTION__);
+            !res.ok())
+            return res;
+        old_data_ptr_ += old_num_bytes;
+
+        if (new_data_size == 0) return {};
+        if (auto res = CheckAdvance(new_data_ptr_, new_end(), new_data_size, __FUNCTION__);
+            !res.ok())
+            return res;
+        memcpy(new_data_ptr_, new_data, new_data_size);
+        new_data_ptr_ += new_data_size;
+        return {};
+    }
+    // Skip |old_skip| from src and |new_skip| from dst, respectively.
+    [[nodiscard]] Result<void> Skip(uint32_t old_skip, uint32_t new_skip) {
+        if (auto res = CheckAdvance(old_data_ptr_, old_end(), old_skip, __FUNCTION__); !res.ok())
+            return res;
+        old_data_ptr_ += old_skip;
+        if (auto res = CheckAdvance(new_data_ptr_, new_end(), new_skip, __FUNCTION__); !res.ok())
+            return res;
+        new_data_ptr_ += new_skip;
+        return {};
+    }
+
+    [[nodiscard]] Result<void> Seek(uint32_t offset) {
+        if (offset > size()) return Errorf("Cannot seek 0x{:x}, size is 0x{:x}", offset, size());
+        old_data_ptr_ = old_begin() + offset;
+        new_data_ptr_ = new_begin() + offset;
+        return {};
+    }
+
+    std::string Finish() {
+        new_data_ptr_ = nullptr;
+        return std::move(new_data_);
+    }
+
+    [[nodiscard]] Result<void> CheckOffset(uint32_t old_offset, uint32_t new_offset) {
+        if (old_begin() + old_offset != old_cur())
+            return Errorf("Old offset mismatch: expected: 0x{:x}, actual: 0x{:x}", old_offset,
+                          old_cur() - old_begin());
+        if (new_begin() + new_offset != new_cur())
+            return Errorf("New offset mismatch: expected: 0x{:x}, actual: 0x{:x}", new_offset,
+                          new_cur() - new_begin());
+        return {};
+    }
+
+    uint64_t size() const { return old_data_->size(); }
+    const char* old_begin() const { return old_data_->data(); }
+    const char* old_cur() { return old_data_ptr_; }
+    const char* old_end() const { return old_data_->data() + old_data_->size(); }
+    char* new_begin() { return new_data_.data(); }
+    char* new_cur() { return new_data_ptr_; }
+    char* new_end() { return new_data_.data() + new_data_.size(); }
+
+  private:
+    // Check if it is okay to advance |num_bytes| from |current|.
+    [[nodiscard]] Result<void> CheckAdvance(const char* current, const char* end,
+                                            uint32_t num_bytes, const char* op) {
+        auto new_end = current + num_bytes;
+        if (new_end < current /* add overflow */)
+            return Errorf("{}: Addition overflow: 0x{} + 0x{:x} < 0x{}", op, fmt::ptr(current),
+                          num_bytes, fmt::ptr(current));
+        if (new_end > end)
+            return Errorf("{}: Boundary overflow: 0x{} + 0x{:x} > 0x{}", op, fmt::ptr(current),
+                          num_bytes, fmt::ptr(end));
+        return {};
+    }
+    const std::string* old_data_;
+    std::string new_data_;
+    const char* old_data_ptr_;
+    char* new_data_ptr_;
+};
+
+// Get the size of vendor boot header.
+[[nodiscard]] Result<uint32_t> get_vendor_boot_header_size(const vendor_boot_img_hdr_v3* hdr) {
+    if (hdr->header_version == 3) return sizeof(vendor_boot_img_hdr_v3);
+    if (hdr->header_version == 4) return sizeof(vendor_boot_img_hdr_v4);
+    return Errorf("Unrecognized vendor boot header version {}", hdr->header_version);
+}
+
+// Check that content contains a valid vendor boot image header with a version at least |version|.
+[[nodiscard]] Result<void> check_vendor_boot_hdr(const std::string& content, uint32_t version) {
+    // get_vendor_boot_header_size reads header_version, so make sure reading it does not
+    // go out of bounds by ensuring that the content has at least the size of V3 header.
+    if (content.size() < sizeof(vendor_boot_img_hdr_v3)) {
+        return Errorf("Size of vendor boot is 0x{:x}, less than size of V3 header: 0x{:x}",
+                      content.size(), sizeof(vendor_boot_img_hdr_v3));
+    }
+    // Now read hdr->header_version and assert the size.
+    auto hdr = reinterpret_cast<const vendor_boot_img_hdr_v3*>(content.data());
+    auto expect_header_size = get_vendor_boot_header_size(hdr);
+    if (!expect_header_size.ok()) return expect_header_size.error();
+    if (content.size() < *expect_header_size) {
+        return Errorf("Size of vendor boot is 0x{:x}, less than size of V{} header: 0x{:x}",
+                      content.size(), version, *expect_header_size);
+    }
+    if (memcmp(hdr->magic, VENDOR_BOOT_MAGIC, VENDOR_BOOT_MAGIC_SIZE) != 0) {
+        return Errorf("Vendor boot image magic mismatch");
+    }
+    if (hdr->header_version < version) {
+        return Errorf("Require vendor boot header V{} but is V{}", version, hdr->header_version);
+    }
+    return {};
+}
+
+// Wrapper of ReadFdToString. Seek to the beginning and read the whole file to string.
+[[nodiscard]] Result<std::string> load_file(android::base::borrowed_fd fd, uint64_t expected_size,
+                                            const char* what) {
+    if (lseek(fd.get(), 0, SEEK_SET) != 0) {
+        return ErrnoErrorf("Can't seek to the beginning of {} image", what);
+    }
+    std::string content;
+    if (!android::base::ReadFdToString(fd, &content)) {
+        return ErrnoErrorf("Cannot read {} to string", what);
+    }
+    if (content.size() != expected_size) {
+        return Errorf("Size of {} does not match, expected 0x{:x}, read 0x{:x}", what,
+                      expected_size, content.size());
+    }
+    return content;
+}
+
+// Wrapper of WriteStringToFd. Seek to the beginning and write the whole file to string.
+[[nodiscard]] Result<void> store_file(android::base::borrowed_fd fd, const std::string& data,
+                                      const char* what) {
+    if (lseek(fd.get(), 0, SEEK_SET) != 0) {
+        return ErrnoErrorf("Cannot seek to beginning of {} before writing", what);
+    }
+    if (!android::base::WriteStringToFd(data, fd)) {
+        return ErrnoErrorf("Cannot write new content to {}", what);
+    }
+    if (TEMP_FAILURE_RETRY(ftruncate64(fd.get(), data.size())) == -1) {
+        return ErrnoErrorf("Truncating new vendor boot image to 0x{:x} fails", data.size());
+    }
+    return {};
+}
+
+// Copy AVB footer if it exists in the old buffer.
+[[nodiscard]] Result<void> copy_avb_footer(DataUpdater* updater) {
+    if (updater->size() < AVB_FOOTER_SIZE) return {};
+    if (auto res = updater->Seek(updater->size() - AVB_FOOTER_SIZE); !res.ok()) return res;
+    if (memcmp(updater->old_cur(), AVB_FOOTER_MAGIC, AVB_FOOTER_MAGIC_LEN) != 0) return {};
+    return updater->Copy(AVB_FOOTER_SIZE);
+}
+
+// round |value| up to a multiple of |page_size|.
+inline uint32_t round_up(uint32_t value, uint32_t page_size) {
+    return (value + page_size - 1) / page_size * page_size;
+}
+
+// Replace the vendor ramdisk as a whole.
+[[nodiscard]] Result<std::string> replace_default_vendor_ramdisk(const std::string& vendor_boot,
+                                                                 const std::string& new_ramdisk) {
+    if (auto res = check_vendor_boot_hdr(vendor_boot, 3); !res.ok()) return res.error();
+    auto hdr = reinterpret_cast<const vendor_boot_img_hdr_v3*>(vendor_boot.data());
+    auto hdr_size = get_vendor_boot_header_size(hdr);
+    if (!hdr_size.ok()) return hdr_size.error();
+    // Refer to bootimg.h for details. Numbers are in bytes.
+    const uint32_t o = round_up(*hdr_size, hdr->page_size);
+    const uint32_t p = round_up(hdr->vendor_ramdisk_size, hdr->page_size);
+    const uint32_t q = round_up(hdr->dtb_size, hdr->page_size);
+
+    DataUpdater updater(vendor_boot);
+
+    // Copy header (O bytes), then update fields in header.
+    if (auto res = updater.Copy(o); !res.ok()) return res.error();
+    auto new_hdr = reinterpret_cast<vendor_boot_img_hdr_v3*>(updater.new_begin());
+    new_hdr->vendor_ramdisk_size = new_ramdisk.size();
+    // Because it is unknown how the new ramdisk is fragmented, the whole table is replaced
+    // with a single entry representing the full ramdisk.
+    if (new_hdr->header_version >= 4) {
+        auto new_hdr_v4 = static_cast<vendor_boot_img_hdr_v4*>(new_hdr);
+        new_hdr_v4->vendor_ramdisk_table_entry_size = sizeof(vendor_ramdisk_table_entry_v4);
+        new_hdr_v4->vendor_ramdisk_table_entry_num = 1;
+        new_hdr_v4->vendor_ramdisk_table_size = new_hdr_v4->vendor_ramdisk_table_entry_num *
+                                                new_hdr_v4->vendor_ramdisk_table_entry_size;
+    }
+
+    // Copy the new ramdisk.
+    if (auto res = updater.Replace(hdr->vendor_ramdisk_size, new_ramdisk); !res.ok())
+        return res.error();
+    const uint32_t new_p = round_up(new_hdr->vendor_ramdisk_size, new_hdr->page_size);
+    if (auto res = updater.Skip(p - hdr->vendor_ramdisk_size, new_p - new_hdr->vendor_ramdisk_size);
+        !res.ok())
+        return res.error();
+    if (auto res = updater.CheckOffset(o + p, o + new_p); !res.ok()) return res.error();
+
+    // Copy DTB (Q bytes).
+    if (auto res = updater.Copy(q); !res.ok()) return res.error();
+
+    if (new_hdr->header_version >= 4) {
+        auto hdr_v4 = static_cast<const vendor_boot_img_hdr_v4*>(hdr);
+        const uint32_t r = round_up(hdr_v4->vendor_ramdisk_table_size, hdr_v4->page_size);
+        const uint32_t s = round_up(hdr_v4->bootconfig_size, hdr_v4->page_size);
+
+        auto new_entry = reinterpret_cast<vendor_ramdisk_table_entry_v4*>(updater.new_cur());
+        auto new_hdr_v4 = static_cast<const vendor_boot_img_hdr_v4*>(new_hdr);
+        auto new_r = round_up(new_hdr_v4->vendor_ramdisk_table_size, new_hdr->page_size);
+        if (auto res = updater.Skip(r, new_r); !res.ok()) return res.error();
+        if (auto res = updater.CheckOffset(o + p + q + r, o + new_p + q + new_r); !res.ok())
+            return res.error();
+
+        // Replace table with single entry representing the full ramdisk.
+        new_entry->ramdisk_size = new_hdr->vendor_ramdisk_size;
+        new_entry->ramdisk_offset = 0;
+        new_entry->ramdisk_type = VENDOR_RAMDISK_TYPE_NONE;
+        memset(new_entry->ramdisk_name, '\0', VENDOR_RAMDISK_NAME_SIZE);
+        memset(new_entry->board_id, '\0', VENDOR_RAMDISK_TABLE_ENTRY_BOARD_ID_SIZE);
+
+        // Copy bootconfig (S bytes).
+        if (auto res = updater.Copy(s); !res.ok()) return res.error();
+    }
+
+    if (auto res = copy_avb_footer(&updater); !res.ok()) return res.error();
+    return updater.Finish();
+}
+
+// Find a ramdisk fragment with a unique name. Abort if none or multiple fragments are found.
+[[nodiscard]] Result<const vendor_ramdisk_table_entry_v4*> find_unique_ramdisk(
+        const std::string& ramdisk_name, const vendor_ramdisk_table_entry_v4* table,
+        uint32_t size) {
+    const vendor_ramdisk_table_entry_v4* ret = nullptr;
+    uint32_t idx = 0;
+    const vendor_ramdisk_table_entry_v4* entry = table;
+    for (; idx < size; idx++, entry++) {
+        auto entry_name_c_str = reinterpret_cast<const char*>(entry->ramdisk_name);
+        auto entry_name_len = strnlen(entry_name_c_str, VENDOR_RAMDISK_NAME_SIZE);
+        std::string_view entry_name(entry_name_c_str, entry_name_len);
+        if (entry_name == ramdisk_name) {
+            if (ret != nullptr) {
+                return Errorf("Multiple vendor ramdisk '{}' found, name should be unique",
+                              ramdisk_name.c_str());
+            }
+            ret = entry;
+        }
+    }
+    if (ret == nullptr) {
+        return Errorf("Vendor ramdisk '{}' not found", ramdisk_name.c_str());
+    }
+    return ret;
+}
+
+// Find the vendor ramdisk fragment with |ramdisk_name| within the content of |vendor_boot|, and
+// replace it with the content of |new_ramdisk|.
+[[nodiscard]] Result<std::string> replace_vendor_ramdisk_fragment(const std::string& ramdisk_name,
+                                                                  const std::string& vendor_boot,
+                                                                  const std::string& new_ramdisk) {
+    if (auto res = check_vendor_boot_hdr(vendor_boot, 4); !res.ok()) return res.error();
+    auto hdr = reinterpret_cast<const vendor_boot_img_hdr_v4*>(vendor_boot.data());
+    auto hdr_size = get_vendor_boot_header_size(hdr);
+    if (!hdr_size.ok()) return hdr_size.error();
+    // Refer to bootimg.h for details. Numbers are in bytes.
+    const uint32_t o = round_up(*hdr_size, hdr->page_size);
+    const uint32_t p = round_up(hdr->vendor_ramdisk_size, hdr->page_size);
+    const uint32_t q = round_up(hdr->dtb_size, hdr->page_size);
+    const uint32_t r = round_up(hdr->vendor_ramdisk_table_size, hdr->page_size);
+    const uint32_t s = round_up(hdr->bootconfig_size, hdr->page_size);
+
+    if (hdr->vendor_ramdisk_table_entry_num == std::numeric_limits<uint32_t>::max()) {
+        return Errorf("Too many vendor ramdisk entries in table, overflow");
+    }
+
+    // Find entry with name |ramdisk_name|.
+    auto old_table_start =
+            reinterpret_cast<const vendor_ramdisk_table_entry_v4*>(vendor_boot.data() + o + p + q);
+    auto find_res =
+            find_unique_ramdisk(ramdisk_name, old_table_start, hdr->vendor_ramdisk_table_entry_num);
+    if (!find_res.ok()) return find_res.error();
+    const vendor_ramdisk_table_entry_v4* replace_entry = *find_res;
+    uint32_t replace_idx = replace_entry - old_table_start;
+
+    // Now reconstruct.
+    DataUpdater updater(vendor_boot);
+
+    // Copy header (O bytes), then update fields in header.
+    if (auto res = updater.Copy(o); !res.ok()) return res.error();
+    auto new_hdr = reinterpret_cast<vendor_boot_img_hdr_v4*>(updater.new_begin());
+
+    // Copy ramdisk fragments, replace for the matching index.
+    {
+        auto old_ramdisk_entry = reinterpret_cast<const vendor_ramdisk_table_entry_v4*>(
+                vendor_boot.data() + o + p + q);
+        uint32_t new_total_ramdisk_size = 0;
+        for (uint32_t new_ramdisk_idx = 0; new_ramdisk_idx < hdr->vendor_ramdisk_table_entry_num;
+             new_ramdisk_idx++, old_ramdisk_entry++) {
+            if (new_ramdisk_idx == replace_idx) {
+                if (auto res = updater.Replace(replace_entry->ramdisk_size, new_ramdisk); !res.ok())
+                    return res.error();
+                new_total_ramdisk_size += new_ramdisk.size();
+            } else {
+                if (auto res = updater.Copy(old_ramdisk_entry->ramdisk_size); !res.ok())
+                    return res.error();
+                new_total_ramdisk_size += old_ramdisk_entry->ramdisk_size;
+            }
+        }
+        new_hdr->vendor_ramdisk_size = new_total_ramdisk_size;
+    }
+
+    // Pad ramdisk to page boundary.
+    const uint32_t new_p = round_up(new_hdr->vendor_ramdisk_size, new_hdr->page_size);
+    if (auto res = updater.Skip(p - hdr->vendor_ramdisk_size, new_p - new_hdr->vendor_ramdisk_size);
+        !res.ok())
+        return res.error();
+    if (auto res = updater.CheckOffset(o + p, o + new_p); !res.ok()) return res.error();
+
+    // Copy DTB (Q bytes).
+    if (auto res = updater.Copy(q); !res.ok()) return res.error();
+
+    // Copy table, but with corresponding entries modified, including:
+    // - ramdisk_size of the entry replaced
+    // - ramdisk_offset of subsequent entries.
+    for (uint32_t new_total_ramdisk_size = 0, new_entry_idx = 0;
+         new_entry_idx < hdr->vendor_ramdisk_table_entry_num; new_entry_idx++) {
+        auto new_entry = reinterpret_cast<vendor_ramdisk_table_entry_v4*>(updater.new_cur());
+        if (auto res = updater.Copy(hdr->vendor_ramdisk_table_entry_size); !res.ok())
+            return res.error();
+        new_entry->ramdisk_offset = new_total_ramdisk_size;
+
+        if (new_entry_idx == replace_idx) {
+            new_entry->ramdisk_size = new_ramdisk.size();
+        }
+        new_total_ramdisk_size += new_entry->ramdisk_size;
+    }
+
+    // Copy padding of R pages; this is okay because table size is not changed.
+    if (auto res = updater.Copy(r - hdr->vendor_ramdisk_table_entry_num *
+                                            hdr->vendor_ramdisk_table_entry_size);
+        !res.ok())
+        return res.error();
+    if (auto res = updater.CheckOffset(o + p + q + r, o + new_p + q + r); !res.ok())
+        return res.error();
+
+    // Copy bootconfig (S bytes).
+    if (auto res = updater.Copy(s); !res.ok()) return res.error();
+
+    if (auto res = copy_avb_footer(&updater); !res.ok()) return res.error();
+    return updater.Finish();
+}
+
+}  // namespace
+
+[[nodiscard]] Result<void> replace_vendor_ramdisk(android::base::borrowed_fd vendor_boot_fd,
+                                                  uint64_t vendor_boot_size,
+                                                  const std::string& ramdisk_name,
+                                                  android::base::borrowed_fd new_ramdisk_fd,
+                                                  uint64_t new_ramdisk_size) {
+    if (new_ramdisk_size > std::numeric_limits<uint32_t>::max()) {
+        return Errorf("New vendor ramdisk is too big");
+    }
+
+    auto vendor_boot = load_file(vendor_boot_fd, vendor_boot_size, "vendor boot");
+    if (!vendor_boot.ok()) return vendor_boot.error();
+    auto new_ramdisk = load_file(new_ramdisk_fd, new_ramdisk_size, "new vendor ramdisk");
+    if (!new_ramdisk.ok()) return new_ramdisk.error();
+
+    Result<std::string> new_vendor_boot;
+    if (ramdisk_name == "default") {
+        new_vendor_boot = replace_default_vendor_ramdisk(*vendor_boot, *new_ramdisk);
+    } else {
+        new_vendor_boot = replace_vendor_ramdisk_fragment(ramdisk_name, *vendor_boot, *new_ramdisk);
+    }
+    if (!new_vendor_boot.ok()) return new_vendor_boot.error();
+    if (auto res = store_file(vendor_boot_fd, *new_vendor_boot, "new vendor boot image"); !res.ok())
+        return res.error();
+
+    return {};
+}
diff --git a/fastboot/vendor_boot_img_utils.h b/fastboot/vendor_boot_img_utils.h
new file mode 100644
index 0000000..0b702bc
--- /dev/null
+++ b/fastboot/vendor_boot_img_utils.h
@@ -0,0 +1,34 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#pragma once
+
+#include <inttypes.h>
+
+#include <string>
+
+#include <android-base/result.h>
+#include <android-base/unique_fd.h>
+
+// Replace the vendor ramdisk named |ramdisk_name| within the vendor boot image,
+// specified by |vendor_boot_fd|, with the ramdisk specified by |new_ramdisk_fd|. Checks
+// that the size of the files are |vendor_boot_size| and |new_ramdisk_size|, respectively.
+// If |ramdisk_name| is "default", replace the vendor ramdisk as a whole. Otherwise, replace
+// a vendor ramdisk fragment with the given unique name.
+[[nodiscard]] android::base::Result<void> replace_vendor_ramdisk(
+        android::base::borrowed_fd vendor_boot_fd, uint64_t vendor_boot_size,
+        const std::string& ramdisk_name, android::base::borrowed_fd new_ramdisk_fd,
+        uint64_t new_ramdisk_size);
diff --git a/fastboot/vendor_boot_img_utils_test.cpp b/fastboot/vendor_boot_img_utils_test.cpp
new file mode 100644
index 0000000..1563b89
--- /dev/null
+++ b/fastboot/vendor_boot_img_utils_test.cpp
@@ -0,0 +1,429 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#include <string.h>
+#include <sys/stat.h>
+#include <sys/types.h>
+#include <unistd.h>
+
+#include <filesystem>
+#include <optional>
+#include <string_view>
+
+#include <android-base/file.h>
+#include <android-base/logging.h>
+#include <android-base/result.h>
+#include <android-base/strings.h>
+#include <bootimg.h>
+#include <gmock/gmock.h>
+#include <gtest/gtest.h>
+#include <libavb/libavb.h>
+
+#include "vendor_boot_img_utils.h"
+
+using android::base::borrowed_fd;
+using android::base::ErrnoError;
+using android::base::GetExecutableDirectory;
+using android::base::ReadFdToString;
+using android::base::Result;
+using testing::AllOf;
+using testing::Each;
+using testing::Eq;
+using testing::HasSubstr;
+using testing::Not;
+using testing::Property;
+using std::string_literals::operator""s;
+
+// Expect that the Result<T> returned by |expr| is successful, and value matches |result_matcher|.
+#define EXPECT_RESULT(expr, result_matcher)                          \
+    EXPECT_THAT(expr, AllOf(Property(&decltype(expr)::ok, Eq(true)), \
+                            Property(&decltype(expr)::value, result_matcher)))
+
+// Expect that the Result<T> returned by |expr| fails, and error message matches |error_matcher|.
+#define EXPECT_ERROR(expr, error_matcher)                                                        \
+    do {                                                                                         \
+        EXPECT_THAT(                                                                             \
+                expr,                                                                            \
+                AllOf(Property(&decltype(expr)::ok, Eq(false)),                                  \
+                      Property(&decltype(expr)::error,                                           \
+                               Property(&decltype(expr)::error_type::message, error_matcher)))); \
+    } while (0)
+
+namespace {
+
+// Wrapper of fstat.
+Result<uint64_t> FileSize(borrowed_fd fd, std::filesystem::path path) {
+    struct stat sb;
+    if (fstat(fd.get(), &sb) == -1) return ErrnoError() << "fstat(" << path << ")";
+    return sb.st_size;
+}
+
+// Seek to beginning then read the whole file.
+Result<std::string> ReadStartOfFdToString(borrowed_fd fd, std::filesystem::path path) {
+    if (lseek64(fd.get(), 0, SEEK_SET) != 0)
+        return ErrnoError() << "lseek64(" << path << ", 0, SEEK_SET)";
+    std::string content;
+    if (!android::base::ReadFdToString(fd, &content)) return ErrnoError() << "read(" << path << ")";
+    return content;
+}
+
+// Round |value| up to page boundary.
+inline uint32_t round_up(uint32_t value, uint32_t page_size) {
+    return (value + page_size - 1) / page_size * page_size;
+}
+
+// Match is successful if |arg| is a zero-padded version of |expected|.
+MATCHER_P(IsPadded, expected, (negation ? "is" : "isn't") + " zero-padded of expected value"s) {
+    if (arg.size() < expected.size()) return false;
+    if (0 != memcmp(arg.data(), expected.data(), expected.size())) return false;
+    auto remainder = std::string_view(arg).substr(expected.size());
+    for (char e : remainder)
+        if (e != '\0') return false;
+    return true;
+}
+
+// Same as Eq, but don't print the content to avoid spam.
+MATCHER_P(MemEq, expected, (negation ? "is" : "isn't") + " expected value"s) {
+    if (arg.size() != expected.size()) return false;
+    return 0 == memcmp(arg.data(), expected.data(), expected.size());
+}
+
+// Expect that |arg| and |expected| has the same AVB footer.
+MATCHER_P(HasSameAvbFooter, expected,
+          (negation ? "has" : "does not have") + "expected AVB footer"s) {
+    if (expected.size() < AVB_FOOTER_SIZE || arg.size() < AVB_FOOTER_SIZE) return false;
+    return std::string_view(expected).substr(expected.size() - AVB_FOOTER_SIZE) ==
+           std::string_view(arg).substr(arg.size() - AVB_FOOTER_SIZE);
+}
+
+// A lazy handle of a file.
+struct TestFileHandle {
+    virtual ~TestFileHandle() = default;
+    // Lazily call OpenImpl(), cache result in open_result_.
+    android::base::Result<void> Open() {
+        if (!open_result_.has_value()) open_result_ = OpenImpl();
+        return open_result_.value();
+    }
+    // The original size at the time when the file is opened. If the file has been modified,
+    // this field is NOT updated.
+    uint64_t size() {
+        CHECK(open_result_.has_value());
+        return size_;
+    }
+    // The current size of the file. If the file has been modified since opened,
+    // this is updated.
+    Result<uint64_t> fsize() {
+        CHECK(open_result_.has_value());
+        return FileSize(fd_, abs_path_);
+    }
+    borrowed_fd fd() {
+        CHECK(open_result_.has_value());
+        return fd_;
+    }
+    Result<std::string> Read() {
+        CHECK(open_result_.has_value());
+        return ReadStartOfFdToString(fd_, abs_path_);
+    }
+
+  private:
+    std::filesystem::path abs_path_;
+    uint64_t size_;
+    std::optional<android::base::Result<void>> open_result_;
+    borrowed_fd fd_{-1};
+    // Opens |rel_path_| as a readonly fd, pass it to Transform, and store result to
+    // |borrowed_fd_|.
+    android::base::Result<void> OpenImpl() {
+        android::base::unique_fd read_fd(TEMP_FAILURE_RETRY(
+                open(abs_path_.c_str(), O_RDONLY | O_CLOEXEC | O_NOFOLLOW | O_BINARY)));
+        if (!read_fd.ok()) return ErrnoError() << "open(" << abs_path_ << ")";
+        auto size = FileSize(read_fd, abs_path_);
+        if (!size.ok()) return size.error();
+        size_ = *size;
+
+        auto borrowed_fd = Transform(abs_path_, std::move(read_fd));
+        if (!borrowed_fd.ok()) return borrowed_fd.error();
+        fd_ = borrowed_fd.value();
+
+        return {};
+    }
+
+  protected:
+    // |rel_path| is the relative path under test data directory.
+    TestFileHandle(const std::filesystem::path& rel_path)
+        : abs_path_(std::move(std::filesystem::path(GetExecutableDirectory()) / rel_path)) {}
+    // Given |read_fd|, the readonly fd on the test file, return an fd that's suitable for client
+    // to use. Implementation is responsible for managing the lifetime of the returned fd.
+    virtual android::base::Result<borrowed_fd> Transform(const std::filesystem::path& abs_path,
+                                                         android::base::unique_fd read_fd) = 0;
+};
+
+// A TestFileHandle where the file is readonly.
+struct ReadOnlyTestFileHandle : TestFileHandle {
+    ReadOnlyTestFileHandle(const std::filesystem::path& rel_path) : TestFileHandle(rel_path) {}
+
+  private:
+    android::base::unique_fd owned_fd_;
+    android::base::Result<borrowed_fd> Transform(const std::filesystem::path&,
+                                                 android::base::unique_fd read_fd) override {
+        owned_fd_ = std::move(read_fd);
+        return owned_fd_;
+    }
+};
+
+// A TestFileHandle where the test file is copies, hence read-writable.
+struct ReadWriteTestFileHandle : TestFileHandle {
+    ReadWriteTestFileHandle(const std::filesystem::path& rel_path) : TestFileHandle(rel_path) {}
+
+  private:
+    std::unique_ptr<TemporaryFile> temp_file_;
+
+    android::base::Result<borrowed_fd> Transform(const std::filesystem::path& abs_path,
+                                                 android::base::unique_fd read_fd) override {
+        // Make a copy to avoid writing to test data. Test files are small, so it is okay
+        // to read the whole file.
+        auto content = ReadStartOfFdToString(read_fd, abs_path);
+        if (!content.ok()) return content.error();
+        temp_file_ = std::make_unique<TemporaryFile>();
+        if (temp_file_->fd == -1)
+            return ErrnoError() << "copy " << abs_path << ": open temp file failed";
+        if (!android::base::WriteStringToFd(*content, temp_file_->fd))
+            return ErrnoError() << "copy " << abs_path << ": write temp file failed";
+
+        return temp_file_->fd;
+    }
+};
+
+class RepackVendorBootImgTestEnv : public ::testing::Environment {
+  public:
+    virtual void SetUp() {
+        OpenTestFile("test_dtb.img", &dtb, &dtb_content);
+        OpenTestFile("test_bootconfig.img", &bootconfig, &bootconfig_content);
+        OpenTestFile("test_vendor_ramdisk_none.img", &none, &none_content);
+        OpenTestFile("test_vendor_ramdisk_platform.img", &platform, &platform_content);
+        OpenTestFile("test_vendor_ramdisk_replace.img", &replace, &replace_content);
+    }
+
+    std::unique_ptr<TestFileHandle> dtb;
+    std::string dtb_content;
+    std::unique_ptr<TestFileHandle> bootconfig;
+    std::string bootconfig_content;
+    std::unique_ptr<TestFileHandle> none;
+    std::string none_content;
+    std::unique_ptr<TestFileHandle> platform;
+    std::string platform_content;
+    std::unique_ptr<TestFileHandle> replace;
+    std::string replace_content;
+
+  private:
+    void OpenTestFile(const char* rel_path, std::unique_ptr<TestFileHandle>* handle,
+                      std::string* content) {
+        *handle = std::make_unique<ReadOnlyTestFileHandle>(rel_path);
+        ASSERT_RESULT_OK((*handle)->Open());
+        auto content_res = (*handle)->Read();
+        ASSERT_RESULT_OK(content_res);
+        *content = *content_res;
+    }
+};
+RepackVendorBootImgTestEnv* env = nullptr;
+
+struct RepackVendorBootImgTestParam {
+    std::string vendor_boot_file_name;
+    uint32_t expected_header_version;
+    friend std::ostream& operator<<(std::ostream& os, const RepackVendorBootImgTestParam& param) {
+        return os << param.vendor_boot_file_name;
+    }
+};
+
+class RepackVendorBootImgTest : public ::testing::TestWithParam<RepackVendorBootImgTestParam> {
+  public:
+    virtual void SetUp() {
+        vboot = std::make_unique<ReadWriteTestFileHandle>(GetParam().vendor_boot_file_name);
+        ASSERT_RESULT_OK(vboot->Open());
+    }
+    std::unique_ptr<TestFileHandle> vboot;
+};
+
+TEST_P(RepackVendorBootImgTest, InvalidSize) {
+    EXPECT_ERROR(replace_vendor_ramdisk(vboot->fd(), vboot->size() + 1, "default",
+                                        env->replace->fd(), env->replace->size()),
+                 HasSubstr("Size of vendor boot does not match"));
+    EXPECT_ERROR(replace_vendor_ramdisk(vboot->fd(), vboot->size(), "default", env->replace->fd(),
+                                        env->replace->size() + 1),
+                 HasSubstr("Size of new vendor ramdisk does not match"));
+}
+
+TEST_P(RepackVendorBootImgTest, ReplaceUnknown) {
+    auto res = replace_vendor_ramdisk(vboot->fd(), vboot->size(), "unknown", env->replace->fd(),
+                                      env->replace->size());
+    if (GetParam().expected_header_version == 3) {
+        EXPECT_ERROR(res, Eq("Require vendor boot header V4 but is V3"));
+    } else if (GetParam().expected_header_version == 4) {
+        EXPECT_ERROR(res, Eq("Vendor ramdisk 'unknown' not found"));
+    }
+}
+
+TEST_P(RepackVendorBootImgTest, ReplaceDefault) {
+    auto old_content = vboot->Read();
+    ASSERT_RESULT_OK(old_content);
+
+    ASSERT_RESULT_OK(replace_vendor_ramdisk(vboot->fd(), vboot->size(), "default",
+                                            env->replace->fd(), env->replace->size()));
+    EXPECT_RESULT(vboot->fsize(), vboot->size()) << "File size should not change after repack";
+
+    auto new_content_res = vboot->Read();
+    ASSERT_RESULT_OK(new_content_res);
+    std::string_view new_content(*new_content_res);
+
+    auto hdr = reinterpret_cast<const vendor_boot_img_hdr_v3*>(new_content.data());
+    ASSERT_EQ(0, memcmp(VENDOR_BOOT_MAGIC, hdr->magic, VENDOR_BOOT_MAGIC_SIZE));
+    ASSERT_EQ(GetParam().expected_header_version, hdr->header_version);
+    EXPECT_EQ(hdr->vendor_ramdisk_size, env->replace->size());
+    EXPECT_EQ(hdr->dtb_size, env->dtb->size());
+
+    auto o = round_up(sizeof(vendor_boot_img_hdr_v3), hdr->page_size);
+    auto p = round_up(hdr->vendor_ramdisk_size, hdr->page_size);
+    auto q = round_up(hdr->dtb_size, hdr->page_size);
+
+    EXPECT_THAT(new_content.substr(o, p), IsPadded(env->replace_content));
+    EXPECT_THAT(new_content.substr(o + p, q), IsPadded(env->dtb_content));
+
+    if (hdr->header_version < 4) return;
+
+    auto hdr_v4 = static_cast<const vendor_boot_img_hdr_v4*>(hdr);
+    EXPECT_EQ(hdr_v4->vendor_ramdisk_table_entry_num, 1);
+    EXPECT_EQ(hdr_v4->vendor_ramdisk_table_size, 1 * hdr_v4->vendor_ramdisk_table_entry_size);
+    EXPECT_GE(hdr_v4->vendor_ramdisk_table_entry_size, sizeof(vendor_ramdisk_table_entry_v4));
+    auto entry = reinterpret_cast<const vendor_ramdisk_table_entry_v4*>(&new_content[o + p + q]);
+    EXPECT_EQ(entry->ramdisk_offset, 0);
+    EXPECT_EQ(entry->ramdisk_size, hdr_v4->vendor_ramdisk_size);
+    EXPECT_EQ(entry->ramdisk_type, VENDOR_RAMDISK_TYPE_NONE);
+
+    EXPECT_EQ(hdr_v4->bootconfig_size, env->bootconfig->size());
+    auto r = round_up(hdr_v4->vendor_ramdisk_table_size, hdr_v4->page_size);
+    auto s = round_up(hdr_v4->bootconfig_size, hdr_v4->page_size);
+    EXPECT_THAT(new_content.substr(o + p + q + r, s), IsPadded(env->bootconfig_content));
+
+    EXPECT_THAT(new_content, HasSameAvbFooter(*old_content));
+}
+
+INSTANTIATE_TEST_SUITE_P(
+        RepackVendorBootImgTest, RepackVendorBootImgTest,
+        ::testing::Values(RepackVendorBootImgTestParam{"vendor_boot_v3.img", 3},
+                          RepackVendorBootImgTestParam{"vendor_boot_v4_with_frag.img", 4},
+                          RepackVendorBootImgTestParam{"vendor_boot_v4_without_frag.img", 4}),
+        [](const auto& info) {
+            return android::base::StringReplace(info.param.vendor_boot_file_name, ".", "_", false);
+        });
+
+std::string_view GetRamdiskName(const vendor_ramdisk_table_entry_v4* entry) {
+    auto ramdisk_name = reinterpret_cast<const char*>(entry->ramdisk_name);
+    return std::string_view(ramdisk_name, strnlen(ramdisk_name, VENDOR_RAMDISK_NAME_SIZE));
+}
+
+class RepackVendorBootImgTestV4 : public ::testing::TestWithParam<uint32_t /* ramdisk type */> {
+  public:
+    virtual void SetUp() {
+        vboot = std::make_unique<ReadWriteTestFileHandle>("vendor_boot_v4_with_frag.img");
+        ASSERT_RESULT_OK(vboot->Open());
+    }
+    std::unique_ptr<TestFileHandle> vboot;
+};
+
+TEST_P(RepackVendorBootImgTestV4, Replace) {
+    uint32_t replace_ramdisk_type = GetParam();
+    std::string replace_ramdisk_name;
+    std::string expect_new_ramdisk_content;
+    uint32_t expect_none_size = env->none->size();
+    uint32_t expect_platform_size = env->platform->size();
+    switch (replace_ramdisk_type) {
+        case VENDOR_RAMDISK_TYPE_NONE:
+            replace_ramdisk_name = "none_ramdisk";
+            expect_new_ramdisk_content = env->replace_content + env->platform_content;
+            expect_none_size = env->replace->size();
+            break;
+        case VENDOR_RAMDISK_TYPE_PLATFORM:
+            replace_ramdisk_name = "platform_ramdisk";
+            expect_new_ramdisk_content = env->none_content + env->replace_content;
+            expect_platform_size = env->replace->size();
+            break;
+        default:
+            LOG(FATAL) << "Ramdisk type " << replace_ramdisk_type
+                       << " is not supported by this test.";
+    }
+
+    auto old_content = vboot->Read();
+    ASSERT_RESULT_OK(old_content);
+
+    ASSERT_RESULT_OK(replace_vendor_ramdisk(vboot->fd(), vboot->size(), replace_ramdisk_name,
+                                            env->replace->fd(), env->replace->size()));
+    EXPECT_RESULT(vboot->fsize(), vboot->size()) << "File size should not change after repack";
+
+    auto new_content_res = vboot->Read();
+    ASSERT_RESULT_OK(new_content_res);
+    std::string_view new_content(*new_content_res);
+
+    auto hdr = reinterpret_cast<const vendor_boot_img_hdr_v4*>(new_content.data());
+    ASSERT_EQ(0, memcmp(VENDOR_BOOT_MAGIC, hdr->magic, VENDOR_BOOT_MAGIC_SIZE));
+    ASSERT_EQ(4, hdr->header_version);
+    EXPECT_EQ(hdr->vendor_ramdisk_size, expect_none_size + expect_platform_size);
+    EXPECT_EQ(hdr->dtb_size, env->dtb->size());
+    EXPECT_EQ(hdr->bootconfig_size, env->bootconfig->size());
+
+    auto o = round_up(sizeof(vendor_boot_img_hdr_v3), hdr->page_size);
+    auto p = round_up(hdr->vendor_ramdisk_size, hdr->page_size);
+    auto q = round_up(hdr->dtb_size, hdr->page_size);
+    auto r = round_up(hdr->vendor_ramdisk_table_size, hdr->page_size);
+    auto s = round_up(hdr->bootconfig_size, hdr->page_size);
+
+    EXPECT_THAT(new_content.substr(o, p), IsPadded(expect_new_ramdisk_content));
+    EXPECT_THAT(new_content.substr(o + p, q), IsPadded(env->dtb_content));
+
+    // Check changes in table.
+    EXPECT_EQ(hdr->vendor_ramdisk_table_entry_num, 2);
+    EXPECT_EQ(hdr->vendor_ramdisk_table_size, 2 * hdr->vendor_ramdisk_table_entry_size);
+    EXPECT_GE(hdr->vendor_ramdisk_table_entry_size, sizeof(vendor_ramdisk_table_entry_v4));
+    auto entry_none =
+            reinterpret_cast<const vendor_ramdisk_table_entry_v4*>(&new_content[o + p + q]);
+    EXPECT_EQ(entry_none->ramdisk_offset, 0);
+    EXPECT_EQ(entry_none->ramdisk_size, expect_none_size);
+    EXPECT_EQ(entry_none->ramdisk_type, VENDOR_RAMDISK_TYPE_NONE);
+    EXPECT_EQ(GetRamdiskName(entry_none), "none_ramdisk");
+
+    auto entry_platform = reinterpret_cast<const vendor_ramdisk_table_entry_v4*>(
+            &new_content[o + p + q + hdr->vendor_ramdisk_table_entry_size]);
+    EXPECT_EQ(entry_platform->ramdisk_offset, expect_none_size);
+    EXPECT_EQ(entry_platform->ramdisk_size, expect_platform_size);
+    EXPECT_EQ(entry_platform->ramdisk_type, VENDOR_RAMDISK_TYPE_PLATFORM);
+    EXPECT_EQ(GetRamdiskName(entry_platform), "platform_ramdisk");
+
+    EXPECT_THAT(new_content.substr(o + p + q + r, s), IsPadded(env->bootconfig_content));
+
+    EXPECT_THAT(new_content, HasSameAvbFooter(*old_content));
+}
+INSTANTIATE_TEST_SUITE_P(RepackVendorBootImgTest, RepackVendorBootImgTestV4,
+                         ::testing::Values(VENDOR_RAMDISK_TYPE_NONE, VENDOR_RAMDISK_TYPE_PLATFORM),
+                         [](const auto& info) {
+                             return info.param == VENDOR_RAMDISK_TYPE_NONE ? "none" : "platform";
+                         });
+
+}  // namespace
+
+int main(int argc, char* argv[]) {
+    ::testing::InitGoogleTest(&argc, argv);
+    env = static_cast<RepackVendorBootImgTestEnv*>(
+            testing::AddGlobalTestEnvironment(new RepackVendorBootImgTestEnv));
+    return RUN_ALL_TESTS();
+}
diff --git a/fs_mgr/fs_mgr_boot_config.cpp b/fs_mgr/fs_mgr_boot_config.cpp
index 75d1e0d..e3ef232 100644
--- a/fs_mgr/fs_mgr_boot_config.cpp
+++ b/fs_mgr/fs_mgr_boot_config.cpp
@@ -91,6 +91,12 @@
         if (key == bootconfig_key) {
             *out_val = value;
             return true;
+        } else if (android_key == "hardware" && android_key == key) {
+            // bootconfig doesn't allow subkeys and values to coexist, so
+            // "androidboot.hardware" cannot be used. It is replaced in
+            // bootconfig with "hardware"
+            *out_val = value;
+            return true;
         }
     }
 
diff --git a/fs_mgr/libfiemap/include/libfiemap/image_manager.h b/fs_mgr/libfiemap/include/libfiemap/image_manager.h
index 50f4f33..7e30509 100644
--- a/fs_mgr/libfiemap/include/libfiemap/image_manager.h
+++ b/fs_mgr/libfiemap/include/libfiemap/image_manager.h
@@ -198,7 +198,7 @@
 
     ~MappedDevice();
 
-    int fd() const { return fd_; }
+    int fd() const { return fd_.get(); }
     const std::string& path() const { return path_; }
 
   protected:
diff --git a/fs_mgr/liblp/builder.cpp b/fs_mgr/liblp/builder.cpp
index 623293e..6cb2c51 100644
--- a/fs_mgr/liblp/builder.cpp
+++ b/fs_mgr/liblp/builder.cpp
@@ -383,11 +383,6 @@
                << " partition alignment is not sector-aligned.";
         return false;
     }
-    if (device_info.alignment_offset > device_info.alignment) {
-        LERROR << "Block device " << device_info.partition_name
-               << " partition alignment offset is greater than its alignment.";
-        return false;
-    }
     return true;
 }
 
@@ -489,7 +484,7 @@
     // Compute the first free sector, factoring in alignment.
     uint64_t free_area_start = total_reserved;
     bool ok;
-    if (super.alignment || super.alignment_offset) {
+    if (super.alignment) {
         ok = AlignTo(free_area_start, super.alignment, &free_area_start);
     } else {
         ok = AlignTo(free_area_start, logical_block_size, &free_area_start);
diff --git a/fs_mgr/liblp/builder_test.cpp b/fs_mgr/liblp/builder_test.cpp
index e4b617a..72827eb 100644
--- a/fs_mgr/liblp/builder_test.cpp
+++ b/fs_mgr/liblp/builder_test.cpp
@@ -176,10 +176,10 @@
     ASSERT_NE(super_device, nullptr);
     EXPECT_EQ(super_device->first_logical_sector, 1536);
 
-    // Alignment offset without alignment doesn't mean anything.
+    // Alignment offset without alignment is ignored.
     device_info.alignment = 0;
     builder = MetadataBuilder::New(device_info, 1024, 2);
-    ASSERT_EQ(builder, nullptr);
+    ASSERT_NE(builder, nullptr);
 
     // Test a small alignment with an alignment offset.
     device_info.alignment = 12 * 1024;
@@ -444,11 +444,6 @@
     device_info.alignment = 131072;
     builder = MetadataBuilder::New(device_info, kMetadataSize, 1);
     EXPECT_EQ(builder, nullptr);
-
-    device_info.alignment = 0;
-    device_info.alignment_offset = 32768 - LP_SECTOR_SIZE;
-    builder = MetadataBuilder::New(device_info, kMetadataSize, 1);
-    EXPECT_EQ(builder, nullptr);
 }
 
 TEST_F(BuilderTest, UpdateBlockDeviceInfo) {
diff --git a/fs_mgr/libsnapshot/Android.bp b/fs_mgr/libsnapshot/Android.bp
index b808609..ea92d25 100644
--- a/fs_mgr/libsnapshot/Android.bp
+++ b/fs_mgr/libsnapshot/Android.bp
@@ -416,6 +416,7 @@
         "snapuserd_server.cpp",
         "snapuserd.cpp",
         "snapuserd_daemon.cpp",
+	"snapuserd_worker.cpp",
     ],
 
     cflags: [
@@ -554,6 +555,7 @@
     srcs: [
         "cow_snapuserd_test.cpp",
         "snapuserd.cpp",
+	"snapuserd_worker.cpp",
     ],
     cflags: [
         "-Wall",
diff --git a/fs_mgr/libsnapshot/android/snapshot/snapshot.proto b/fs_mgr/libsnapshot/android/snapshot/snapshot.proto
index b4e92a2..1ebc29f 100644
--- a/fs_mgr/libsnapshot/android/snapshot/snapshot.proto
+++ b/fs_mgr/libsnapshot/android/snapshot/snapshot.proto
@@ -46,7 +46,7 @@
     SECOND_PHASE = 2;
 }
 
-// Next: 12
+// Next: 13
 message SnapshotStatus {
     // Name of the snapshot. This is usually the name of the snapshotted
     // logical partition; for example, "system_b".
@@ -105,6 +105,9 @@
 
     // Compression algorithm (none, gz, or brotli).
     string compression_algorithm = 11;
+
+    // Estimated COW size from OTA manifest.
+    uint64 estimated_cow_size = 12;
 }
 
 // Next: 8
@@ -159,7 +162,7 @@
     MergePhase merge_phase = 6;
 }
 
-// Next: 5
+// Next: 9
 message SnapshotMergeReport {
     // Status of the update after the merge attempts.
     UpdateState state = 1;
@@ -173,4 +176,16 @@
 
     // Whether compression/dm-user was used for any snapshots.
     bool compression_enabled = 4;
+
+    // Total size used by COWs, including /data and the super partition.
+    uint64 total_cow_size_bytes = 5;
+
+    // Sum of the estimated COW fields in the OTA manifest.
+    uint64 estimated_cow_size_bytes = 6;
+
+    // Time from boot to sys.boot_completed, in milliseconds.
+    uint32 boot_complete_time_ms = 7;
+
+    // Time from sys.boot_completed to merge start, in milliseconds.
+    uint32 boot_complete_to_merge_start_time_ms = 8;
 }
diff --git a/fs_mgr/libsnapshot/cow_api_test.cpp b/fs_mgr/libsnapshot/cow_api_test.cpp
index a96352a..5d63220 100644
--- a/fs_mgr/libsnapshot/cow_api_test.cpp
+++ b/fs_mgr/libsnapshot/cow_api_test.cpp
@@ -757,6 +757,30 @@
     ASSERT_TRUE(iter->Done());
 }
 
+TEST_F(CowTest, AppendAfterFinalize) {
+    CowOptions options;
+    options.cluster_ops = 0;
+    auto writer = std::make_unique<CowWriter>(options);
+    ASSERT_TRUE(writer->Initialize(cow_->fd));
+
+    std::string data = "This is some data, believe it";
+    data.resize(options.block_size, '\0');
+    ASSERT_TRUE(writer->AddRawBlocks(50, data.data(), data.size()));
+    ASSERT_TRUE(writer->AddLabel(3));
+    ASSERT_TRUE(writer->Finalize());
+
+    std::string data2 = "More data!";
+    data2.resize(options.block_size, '\0');
+    ASSERT_TRUE(writer->AddRawBlocks(51, data2.data(), data2.size()));
+    ASSERT_TRUE(writer->Finalize());
+
+    ASSERT_EQ(lseek(cow_->fd, 0, SEEK_SET), 0);
+
+    // COW should be valid.
+    CowReader reader;
+    ASSERT_TRUE(reader.Parse(cow_->fd));
+}
+
 }  // namespace snapshot
 }  // namespace android
 
diff --git a/fs_mgr/libsnapshot/cow_reader.cpp b/fs_mgr/libsnapshot/cow_reader.cpp
index cf9f6ea..7199b38 100644
--- a/fs_mgr/libsnapshot/cow_reader.cpp
+++ b/fs_mgr/libsnapshot/cow_reader.cpp
@@ -42,6 +42,29 @@
 #endif
 }
 
+bool CowReader::InitForMerge(android::base::unique_fd&& fd) {
+    owned_fd_ = std::move(fd);
+    fd_ = owned_fd_.get();
+
+    auto pos = lseek(fd_.get(), 0, SEEK_END);
+    if (pos < 0) {
+        PLOG(ERROR) << "lseek end failed";
+        return false;
+    }
+    fd_size_ = pos;
+
+    if (lseek(fd_.get(), 0, SEEK_SET) < 0) {
+        PLOG(ERROR) << "lseek header failed";
+        return false;
+    }
+    if (!android::base::ReadFully(fd_, &header_, sizeof(header_))) {
+        PLOG(ERROR) << "read header failed";
+        return false;
+    }
+
+    return true;
+}
+
 bool CowReader::Parse(android::base::unique_fd&& fd, std::optional<uint64_t> label) {
     owned_fd_ = std::move(fd);
     return Parse(android::base::borrowed_fd{owned_fd_}, label);
@@ -206,7 +229,8 @@
 
     if (footer_) {
         if (ops_buffer->size() != footer_->op.num_ops) {
-            LOG(ERROR) << "num ops does not match";
+            LOG(ERROR) << "num ops does not match, expected " << footer_->op.num_ops << ", found "
+                       << ops_buffer->size();
             return false;
         }
         if (ops_buffer->size() * sizeof(CowOperation) != footer_->op.ops_size) {
@@ -226,6 +250,8 @@
     }
 
     ops_ = ops_buffer;
+    ops_->shrink_to_fit();
+
     return true;
 }
 
diff --git a/fs_mgr/libsnapshot/cow_writer.cpp b/fs_mgr/libsnapshot/cow_writer.cpp
index 81edc79..59f6d6f 100644
--- a/fs_mgr/libsnapshot/cow_writer.cpp
+++ b/fs_mgr/libsnapshot/cow_writer.cpp
@@ -376,6 +376,7 @@
     auto continue_data_pos = next_data_pos_;
     auto continue_op_pos = next_op_pos_;
     auto continue_size = ops_.size();
+    auto continue_num_ops = footer_.op.num_ops;
     bool extra_cluster = false;
 
     // Footer should be at the end of a file, so if there is data after the current block, end it
@@ -408,9 +409,9 @@
         current_data_size_ = continue_data_size;
         next_data_pos_ = continue_data_pos;
         next_op_pos_ = continue_op_pos;
+        footer_.op.num_ops = continue_num_ops;
         ops_.resize(continue_size);
     }
-
     return Sync();
 }
 
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/cow_reader.h b/fs_mgr/libsnapshot/include/libsnapshot/cow_reader.h
index 1de7473..552fd96 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/cow_reader.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/cow_reader.h
@@ -116,12 +116,15 @@
 class CowReader : public ICowReader {
   public:
     CowReader();
+    ~CowReader() { owned_fd_ = {}; }
 
     // Parse the COW, optionally, up to the given label. If no label is
     // specified, the COW must have an intact footer.
     bool Parse(android::base::unique_fd&& fd, std::optional<uint64_t> label = {});
     bool Parse(android::base::borrowed_fd fd, std::optional<uint64_t> label = {});
 
+    bool InitForMerge(android::base::unique_fd&& fd);
+
     bool GetHeader(CowHeader* header) override;
     bool GetFooter(CowFooter* footer) override;
 
@@ -146,6 +149,8 @@
 
     uint64_t total_data_ops() { return total_data_ops_; }
 
+    void CloseCowFd() { owned_fd_ = {}; }
+
   private:
     bool ParseOps(std::optional<uint64_t> label);
 
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/mock_snapshot.h b/fs_mgr/libsnapshot/include/libsnapshot/mock_snapshot.h
index 1e420cb..1cb966b 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/mock_snapshot.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/mock_snapshot.h
@@ -26,7 +26,8 @@
     MOCK_METHOD(bool, BeginUpdate, (), (override));
     MOCK_METHOD(bool, CancelUpdate, (), (override));
     MOCK_METHOD(bool, FinishedSnapshotWrites, (bool wipe), (override));
-    MOCK_METHOD(bool, InitiateMerge, (uint64_t * cow_file_size), (override));
+    MOCK_METHOD(void, UpdateCowStats, (ISnapshotMergeStats * stats), (override));
+    MOCK_METHOD(bool, InitiateMerge, (), (override));
 
     MOCK_METHOD(UpdateState, ProcessUpdateState,
                 (const std::function<bool()>& callback, const std::function<bool()>& before_cancel),
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h b/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
index a79a86d..7e74fac 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
@@ -127,9 +127,14 @@
     // may need to be merged before wiping.
     virtual bool FinishedSnapshotWrites(bool wipe) = 0;
 
+    // Update an ISnapshotMergeStats object with statistics about COW usage.
+    // This should be called before the merge begins as otherwise snapshots
+    // may be deleted.
+    virtual void UpdateCowStats(ISnapshotMergeStats* stats) = 0;
+
     // Initiate a merge on all snapshot devices. This should only be used after an
     // update has been marked successful after booting.
-    virtual bool InitiateMerge(uint64_t* cow_file_size = nullptr) = 0;
+    virtual bool InitiateMerge() = 0;
 
     // Perform any necessary post-boot actions. This should be run soon after
     // /data is mounted.
@@ -326,7 +331,8 @@
     bool BeginUpdate() override;
     bool CancelUpdate() override;
     bool FinishedSnapshotWrites(bool wipe) override;
-    bool InitiateMerge(uint64_t* cow_file_size = nullptr) override;
+    void UpdateCowStats(ISnapshotMergeStats* stats) override;
+    bool InitiateMerge() override;
     UpdateState ProcessUpdateState(const std::function<bool()>& callback = {},
                                    const std::function<bool()>& before_cancel = {}) override;
     UpdateState GetUpdateState(double* progress = nullptr) override;
@@ -491,7 +497,8 @@
     bool RemoveAllSnapshots(LockedFile* lock);
 
     // List the known snapshot names.
-    bool ListSnapshots(LockedFile* lock, std::vector<std::string>* snapshots);
+    bool ListSnapshots(LockedFile* lock, std::vector<std::string>* snapshots,
+                       const std::string& suffix = "");
 
     // Check for a cancelled or rolled back merge, returning true if such a
     // condition was detected and handled.
@@ -679,6 +686,9 @@
     friend std::ostream& operator<<(std::ostream& os, SnapshotManager::Slot slot);
     Slot GetCurrentSlot();
 
+    // Return the suffix we expect snapshots to have.
+    std::string GetSnapshotSlotSuffix();
+
     std::string ReadUpdateSourceSlotSuffix();
 
     // Helper for RemoveAllSnapshots.
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapshot_stats.h b/fs_mgr/libsnapshot/include/libsnapshot/snapshot_stats.h
index 96d2deb..e617d7a 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapshot_stats.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapshot_stats.h
@@ -30,7 +30,15 @@
     virtual bool Start() = 0;
     virtual void set_state(android::snapshot::UpdateState state, bool using_compression) = 0;
     virtual void set_cow_file_size(uint64_t cow_file_size) = 0;
+    virtual void set_total_cow_size_bytes(uint64_t bytes) = 0;
+    virtual void set_estimated_cow_size_bytes(uint64_t bytes) = 0;
+    virtual void set_boot_complete_time_ms(uint32_t ms) = 0;
+    virtual void set_boot_complete_to_merge_start_time_ms(uint32_t ms) = 0;
     virtual uint64_t cow_file_size() = 0;
+    virtual uint64_t total_cow_size_bytes() = 0;
+    virtual uint64_t estimated_cow_size_bytes() = 0;
+    virtual uint32_t boot_complete_time_ms() = 0;
+    virtual uint32_t boot_complete_to_merge_start_time_ms() = 0;
 
     // Called when merge ends. Properly clean up permanent storage.
     class Result {
@@ -54,6 +62,14 @@
     void set_state(android::snapshot::UpdateState state, bool using_compression) override;
     void set_cow_file_size(uint64_t cow_file_size) override;
     uint64_t cow_file_size() override;
+    void set_total_cow_size_bytes(uint64_t bytes) override;
+    void set_estimated_cow_size_bytes(uint64_t bytes) override;
+    uint64_t total_cow_size_bytes() override;
+    uint64_t estimated_cow_size_bytes() override;
+    void set_boot_complete_time_ms(uint32_t ms) override;
+    uint32_t boot_complete_time_ms() override;
+    void set_boot_complete_to_merge_start_time_ms(uint32_t ms) override;
+    uint32_t boot_complete_to_merge_start_time_ms() override;
     std::unique_ptr<Result> Finish() override;
 
   private:
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapshot_stub.h b/fs_mgr/libsnapshot/include/libsnapshot/snapshot_stub.h
index 3365ceb..cc75db8 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapshot_stub.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapshot_stub.h
@@ -28,7 +28,8 @@
     bool BeginUpdate() override;
     bool CancelUpdate() override;
     bool FinishedSnapshotWrites(bool wipe) override;
-    bool InitiateMerge(uint64_t* cow_file_size = nullptr) override;
+    void UpdateCowStats(ISnapshotMergeStats* stats) override;
+    bool InitiateMerge() override;
     UpdateState ProcessUpdateState(const std::function<bool()>& callback = {},
                                    const std::function<bool()>& before_cancel = {}) override;
     UpdateState GetUpdateState(double* progress = nullptr) override;
diff --git a/fs_mgr/libsnapshot/partition_cow_creator.cpp b/fs_mgr/libsnapshot/partition_cow_creator.cpp
index 6002043..5569da0 100644
--- a/fs_mgr/libsnapshot/partition_cow_creator.cpp
+++ b/fs_mgr/libsnapshot/partition_cow_creator.cpp
@@ -202,6 +202,10 @@
     ret.snapshot_status.set_device_size(target_partition->size());
     ret.snapshot_status.set_snapshot_size(target_partition->size());
 
+    if (update && update->has_estimate_cow_size()) {
+        ret.snapshot_status.set_estimated_cow_size(update->estimate_cow_size());
+    }
+
     if (ret.snapshot_status.snapshot_size() == 0) {
         LOG(INFO) << "Not creating snapshot for partition " << ret.snapshot_status.name();
         ret.snapshot_status.set_cow_partition_size(0);
diff --git a/fs_mgr/libsnapshot/snapshot.cpp b/fs_mgr/libsnapshot/snapshot.cpp
index ca4c265..bd1e284 100644
--- a/fs_mgr/libsnapshot/snapshot.cpp
+++ b/fs_mgr/libsnapshot/snapshot.cpp
@@ -230,6 +230,15 @@
     return Slot::Target;
 }
 
+std::string SnapshotManager::GetSnapshotSlotSuffix() {
+    switch (GetCurrentSlot()) {
+        case Slot::Target:
+            return device_->GetSlotSuffix();
+        default:
+            return device_->GetOtherSlotSuffix();
+    }
+}
+
 static bool RemoveFileIfExists(const std::string& path) {
     std::string message;
     if (!android::base::RemoveFileIfExists(path, &message)) {
@@ -624,7 +633,7 @@
     return true;
 }
 
-bool SnapshotManager::InitiateMerge(uint64_t* cow_file_size) {
+bool SnapshotManager::InitiateMerge() {
     auto lock = LockExclusive();
     if (!lock) return false;
 
@@ -691,7 +700,6 @@
 
     std::vector<std::string> first_merge_group;
 
-    uint64_t total_cow_file_size = 0;
     DmTargetSnapshot::Status initial_target_values = {};
     for (const auto& snapshot : snapshots) {
         DmTargetSnapshot::Status current_status;
@@ -706,7 +714,6 @@
         if (!ReadSnapshotStatus(lock.get(), snapshot, &snapshot_status)) {
             return false;
         }
-        total_cow_file_size += snapshot_status.cow_file_size();
 
         compression_enabled |= snapshot_status.compression_enabled();
         if (DecideMergePhase(snapshot_status) == MergePhase::FIRST_PHASE) {
@@ -714,10 +721,6 @@
         }
     }
 
-    if (cow_file_size) {
-        *cow_file_size = total_cow_file_size;
-    }
-
     SnapshotUpdateStatus initial_status;
     initial_status.set_state(UpdateState::Merging);
     initial_status.set_sectors_allocated(initial_target_values.sectors_allocated);
@@ -1262,7 +1265,7 @@
             LOG(ERROR) << "DeleteDevice timeout: " << name;
             return false;
         }
-        std::this_thread::sleep_for(250ms);
+        std::this_thread::sleep_for(400ms);
     }
 
     return true;
@@ -1732,7 +1735,8 @@
     return update_status.compression_enabled();
 }
 
-bool SnapshotManager::ListSnapshots(LockedFile* lock, std::vector<std::string>* snapshots) {
+bool SnapshotManager::ListSnapshots(LockedFile* lock, std::vector<std::string>* snapshots,
+                                    const std::string& suffix) {
     CHECK(lock);
 
     auto dir_path = metadata_dir_ + "/snapshots"s;
@@ -1745,7 +1749,12 @@
     struct dirent* dp;
     while ((dp = readdir(dir.get())) != nullptr) {
         if (dp->d_type != DT_REG) continue;
-        snapshots->emplace_back(dp->d_name);
+
+        std::string name(dp->d_name);
+        if (!suffix.empty() && !android::base::EndsWith(name, suffix)) {
+            continue;
+        }
+        snapshots->emplace_back(std::move(name));
     }
     return true;
 }
@@ -3565,5 +3574,34 @@
     return MergePhase::SECOND_PHASE;
 }
 
+void SnapshotManager::UpdateCowStats(ISnapshotMergeStats* stats) {
+    auto lock = LockExclusive();
+    if (!lock) return;
+
+    std::vector<std::string> snapshots;
+    if (!ListSnapshots(lock.get(), &snapshots, GetSnapshotSlotSuffix())) {
+        LOG(ERROR) << "Could not list snapshots";
+        return;
+    }
+
+    uint64_t cow_file_size = 0;
+    uint64_t total_cow_size = 0;
+    uint64_t estimated_cow_size = 0;
+    for (const auto& snapshot : snapshots) {
+        SnapshotStatus status;
+        if (!ReadSnapshotStatus(lock.get(), snapshot, &status)) {
+            return;
+        }
+
+        cow_file_size += status.cow_file_size();
+        total_cow_size += status.cow_file_size() + status.cow_partition_size();
+        estimated_cow_size += status.estimated_cow_size();
+    }
+
+    stats->set_cow_file_size(cow_file_size);
+    stats->set_total_cow_size_bytes(total_cow_size);
+    stats->set_estimated_cow_size_bytes(estimated_cow_size);
+}
+
 }  // namespace snapshot
 }  // namespace android
diff --git a/fs_mgr/libsnapshot/snapshot_stats.cpp b/fs_mgr/libsnapshot/snapshot_stats.cpp
index 513700d..7fcfcea 100644
--- a/fs_mgr/libsnapshot/snapshot_stats.cpp
+++ b/fs_mgr/libsnapshot/snapshot_stats.cpp
@@ -98,6 +98,38 @@
     return report_.cow_file_size();
 }
 
+void SnapshotMergeStats::set_total_cow_size_bytes(uint64_t bytes) {
+    report_.set_total_cow_size_bytes(bytes);
+}
+
+void SnapshotMergeStats::set_estimated_cow_size_bytes(uint64_t bytes) {
+    report_.set_estimated_cow_size_bytes(bytes);
+}
+
+uint64_t SnapshotMergeStats::total_cow_size_bytes() {
+    return report_.total_cow_size_bytes();
+}
+
+uint64_t SnapshotMergeStats::estimated_cow_size_bytes() {
+    return report_.estimated_cow_size_bytes();
+}
+
+void SnapshotMergeStats::set_boot_complete_time_ms(uint32_t ms) {
+    report_.set_boot_complete_time_ms(ms);
+}
+
+uint32_t SnapshotMergeStats::boot_complete_time_ms() {
+    return report_.boot_complete_time_ms();
+}
+
+void SnapshotMergeStats::set_boot_complete_to_merge_start_time_ms(uint32_t ms) {
+    report_.set_boot_complete_to_merge_start_time_ms(ms);
+}
+
+uint32_t SnapshotMergeStats::boot_complete_to_merge_start_time_ms() {
+    return report_.boot_complete_to_merge_start_time_ms();
+}
+
 class SnapshotMergeStatsResultImpl : public SnapshotMergeStats::Result {
   public:
     SnapshotMergeStatsResultImpl(const SnapshotMergeReport& report,
diff --git a/fs_mgr/libsnapshot/snapshot_stub.cpp b/fs_mgr/libsnapshot/snapshot_stub.cpp
index 8a254c9..43825cc 100644
--- a/fs_mgr/libsnapshot/snapshot_stub.cpp
+++ b/fs_mgr/libsnapshot/snapshot_stub.cpp
@@ -43,7 +43,7 @@
     return false;
 }
 
-bool SnapshotManagerStub::InitiateMerge(uint64_t*) {
+bool SnapshotManagerStub::InitiateMerge() {
     LOG(ERROR) << __FUNCTION__ << " should never be called.";
     return false;
 }
@@ -127,6 +127,14 @@
     void set_cow_file_size(uint64_t) override {}
     uint64_t cow_file_size() override { return 0; }
     std::unique_ptr<Result> Finish() override { return nullptr; }
+    void set_total_cow_size_bytes(uint64_t) override {}
+    void set_estimated_cow_size_bytes(uint64_t) override {}
+    uint64_t total_cow_size_bytes() override { return 0; }
+    uint64_t estimated_cow_size_bytes() override { return 0; }
+    void set_boot_complete_time_ms(uint32_t) override {}
+    uint32_t boot_complete_time_ms() override { return 0; }
+    void set_boot_complete_to_merge_start_time_ms(uint32_t) override {}
+    uint32_t boot_complete_to_merge_start_time_ms() override { return 0; }
 };
 
 ISnapshotMergeStats* SnapshotManagerStub::GetSnapshotMergeStatsInstance() {
@@ -151,4 +159,8 @@
     return false;
 }
 
+void SnapshotManagerStub::UpdateCowStats(ISnapshotMergeStats*) {
+    LOG(ERROR) << __FUNCTION__ << " should never be called.";
+}
+
 }  // namespace android::snapshot
diff --git a/fs_mgr/libsnapshot/snapshotctl.cpp b/fs_mgr/libsnapshot/snapshotctl.cpp
index a44de84..5eb2003 100644
--- a/fs_mgr/libsnapshot/snapshotctl.cpp
+++ b/fs_mgr/libsnapshot/snapshotctl.cpp
@@ -48,6 +48,17 @@
     return SnapshotManager::New()->Dump(std::cout);
 }
 
+bool MapCmdHandler(int, char** argv) {
+    android::base::InitLogging(argv, &android::base::StderrLogger);
+    using namespace std::chrono_literals;
+    return SnapshotManager::New()->MapAllSnapshots(5000ms);
+}
+
+bool UnmapCmdHandler(int, char** argv) {
+    android::base::InitLogging(argv, &android::base::StderrLogger);
+    return SnapshotManager::New()->UnmapAllSnapshots();
+}
+
 bool MergeCmdHandler(int /*argc*/, char** argv) {
     android::base::InitLogging(argv, &android::base::StderrLogger);
     LOG(WARNING) << "Deprecated. Call update_engine_client --merge instead.";
@@ -58,6 +69,8 @@
         // clang-format off
         {"dump", DumpCmdHandler},
         {"merge", MergeCmdHandler},
+        {"map", MapCmdHandler},
+        {"unmap", UnmapCmdHandler},
         // clang-format on
 };
 
diff --git a/fs_mgr/libsnapshot/snapuserd.cpp b/fs_mgr/libsnapshot/snapuserd.cpp
index 4c4a342..5ef9e29 100644
--- a/fs_mgr/libsnapshot/snapuserd.cpp
+++ b/fs_mgr/libsnapshot/snapuserd.cpp
@@ -32,41 +32,6 @@
 #define SNAP_LOG(level) LOG(level) << misc_name_ << ": "
 #define SNAP_PLOG(level) PLOG(level) << misc_name_ << ": "
 
-static constexpr size_t PAYLOAD_SIZE = (1UL << 20);
-
-static_assert(PAYLOAD_SIZE >= BLOCK_SZ);
-
-void BufferSink::Initialize(size_t size) {
-    buffer_size_ = size;
-    buffer_offset_ = 0;
-    buffer_ = std::make_unique<uint8_t[]>(size);
-}
-
-void* BufferSink::GetPayloadBuffer(size_t size) {
-    if ((buffer_size_ - buffer_offset_) < size) return nullptr;
-
-    char* buffer = reinterpret_cast<char*>(GetBufPtr());
-    struct dm_user_message* msg = (struct dm_user_message*)(&(buffer[0]));
-    return (char*)msg->payload.buf + buffer_offset_;
-}
-
-void* BufferSink::GetBuffer(size_t requested, size_t* actual) {
-    void* buf = GetPayloadBuffer(requested);
-    if (!buf) {
-        *actual = 0;
-        return nullptr;
-    }
-    *actual = requested;
-    return buf;
-}
-
-struct dm_user_header* BufferSink::GetHeaderPtr() {
-    CHECK(sizeof(struct dm_user_header) <= buffer_size_);
-    char* buf = reinterpret_cast<char*>(GetBufPtr());
-    struct dm_user_header* header = (struct dm_user_header*)(&(buf[0]));
-    return header;
-}
-
 Snapuserd::Snapuserd(const std::string& misc_name, const std::string& cow_device,
                      const std::string& backing_device) {
     misc_name_ = misc_name;
@@ -75,356 +40,32 @@
     control_device_ = "/dev/dm-user/" + misc_name;
 }
 
-// Construct kernel COW header in memory
-// This header will be in sector 0. The IO
-// request will always be 4k. After constructing
-// the header, zero out the remaining block.
-void Snapuserd::ConstructKernelCowHeader() {
-    void* buffer = bufsink_.GetPayloadBuffer(BLOCK_SZ);
-    CHECK(buffer != nullptr);
+bool Snapuserd::InitializeWorkers() {
+    for (int i = 0; i < NUM_THREADS_PER_PARTITION; i++) {
+        std::unique_ptr<WorkerThread> wt = std::make_unique<WorkerThread>(
+                cow_device_, backing_store_device_, control_device_, misc_name_, GetSharedPtr());
 
-    memset(buffer, 0, BLOCK_SZ);
-
-    struct disk_header* dh = reinterpret_cast<struct disk_header*>(buffer);
-
-    dh->magic = SNAP_MAGIC;
-    dh->valid = SNAPSHOT_VALID;
-    dh->version = SNAPSHOT_DISK_VERSION;
-    dh->chunk_size = CHUNK_SIZE;
-}
-
-// Start the replace operation. This will read the
-// internal COW format and if the block is compressed,
-// it will be de-compressed.
-bool Snapuserd::ProcessReplaceOp(const CowOperation* cow_op) {
-    if (!reader_->ReadData(*cow_op, &bufsink_)) {
-        SNAP_LOG(ERROR) << "ProcessReplaceOp failed for block " << cow_op->new_block;
-        return false;
+        worker_threads_.push_back(std::move(wt));
     }
-
     return true;
 }
 
-// Start the copy operation. This will read the backing
-// block device which is represented by cow_op->source.
-bool Snapuserd::ProcessCopyOp(const CowOperation* cow_op) {
-    void* buffer = bufsink_.GetPayloadBuffer(BLOCK_SZ);
-    CHECK(buffer != nullptr);
+bool Snapuserd::CommitMerge(int num_merge_ops) {
+    {
+        std::lock_guard<std::mutex> lock(lock_);
+        CowHeader header;
 
-    // Issue a single 4K IO. However, this can be optimized
-    // if the successive blocks are contiguous.
-    if (!android::base::ReadFullyAtOffset(backing_store_fd_, buffer, BLOCK_SZ,
-                                          cow_op->source * BLOCK_SZ)) {
-        SNAP_PLOG(ERROR) << "Copy-op failed. Read from backing store: " << backing_store_device_
-                         << "at block :" << cow_op->source;
-        return false;
-    }
-
-    return true;
-}
-
-bool Snapuserd::ProcessZeroOp() {
-    // Zero out the entire block
-    void* buffer = bufsink_.GetPayloadBuffer(BLOCK_SZ);
-    CHECK(buffer != nullptr);
-
-    memset(buffer, 0, BLOCK_SZ);
-    return true;
-}
-
-bool Snapuserd::ProcessCowOp(const CowOperation* cow_op) {
-    CHECK(cow_op != nullptr);
-
-    switch (cow_op->type) {
-        case kCowReplaceOp: {
-            return ProcessReplaceOp(cow_op);
+        reader_->GetHeader(&header);
+        header.num_merge_ops += num_merge_ops;
+        reader_->UpdateMergeProgress(num_merge_ops);
+        if (!writer_->CommitMerge(num_merge_ops)) {
+            SNAP_LOG(ERROR) << "CommitMerge failed... merged_ops_cur_iter: " << num_merge_ops
+                            << " Total-merged-ops: " << header.num_merge_ops;
+            return false;
         }
-
-        case kCowZeroOp: {
-            return ProcessZeroOp();
-        }
-
-        case kCowCopyOp: {
-            return ProcessCopyOp(cow_op);
-        }
-
-        default: {
-            SNAP_LOG(ERROR) << "Unknown operation-type found: " << cow_op->type;
-        }
-    }
-    return false;
-}
-
-int Snapuserd::ReadUnalignedSector(sector_t sector, size_t size,
-                                   std::map<sector_t, const CowOperation*>::iterator& it) {
-    size_t skip_sector_size = 0;
-
-    SNAP_LOG(DEBUG) << "ReadUnalignedSector: sector " << sector << " size: " << size
-                    << " Aligned sector: " << it->second;
-
-    if (!ProcessCowOp(it->second)) {
-        SNAP_LOG(ERROR) << "ReadUnalignedSector: " << sector << " failed of size: " << size;
-        return -1;
+        merge_initiated_ = true;
     }
 
-    int num_sectors_skip = sector - it->first;
-
-    if (num_sectors_skip > 0) {
-        skip_sector_size = num_sectors_skip << SECTOR_SHIFT;
-        char* buffer = reinterpret_cast<char*>(bufsink_.GetBufPtr());
-        struct dm_user_message* msg = (struct dm_user_message*)(&(buffer[0]));
-
-        memmove(msg->payload.buf, (char*)msg->payload.buf + skip_sector_size,
-                (BLOCK_SZ - skip_sector_size));
-    }
-
-    bufsink_.ResetBufferOffset();
-    return std::min(size, (BLOCK_SZ - skip_sector_size));
-}
-
-/*
- * Read the data for a given COW Operation.
- *
- * Kernel can issue IO at a sector granularity.
- * Hence, an IO may end up with reading partial
- * data from a COW operation or we may also
- * end up with interspersed request between
- * two COW operations.
- *
- */
-int Snapuserd::ReadData(sector_t sector, size_t size) {
-    /*
-     * chunk_map stores COW operation at 4k granularity.
-     * If the requested IO with the sector falls on the 4k
-     * boundary, then we can read the COW op directly without
-     * any issue.
-     *
-     * However, if the requested sector is not 4K aligned,
-     * then we will have the find the nearest COW operation
-     * and chop the 4K block to fetch the requested sector.
-     */
-    std::map<sector_t, const CowOperation*>::iterator it = chunk_map_.find(sector);
-    if (it == chunk_map_.end()) {
-        it = chunk_map_.lower_bound(sector);
-        if (it != chunk_map_.begin()) {
-            --it;
-        }
-
-        /*
-         * If the IO is spanned between two COW operations,
-         * split the IO into two parts:
-         *
-         * 1: Read the first part from the single COW op
-         * 2: Read the second part from the next COW op.
-         *
-         * Ex: Let's say we have a 1024 Bytes IO request.
-         *
-         * 0       COW OP-1  4096     COW OP-2  8192
-         * |******************|*******************|
-         *              |*****|*****|
-         *           3584           4608
-         *              <- 1024B - >
-         *
-         * We have two COW operations which are 4k blocks.
-         * The IO is requested for 1024 Bytes which are spanned
-         * between two COW operations. We will split this IO
-         * into two parts:
-         *
-         * 1: IO of size 512B from offset 3584 bytes (COW OP-1)
-         * 2: IO of size 512B from offset 4096 bytes (COW OP-2)
-         */
-        return ReadUnalignedSector(sector, size, it);
-    }
-
-    int num_ops = DIV_ROUND_UP(size, BLOCK_SZ);
-    while (num_ops) {
-        if (!ProcessCowOp(it->second)) {
-            return -1;
-        }
-        num_ops -= 1;
-        it++;
-        // Update the buffer offset
-        bufsink_.UpdateBufferOffset(BLOCK_SZ);
-
-        SNAP_LOG(DEBUG) << "ReadData at sector: " << sector << " size: " << size;
-    }
-
-    // Reset the buffer offset
-    bufsink_.ResetBufferOffset();
-    return size;
-}
-
-/*
- * dm-snap does prefetch reads while reading disk-exceptions.
- * By default, prefetch value is set to 12; this means that
- * dm-snap will issue 12 areas wherein each area is a 4k page
- * of disk-exceptions.
- *
- * If during prefetch, if the chunk-id seen is beyond the
- * actual number of metadata page, fill the buffer with zero.
- * When dm-snap starts parsing the buffer, it will stop
- * reading metadata page once the buffer content is zero.
- */
-bool Snapuserd::ZerofillDiskExceptions(size_t read_size) {
-    size_t size = exceptions_per_area_ * sizeof(struct disk_exception);
-
-    if (read_size > size) {
-        return false;
-    }
-
-    void* buffer = bufsink_.GetPayloadBuffer(size);
-    CHECK(buffer != nullptr);
-
-    memset(buffer, 0, size);
-    return true;
-}
-
-/*
- * A disk exception is a simple mapping of old_chunk to new_chunk.
- * When dm-snapshot device is created, kernel requests these mapping.
- *
- * Each disk exception is of size 16 bytes. Thus a single 4k page can
- * have:
- *
- * exceptions_per_area_ = 4096/16 = 256. This entire 4k page
- * is considered a metadata page and it is represented by chunk ID.
- *
- * Convert the chunk ID to index into the vector which gives us
- * the metadata page.
- */
-bool Snapuserd::ReadDiskExceptions(chunk_t chunk, size_t read_size) {
-    uint32_t stride = exceptions_per_area_ + 1;
-    size_t size;
-
-    // ChunkID to vector index
-    lldiv_t divresult = lldiv(chunk, stride);
-
-    if (divresult.quot < vec_.size()) {
-        size = exceptions_per_area_ * sizeof(struct disk_exception);
-
-        CHECK(read_size == size);
-
-        void* buffer = bufsink_.GetPayloadBuffer(size);
-        CHECK(buffer != nullptr);
-
-        memcpy(buffer, vec_[divresult.quot].get(), size);
-    } else {
-        return ZerofillDiskExceptions(read_size);
-    }
-
-    return true;
-}
-
-loff_t Snapuserd::GetMergeStartOffset(void* merged_buffer, void* unmerged_buffer,
-                                      int* unmerged_exceptions) {
-    loff_t offset = 0;
-    *unmerged_exceptions = 0;
-
-    while (*unmerged_exceptions <= exceptions_per_area_) {
-        struct disk_exception* merged_de =
-                reinterpret_cast<struct disk_exception*>((char*)merged_buffer + offset);
-        struct disk_exception* cow_de =
-                reinterpret_cast<struct disk_exception*>((char*)unmerged_buffer + offset);
-
-        // Unmerged op by the kernel
-        if (merged_de->old_chunk != 0 || merged_de->new_chunk != 0) {
-            CHECK(merged_de->old_chunk == cow_de->old_chunk);
-            CHECK(merged_de->new_chunk == cow_de->new_chunk);
-
-            offset += sizeof(struct disk_exception);
-            *unmerged_exceptions += 1;
-            continue;
-        }
-
-        break;
-    }
-
-    CHECK(!(*unmerged_exceptions == exceptions_per_area_));
-
-    SNAP_LOG(DEBUG) << "Unmerged_Exceptions: " << *unmerged_exceptions << " Offset: " << offset;
-    return offset;
-}
-
-int Snapuserd::GetNumberOfMergedOps(void* merged_buffer, void* unmerged_buffer, loff_t offset,
-                                    int unmerged_exceptions) {
-    int merged_ops_cur_iter = 0;
-
-    // Find the operations which are merged in this cycle.
-    while ((unmerged_exceptions + merged_ops_cur_iter) < exceptions_per_area_) {
-        struct disk_exception* merged_de =
-                reinterpret_cast<struct disk_exception*>((char*)merged_buffer + offset);
-        struct disk_exception* cow_de =
-                reinterpret_cast<struct disk_exception*>((char*)unmerged_buffer + offset);
-
-        CHECK(merged_de->new_chunk == 0);
-        CHECK(merged_de->old_chunk == 0);
-
-        if (cow_de->new_chunk != 0) {
-            merged_ops_cur_iter += 1;
-            offset += sizeof(struct disk_exception);
-            const CowOperation* cow_op = chunk_map_[ChunkToSector(cow_de->new_chunk)];
-            CHECK(cow_op != nullptr);
-
-            CHECK(cow_op->new_block == cow_de->old_chunk);
-            // zero out to indicate that operation is merged.
-            cow_de->old_chunk = 0;
-            cow_de->new_chunk = 0;
-        } else if (cow_de->old_chunk == 0) {
-            // Already merged op in previous iteration or
-            // This could also represent a partially filled area.
-            //
-            // If the op was merged in previous cycle, we don't have
-            // to count them.
-            CHECK(cow_de->new_chunk == 0);
-            break;
-        } else {
-            SNAP_LOG(ERROR) << "Error in merge operation. Found invalid metadata: "
-                            << " merged_de-old-chunk: " << merged_de->old_chunk
-                            << " merged_de-new-chunk: " << merged_de->new_chunk
-                            << " cow_de-old-chunk: " << cow_de->old_chunk
-                            << " cow_de-new-chunk: " << cow_de->new_chunk
-                            << " unmerged_exceptions: " << unmerged_exceptions
-                            << " merged_ops_cur_iter: " << merged_ops_cur_iter
-                            << " offset: " << offset;
-            return -1;
-        }
-    }
-    return merged_ops_cur_iter;
-}
-
-bool Snapuserd::ProcessMergeComplete(chunk_t chunk, void* buffer) {
-    uint32_t stride = exceptions_per_area_ + 1;
-    CowHeader header;
-
-    if (!reader_->GetHeader(&header)) {
-        SNAP_LOG(ERROR) << "Failed to get header";
-        return false;
-    }
-
-    // ChunkID to vector index
-    lldiv_t divresult = lldiv(chunk, stride);
-    CHECK(divresult.quot < vec_.size());
-    SNAP_LOG(DEBUG) << "ProcessMergeComplete: chunk: " << chunk
-                    << " Metadata-Index: " << divresult.quot;
-
-    int unmerged_exceptions = 0;
-    loff_t offset = GetMergeStartOffset(buffer, vec_[divresult.quot].get(), &unmerged_exceptions);
-
-    int merged_ops_cur_iter =
-            GetNumberOfMergedOps(buffer, vec_[divresult.quot].get(), offset, unmerged_exceptions);
-
-    // There should be at least one operation merged in this cycle
-    CHECK(merged_ops_cur_iter > 0);
-
-    header.num_merge_ops += merged_ops_cur_iter;
-    reader_->UpdateMergeProgress(merged_ops_cur_iter);
-    if (!writer_->CommitMerge(merged_ops_cur_iter)) {
-        SNAP_LOG(ERROR) << "CommitMerge failed... merged_ops_cur_iter: " << merged_ops_cur_iter;
-        return false;
-    }
-
-    SNAP_LOG(DEBUG) << "Merge success: " << merged_ops_cur_iter << "chunk: " << chunk;
-    merge_initiated_ = true;
     return true;
 }
 
@@ -589,7 +230,7 @@
 
 
         // Store operation pointer.
-        chunk_map_[ChunkToSector(data_chunk_id)] = cow_op;
+        chunk_vec_.push_back(std::make_pair(ChunkToSector(data_chunk_id), cow_op));
         num_ops += 1;
         offset += sizeof(struct disk_exception);
         cowop_riter_->Next();
@@ -781,7 +422,7 @@
             de->new_chunk = data_chunk_id;
 
             // Store operation pointer.
-            chunk_map_[ChunkToSector(data_chunk_id)] = it->second;
+            chunk_vec_.push_back(std::make_pair(ChunkToSector(data_chunk_id), it->second));
             offset += sizeof(struct disk_exception);
             num_ops += 1;
             copy_ops++;
@@ -827,6 +468,12 @@
                         << "Areas : " << vec_.size();
     }
 
+    chunk_vec_.shrink_to_fit();
+    vec_.shrink_to_fit();
+
+    // Sort the vector based on sectors as we need this during un-aligned access
+    std::sort(chunk_vec_.begin(), chunk_vec_.end(), compare);
+
     SNAP_LOG(INFO) << "ReadMetadata completed. Final-chunk-id: " << data_chunk_id
                    << " Num Sector: " << ChunkToSector(data_chunk_id)
                    << " Replace-ops: " << replace_ops << " Zero-ops: " << zero_ops
@@ -836,7 +483,6 @@
 
     // Total number of sectors required for creating dm-user device
     num_sectors_ = ChunkToSector(data_chunk_id);
-    metadata_read_done_ = true;
     merge_initiated_ = false;
     return true;
 }
@@ -850,37 +496,6 @@
     }
 }
 
-// Read Header from dm-user misc device. This gives
-// us the sector number for which IO is issued by dm-snapshot device
-bool Snapuserd::ReadDmUserHeader() {
-    if (!android::base::ReadFully(ctrl_fd_, bufsink_.GetBufPtr(), sizeof(struct dm_user_header))) {
-        SNAP_PLOG(ERROR) << "Control-read failed";
-        return false;
-    }
-
-    return true;
-}
-
-// Send the payload/data back to dm-user misc device.
-bool Snapuserd::WriteDmUserPayload(size_t size) {
-    if (!android::base::WriteFully(ctrl_fd_, bufsink_.GetBufPtr(),
-                                   sizeof(struct dm_user_header) + size)) {
-        SNAP_PLOG(ERROR) << "Write to dm-user failed size: " << size;
-        return false;
-    }
-
-    return true;
-}
-
-bool Snapuserd::ReadDmUserPayload(void* buffer, size_t size) {
-    if (!android::base::ReadFully(ctrl_fd_, buffer, size)) {
-        SNAP_PLOG(ERROR) << "ReadDmUserPayload failed size: " << size;
-        return false;
-    }
-
-    return true;
-}
-
 bool Snapuserd::InitCowDevice() {
     cow_fd_.reset(open(cow_device_.c_str(), O_RDWR));
     if (cow_fd_ < 0) {
@@ -888,186 +503,26 @@
         return false;
     }
 
-    // Allocate the buffer which is used to communicate between
-    // daemon and dm-user. The buffer comprises of header and a fixed payload.
-    // If the dm-user requests a big IO, the IO will be broken into chunks
-    // of PAYLOAD_SIZE.
-    size_t buf_size = sizeof(struct dm_user_header) + PAYLOAD_SIZE;
-    bufsink_.Initialize(buf_size);
-
     return ReadMetadata();
 }
 
-bool Snapuserd::InitBackingAndControlDevice() {
-    backing_store_fd_.reset(open(backing_store_device_.c_str(), O_RDONLY));
-    if (backing_store_fd_ < 0) {
-        SNAP_PLOG(ERROR) << "Open Failed: " << backing_store_device_;
-        return false;
+/*
+ * Entry point to launch worker threads
+ */
+bool Snapuserd::Start() {
+    std::vector<std::future<bool>> threads;
+
+    for (int i = 0; i < worker_threads_.size(); i++) {
+        threads.emplace_back(
+                std::async(std::launch::async, &WorkerThread::RunThread, worker_threads_[i].get()));
     }
 
-    ctrl_fd_.reset(open(control_device_.c_str(), O_RDWR));
-    if (ctrl_fd_ < 0) {
-        SNAP_PLOG(ERROR) << "Unable to open " << control_device_;
-        return false;
+    bool ret = true;
+    for (auto& t : threads) {
+        ret = t.get() && ret;
     }
 
-    return true;
-}
-
-bool Snapuserd::DmuserWriteRequest() {
-    struct dm_user_header* header = bufsink_.GetHeaderPtr();
-
-    // device mapper has the capability to allow
-    // targets to flush the cache when writes are completed. This
-    // is controlled by each target by a flag "flush_supported".
-    // This flag is set by dm-user. When flush is supported,
-    // a number of zero-length bio's will be submitted to
-    // the target for the purpose of flushing cache. It is the
-    // responsibility of the target driver - which is dm-user in this
-    // case, to remap these bio's to the underlying device. Since,
-    // there is no underlying device for dm-user, this zero length
-    // bio's gets routed to daemon.
-    //
-    // Flush operations are generated post merge by dm-snap by having
-    // REQ_PREFLUSH flag set. Snapuser daemon doesn't have anything
-    // to flush per se; hence, just respond back with a success message.
-    if (header->sector == 0) {
-        CHECK(header->len == 0);
-        header->type = DM_USER_RESP_SUCCESS;
-        if (!WriteDmUserPayload(0)) {
-            return false;
-        }
-        return true;
-    }
-
-    size_t remaining_size = header->len;
-    size_t read_size = std::min(PAYLOAD_SIZE, remaining_size);
-    CHECK(read_size == BLOCK_SZ);
-
-    CHECK(header->sector > 0);
-    chunk_t chunk = SectorToChunk(header->sector);
-    CHECK(chunk_map_.find(header->sector) == chunk_map_.end());
-
-    void* buffer = bufsink_.GetPayloadBuffer(read_size);
-    CHECK(buffer != nullptr);
-    header->type = DM_USER_RESP_SUCCESS;
-
-    if (!ReadDmUserPayload(buffer, read_size)) {
-        SNAP_LOG(ERROR) << "ReadDmUserPayload failed for chunk id: " << chunk
-                        << "Sector: " << header->sector;
-        header->type = DM_USER_RESP_ERROR;
-    }
-
-    if (header->type == DM_USER_RESP_SUCCESS && !ProcessMergeComplete(chunk, buffer)) {
-        SNAP_LOG(ERROR) << "ProcessMergeComplete failed for chunk id: " << chunk
-                        << "Sector: " << header->sector;
-        header->type = DM_USER_RESP_ERROR;
-    } else {
-        SNAP_LOG(DEBUG) << "ProcessMergeComplete success for chunk id: " << chunk
-                        << "Sector: " << header->sector;
-    }
-
-    if (!WriteDmUserPayload(0)) {
-        return false;
-    }
-
-    return true;
-}
-
-bool Snapuserd::DmuserReadRequest() {
-    struct dm_user_header* header = bufsink_.GetHeaderPtr();
-    size_t remaining_size = header->len;
-    loff_t offset = 0;
-    sector_t sector = header->sector;
-    do {
-        size_t read_size = std::min(PAYLOAD_SIZE, remaining_size);
-
-        int ret = read_size;
-        header->type = DM_USER_RESP_SUCCESS;
-        chunk_t chunk = SectorToChunk(header->sector);
-
-        // Request to sector 0 is always for kernel
-        // representation of COW header. This IO should be only
-        // once during dm-snapshot device creation. We should
-        // never see multiple IO requests. Additionally this IO
-        // will always be a single 4k.
-        if (header->sector == 0) {
-            CHECK(metadata_read_done_ == true);
-            CHECK(read_size == BLOCK_SZ);
-            ConstructKernelCowHeader();
-            SNAP_LOG(DEBUG) << "Kernel header constructed";
-        } else {
-            if (!offset && (read_size == BLOCK_SZ) &&
-                chunk_map_.find(header->sector) == chunk_map_.end()) {
-                if (!ReadDiskExceptions(chunk, read_size)) {
-                    SNAP_LOG(ERROR) << "ReadDiskExceptions failed for chunk id: " << chunk
-                                    << "Sector: " << header->sector;
-                    header->type = DM_USER_RESP_ERROR;
-                } else {
-                    SNAP_LOG(DEBUG) << "ReadDiskExceptions success for chunk id: " << chunk
-                                    << "Sector: " << header->sector;
-                }
-            } else {
-                chunk_t num_sectors_read = (offset >> SECTOR_SHIFT);
-                ret = ReadData(sector + num_sectors_read, read_size);
-                if (ret < 0) {
-                    SNAP_LOG(ERROR) << "ReadData failed for chunk id: " << chunk
-                                    << " Sector: " << (sector + num_sectors_read)
-                                    << " size: " << read_size << " header-len: " << header->len;
-                    header->type = DM_USER_RESP_ERROR;
-                } else {
-                    SNAP_LOG(DEBUG) << "ReadData success for chunk id: " << chunk
-                                    << "Sector: " << header->sector;
-                }
-            }
-        }
-
-        // Daemon will not be terminated if there is any error. We will
-        // just send the error back to dm-user.
-        if (!WriteDmUserPayload(ret)) {
-            return false;
-        }
-
-        remaining_size -= ret;
-        offset += ret;
-    } while (remaining_size > 0);
-
-    return true;
-}
-
-bool Snapuserd::Run() {
-    struct dm_user_header* header = bufsink_.GetHeaderPtr();
-
-    bufsink_.Clear();
-
-    if (!ReadDmUserHeader()) {
-        SNAP_LOG(ERROR) << "ReadDmUserHeader failed";
-        return false;
-    }
-
-    SNAP_LOG(DEBUG) << "msg->seq: " << std::hex << header->seq;
-    SNAP_LOG(DEBUG) << "msg->type: " << std::hex << header->type;
-    SNAP_LOG(DEBUG) << "msg->flags: " << std::hex << header->flags;
-    SNAP_LOG(DEBUG) << "msg->sector: " << std::hex << header->sector;
-    SNAP_LOG(DEBUG) << "msg->len: " << std::hex << header->len;
-
-    switch (header->type) {
-        case DM_USER_REQ_MAP_READ: {
-            if (!DmuserReadRequest()) {
-                return false;
-            }
-            break;
-        }
-
-        case DM_USER_REQ_MAP_WRITE: {
-            if (!DmuserWriteRequest()) {
-                return false;
-            }
-            break;
-        }
-    }
-
-    return true;
+    return ret;
 }
 
 }  // namespace snapshot
diff --git a/fs_mgr/libsnapshot/snapuserd.h b/fs_mgr/libsnapshot/snapuserd.h
index 518d08b..9335364 100644
--- a/fs_mgr/libsnapshot/snapuserd.h
+++ b/fs_mgr/libsnapshot/snapuserd.h
@@ -18,13 +18,17 @@
 #include <stdint.h>
 #include <stdlib.h>
 
+#include <bitset>
 #include <csignal>
 #include <cstring>
+#include <future>
 #include <iostream>
 #include <limits>
 #include <map>
+#include <mutex>
 #include <string>
 #include <thread>
+#include <unordered_map>
 #include <vector>
 
 #include <android-base/file.h>
@@ -40,6 +44,17 @@
 namespace snapshot {
 
 using android::base::unique_fd;
+using namespace std::chrono_literals;
+
+static constexpr size_t PAYLOAD_SIZE = (1UL << 20);
+static_assert(PAYLOAD_SIZE >= BLOCK_SZ);
+
+/*
+ * With 4 threads, we get optimal performance
+ * when update_verifier reads the partition during
+ * boot.
+ */
+static constexpr int NUM_THREADS_PER_PARTITION = 4;
 
 class BufferSink : public IByteSink {
   public:
@@ -59,53 +74,111 @@
     size_t buffer_size_;
 };
 
-class Snapuserd final {
+class Snapuserd;
+
+class WorkerThread {
   public:
-    Snapuserd(const std::string& misc_name, const std::string& cow_device,
-              const std::string& backing_device);
-    bool InitBackingAndControlDevice();
-    bool InitCowDevice();
-    bool Run();
-    const std::string& GetControlDevicePath() { return control_device_; }
-    const std::string& GetMiscName() { return misc_name_; }
-    uint64_t GetNumSectors() { return num_sectors_; }
-    bool IsAttached() const { return ctrl_fd_ >= 0; }
-    void CheckMergeCompletionStatus();
-    void CloseFds() {
-        ctrl_fd_ = {};
-        cow_fd_ = {};
-        backing_store_fd_ = {};
-    }
-    size_t GetMetadataAreaSize() { return vec_.size(); }
-    void* GetExceptionBuffer(size_t i) { return vec_[i].get(); }
+    WorkerThread(const std::string& cow_device, const std::string& backing_device,
+                 const std::string& control_device, const std::string& misc_name,
+                 std::shared_ptr<Snapuserd> snapuserd);
+    bool RunThread();
 
   private:
+    // Initialization
+    void InitializeBufsink();
+    bool InitializeFds();
+    bool InitReader();
+    void CloseFds() {
+        ctrl_fd_ = {};
+        backing_store_fd_ = {};
+    }
+
+    // Functions interacting with dm-user
+    bool ReadDmUserHeader();
     bool DmuserReadRequest();
     bool DmuserWriteRequest();
-
-    bool ReadDmUserHeader();
     bool ReadDmUserPayload(void* buffer, size_t size);
     bool WriteDmUserPayload(size_t size);
-    void ConstructKernelCowHeader();
-    bool ReadMetadata();
-    bool ZerofillDiskExceptions(size_t read_size);
-    bool ReadDiskExceptions(chunk_t chunk, size_t size);
-    int ReadUnalignedSector(sector_t sector, size_t size,
-                            std::map<sector_t, const CowOperation*>::iterator& it);
-    int ReadData(sector_t sector, size_t size);
-    bool IsChunkIdMetadata(chunk_t chunk);
-    chunk_t GetNextAllocatableChunkId(chunk_t chunk_id);
 
+    bool ReadDiskExceptions(chunk_t chunk, size_t size);
+    bool ZerofillDiskExceptions(size_t read_size);
+    void ConstructKernelCowHeader();
+
+    // IO Path
+    bool ProcessIORequest();
+    int ReadData(sector_t sector, size_t size);
+    int ReadUnalignedSector(sector_t sector, size_t size,
+                            std::vector<std::pair<sector_t, const CowOperation*>>::iterator& it);
+
+    // Processing COW operations
     bool ProcessCowOp(const CowOperation* cow_op);
     bool ProcessReplaceOp(const CowOperation* cow_op);
     bool ProcessCopyOp(const CowOperation* cow_op);
     bool ProcessZeroOp();
 
+    // Merge related functions
+    bool ProcessMergeComplete(chunk_t chunk, void* buffer);
     loff_t GetMergeStartOffset(void* merged_buffer, void* unmerged_buffer,
                                int* unmerged_exceptions);
     int GetNumberOfMergedOps(void* merged_buffer, void* unmerged_buffer, loff_t offset,
                              int unmerged_exceptions);
-    bool ProcessMergeComplete(chunk_t chunk, void* buffer);
+
+    sector_t ChunkToSector(chunk_t chunk) { return chunk << CHUNK_SHIFT; }
+    chunk_t SectorToChunk(sector_t sector) { return sector >> CHUNK_SHIFT; }
+
+    std::unique_ptr<CowReader> reader_;
+    BufferSink bufsink_;
+
+    std::string cow_device_;
+    std::string backing_store_device_;
+    std::string control_device_;
+    std::string misc_name_;
+
+    unique_fd cow_fd_;
+    unique_fd backing_store_fd_;
+    unique_fd ctrl_fd_;
+
+    std::shared_ptr<Snapuserd> snapuserd_;
+    uint32_t exceptions_per_area_;
+};
+
+class Snapuserd : public std::enable_shared_from_this<Snapuserd> {
+  public:
+    Snapuserd(const std::string& misc_name, const std::string& cow_device,
+              const std::string& backing_device);
+    bool InitCowDevice();
+    bool Start();
+    const std::string& GetControlDevicePath() { return control_device_; }
+    const std::string& GetMiscName() { return misc_name_; }
+    uint64_t GetNumSectors() { return num_sectors_; }
+    bool IsAttached() const { return attached_; }
+    void AttachControlDevice() { attached_ = true; }
+
+    void CheckMergeCompletionStatus();
+    bool CommitMerge(int num_merge_ops);
+
+    void CloseFds() { cow_fd_ = {}; }
+    size_t GetMetadataAreaSize() { return vec_.size(); }
+    void* GetExceptionBuffer(size_t i) { return vec_[i].get(); }
+
+    bool InitializeWorkers();
+    std::shared_ptr<Snapuserd> GetSharedPtr() { return shared_from_this(); }
+
+    std::vector<std::pair<sector_t, const CowOperation*>>& GetChunkVec() { return chunk_vec_; }
+    const std::vector<std::unique_ptr<uint8_t[]>>& GetMetadataVec() const { return vec_; }
+
+    static bool compare(std::pair<sector_t, const CowOperation*> p1,
+                        std::pair<sector_t, const CowOperation*> p2) {
+        return p1.first < p2.first;
+    }
+
+  private:
+    std::vector<std::unique_ptr<WorkerThread>> worker_threads_;
+
+    bool IsChunkIdMetadata(chunk_t chunk);
+    chunk_t GetNextAllocatableChunkId(chunk_t chunk_id);
+
+    bool ReadMetadata();
     sector_t ChunkToSector(chunk_t chunk) { return chunk << CHUNK_SHIFT; }
     chunk_t SectorToChunk(sector_t sector) { return sector >> CHUNK_SHIFT; }
     bool IsBlockAligned(int read_size) { return ((read_size & (BLOCK_SZ - 1)) == 0); }
@@ -116,8 +189,6 @@
     std::string misc_name_;
 
     unique_fd cow_fd_;
-    unique_fd backing_store_fd_;
-    unique_fd ctrl_fd_;
 
     uint32_t exceptions_per_area_;
     uint64_t num_sectors_;
@@ -131,19 +202,14 @@
     // mapping of old-chunk to new-chunk
     std::vector<std::unique_ptr<uint8_t[]>> vec_;
 
-    // Key - Sector
-    // Value - cow operation
-    //
-    // chunk_map stores the pseudo mapping of sector
-    // to COW operations. Each COW op is 4k; however,
-    // we can get a read request which are as small
-    // as 512 bytes. Hence, we need to binary search
-    // in the chunk_map to find the nearest COW op.
-    std::map<sector_t, const CowOperation*> chunk_map_;
+    // chunk_vec stores the pseudo mapping of sector
+    // to COW operations.
+    std::vector<std::pair<sector_t, const CowOperation*>> chunk_vec_;
 
-    bool metadata_read_done_ = false;
+    std::mutex lock_;
+
     bool merge_initiated_ = false;
-    BufferSink bufsink_;
+    bool attached_ = false;
 };
 
 }  // namespace snapshot
diff --git a/fs_mgr/libsnapshot/snapuserd_server.cpp b/fs_mgr/libsnapshot/snapuserd_server.cpp
index 017de3b..167895e 100644
--- a/fs_mgr/libsnapshot/snapuserd_server.cpp
+++ b/fs_mgr/libsnapshot/snapuserd_server.cpp
@@ -77,8 +77,8 @@
     JoinAllThreads();
 }
 
-DmUserHandler::DmUserHandler(std::unique_ptr<Snapuserd>&& snapuserd)
-    : snapuserd_(std::move(snapuserd)), misc_name_(snapuserd_->GetMiscName()) {}
+DmUserHandler::DmUserHandler(std::shared_ptr<Snapuserd> snapuserd)
+    : snapuserd_(snapuserd), misc_name_(snapuserd_->GetMiscName()) {}
 
 bool SnapuserdServer::Sendmsg(android::base::borrowed_fd fd, const std::string& msg) {
     ssize_t ret = TEMP_FAILURE_RETRY(send(fd.get(), msg.data(), msg.size(), 0));
@@ -204,10 +204,8 @@
 void SnapuserdServer::RunThread(std::shared_ptr<DmUserHandler> handler) {
     LOG(INFO) << "Entering thread for handler: " << handler->misc_name();
 
-    while (!StopRequested()) {
-        if (!handler->snapuserd()->Run()) {
-            break;
-        }
+    if (!handler->snapuserd()->Start()) {
+        LOG(ERROR) << " Failed to launch all worker threads";
     }
 
     handler->snapuserd()->CloseFds();
@@ -349,13 +347,18 @@
 std::shared_ptr<DmUserHandler> SnapuserdServer::AddHandler(const std::string& misc_name,
                                                            const std::string& cow_device_path,
                                                            const std::string& backing_device) {
-    auto snapuserd = std::make_unique<Snapuserd>(misc_name, cow_device_path, backing_device);
+    auto snapuserd = std::make_shared<Snapuserd>(misc_name, cow_device_path, backing_device);
     if (!snapuserd->InitCowDevice()) {
         LOG(ERROR) << "Failed to initialize Snapuserd";
         return nullptr;
     }
 
-    auto handler = std::make_shared<DmUserHandler>(std::move(snapuserd));
+    if (!snapuserd->InitializeWorkers()) {
+        LOG(ERROR) << "Failed to initialize workers";
+        return nullptr;
+    }
+
+    auto handler = std::make_shared<DmUserHandler>(snapuserd);
     {
         std::lock_guard<std::mutex> lock(lock_);
         if (FindHandler(&lock, misc_name) != dm_users_.end()) {
@@ -370,10 +373,7 @@
 bool SnapuserdServer::StartHandler(const std::shared_ptr<DmUserHandler>& handler) {
     CHECK(!handler->snapuserd()->IsAttached());
 
-    if (!handler->snapuserd()->InitBackingAndControlDevice()) {
-        LOG(ERROR) << "Failed to initialize control device: " << handler->misc_name();
-        return false;
-    }
+    handler->snapuserd()->AttachControlDevice();
 
     handler->thread() = std::thread(std::bind(&SnapuserdServer::RunThread, this, handler));
     return true;
diff --git a/fs_mgr/libsnapshot/snapuserd_server.h b/fs_mgr/libsnapshot/snapuserd_server.h
index 7cbc2de..e9d575d 100644
--- a/fs_mgr/libsnapshot/snapuserd_server.h
+++ b/fs_mgr/libsnapshot/snapuserd_server.h
@@ -47,17 +47,17 @@
 
 class DmUserHandler {
   public:
-    explicit DmUserHandler(std::unique_ptr<Snapuserd>&& snapuserd);
+    explicit DmUserHandler(std::shared_ptr<Snapuserd> snapuserd);
 
     void FreeResources() { snapuserd_ = nullptr; }
-    const std::unique_ptr<Snapuserd>& snapuserd() const { return snapuserd_; }
+    const std::shared_ptr<Snapuserd>& snapuserd() const { return snapuserd_; }
     std::thread& thread() { return thread_; }
 
     const std::string& misc_name() const { return misc_name_; }
 
   private:
     std::thread thread_;
-    std::unique_ptr<Snapuserd> snapuserd_;
+    std::shared_ptr<Snapuserd> snapuserd_;
     std::string misc_name_;
 };
 
diff --git a/fs_mgr/libsnapshot/snapuserd_worker.cpp b/fs_mgr/libsnapshot/snapuserd_worker.cpp
new file mode 100644
index 0000000..1002569
--- /dev/null
+++ b/fs_mgr/libsnapshot/snapuserd_worker.cpp
@@ -0,0 +1,695 @@
+/*
+ * Copyright (C) 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.
+ */
+
+#include "snapuserd.h"
+
+#include <csignal>
+#include <optional>
+#include <set>
+
+#include <libsnapshot/snapuserd_client.h>
+
+namespace android {
+namespace snapshot {
+
+using namespace android;
+using namespace android::dm;
+using android::base::unique_fd;
+
+#define SNAP_LOG(level) LOG(level) << misc_name_ << ": "
+#define SNAP_PLOG(level) PLOG(level) << misc_name_ << ": "
+
+void BufferSink::Initialize(size_t size) {
+    buffer_size_ = size;
+    buffer_offset_ = 0;
+    buffer_ = std::make_unique<uint8_t[]>(size);
+}
+
+void* BufferSink::GetPayloadBuffer(size_t size) {
+    if ((buffer_size_ - buffer_offset_) < size) return nullptr;
+
+    char* buffer = reinterpret_cast<char*>(GetBufPtr());
+    struct dm_user_message* msg = (struct dm_user_message*)(&(buffer[0]));
+    return (char*)msg->payload.buf + buffer_offset_;
+}
+
+void* BufferSink::GetBuffer(size_t requested, size_t* actual) {
+    void* buf = GetPayloadBuffer(requested);
+    if (!buf) {
+        *actual = 0;
+        return nullptr;
+    }
+    *actual = requested;
+    return buf;
+}
+
+struct dm_user_header* BufferSink::GetHeaderPtr() {
+    CHECK(sizeof(struct dm_user_header) <= buffer_size_);
+    char* buf = reinterpret_cast<char*>(GetBufPtr());
+    struct dm_user_header* header = (struct dm_user_header*)(&(buf[0]));
+    return header;
+}
+
+WorkerThread::WorkerThread(const std::string& cow_device, const std::string& backing_device,
+                           const std::string& control_device, const std::string& misc_name,
+                           std::shared_ptr<Snapuserd> snapuserd) {
+    cow_device_ = cow_device;
+    backing_store_device_ = backing_device;
+    control_device_ = control_device;
+    misc_name_ = misc_name;
+    snapuserd_ = snapuserd;
+    exceptions_per_area_ = (CHUNK_SIZE << SECTOR_SHIFT) / sizeof(struct disk_exception);
+}
+
+bool WorkerThread::InitializeFds() {
+    backing_store_fd_.reset(open(backing_store_device_.c_str(), O_RDONLY));
+    if (backing_store_fd_ < 0) {
+        SNAP_PLOG(ERROR) << "Open Failed: " << backing_store_device_;
+        return false;
+    }
+
+    cow_fd_.reset(open(cow_device_.c_str(), O_RDWR));
+    if (cow_fd_ < 0) {
+        SNAP_PLOG(ERROR) << "Open Failed: " << cow_device_;
+        return false;
+    }
+
+    ctrl_fd_.reset(open(control_device_.c_str(), O_RDWR));
+    if (ctrl_fd_ < 0) {
+        SNAP_PLOG(ERROR) << "Unable to open " << control_device_;
+        return false;
+    }
+
+    return true;
+}
+
+bool WorkerThread::InitReader() {
+    reader_ = std::make_unique<CowReader>();
+    if (!reader_->InitForMerge(std::move(cow_fd_))) {
+        return false;
+    }
+
+    return true;
+}
+
+// Construct kernel COW header in memory
+// This header will be in sector 0. The IO
+// request will always be 4k. After constructing
+// the header, zero out the remaining block.
+void WorkerThread::ConstructKernelCowHeader() {
+    void* buffer = bufsink_.GetPayloadBuffer(BLOCK_SZ);
+    CHECK(buffer != nullptr);
+
+    memset(buffer, 0, BLOCK_SZ);
+
+    struct disk_header* dh = reinterpret_cast<struct disk_header*>(buffer);
+
+    dh->magic = SNAP_MAGIC;
+    dh->valid = SNAPSHOT_VALID;
+    dh->version = SNAPSHOT_DISK_VERSION;
+    dh->chunk_size = CHUNK_SIZE;
+}
+
+// Start the replace operation. This will read the
+// internal COW format and if the block is compressed,
+// it will be de-compressed.
+bool WorkerThread::ProcessReplaceOp(const CowOperation* cow_op) {
+    if (!reader_->ReadData(*cow_op, &bufsink_)) {
+        SNAP_LOG(ERROR) << "ProcessReplaceOp failed for block " << cow_op->new_block;
+        return false;
+    }
+
+    return true;
+}
+
+// Start the copy operation. This will read the backing
+// block device which is represented by cow_op->source.
+bool WorkerThread::ProcessCopyOp(const CowOperation* cow_op) {
+    void* buffer = bufsink_.GetPayloadBuffer(BLOCK_SZ);
+    CHECK(buffer != nullptr);
+
+    // Issue a single 4K IO. However, this can be optimized
+    // if the successive blocks are contiguous.
+    if (!android::base::ReadFullyAtOffset(backing_store_fd_, buffer, BLOCK_SZ,
+                                          cow_op->source * BLOCK_SZ)) {
+        SNAP_PLOG(ERROR) << "Copy-op failed. Read from backing store: " << backing_store_device_
+                         << "at block :" << cow_op->source;
+        return false;
+    }
+
+    return true;
+}
+
+bool WorkerThread::ProcessZeroOp() {
+    // Zero out the entire block
+    void* buffer = bufsink_.GetPayloadBuffer(BLOCK_SZ);
+    CHECK(buffer != nullptr);
+
+    memset(buffer, 0, BLOCK_SZ);
+    return true;
+}
+
+bool WorkerThread::ProcessCowOp(const CowOperation* cow_op) {
+    CHECK(cow_op != nullptr);
+
+    switch (cow_op->type) {
+        case kCowReplaceOp: {
+            return ProcessReplaceOp(cow_op);
+        }
+
+        case kCowZeroOp: {
+            return ProcessZeroOp();
+        }
+
+        case kCowCopyOp: {
+            return ProcessCopyOp(cow_op);
+        }
+
+        default: {
+            SNAP_LOG(ERROR) << "Unknown operation-type found: " << cow_op->type;
+        }
+    }
+    return false;
+}
+
+int WorkerThread::ReadUnalignedSector(
+        sector_t sector, size_t size,
+        std::vector<std::pair<sector_t, const CowOperation*>>::iterator& it) {
+    size_t skip_sector_size = 0;
+
+    SNAP_LOG(DEBUG) << "ReadUnalignedSector: sector " << sector << " size: " << size
+                    << " Aligned sector: " << it->first;
+
+    if (!ProcessCowOp(it->second)) {
+        SNAP_LOG(ERROR) << "ReadUnalignedSector: " << sector << " failed of size: " << size;
+        return -1;
+    }
+
+    int num_sectors_skip = sector - it->first;
+
+    if (num_sectors_skip > 0) {
+        skip_sector_size = num_sectors_skip << SECTOR_SHIFT;
+        char* buffer = reinterpret_cast<char*>(bufsink_.GetBufPtr());
+        struct dm_user_message* msg = (struct dm_user_message*)(&(buffer[0]));
+
+        memmove(msg->payload.buf, (char*)msg->payload.buf + skip_sector_size,
+                (BLOCK_SZ - skip_sector_size));
+    }
+
+    bufsink_.ResetBufferOffset();
+    return std::min(size, (BLOCK_SZ - skip_sector_size));
+}
+
+/*
+ * Read the data for a given COW Operation.
+ *
+ * Kernel can issue IO at a sector granularity.
+ * Hence, an IO may end up with reading partial
+ * data from a COW operation or we may also
+ * end up with interspersed request between
+ * two COW operations.
+ *
+ */
+int WorkerThread::ReadData(sector_t sector, size_t size) {
+    std::vector<std::pair<sector_t, const CowOperation*>>& chunk_vec = snapuserd_->GetChunkVec();
+    std::vector<std::pair<sector_t, const CowOperation*>>::iterator it;
+    /*
+     * chunk_map stores COW operation at 4k granularity.
+     * If the requested IO with the sector falls on the 4k
+     * boundary, then we can read the COW op directly without
+     * any issue.
+     *
+     * However, if the requested sector is not 4K aligned,
+     * then we will have the find the nearest COW operation
+     * and chop the 4K block to fetch the requested sector.
+     */
+    it = std::lower_bound(chunk_vec.begin(), chunk_vec.end(), std::make_pair(sector, nullptr),
+                          Snapuserd::compare);
+
+    CHECK(it != chunk_vec.end());
+
+    // We didn't find the required sector; hence find the previous sector
+    // as lower_bound will gives us the value greater than
+    // the requested sector
+    if (it->first != sector) {
+        if (it != chunk_vec.begin()) {
+            --it;
+        }
+
+        /*
+         * If the IO is spanned between two COW operations,
+         * split the IO into two parts:
+         *
+         * 1: Read the first part from the single COW op
+         * 2: Read the second part from the next COW op.
+         *
+         * Ex: Let's say we have a 1024 Bytes IO request.
+         *
+         * 0       COW OP-1  4096     COW OP-2  8192
+         * |******************|*******************|
+         *              |*****|*****|
+         *           3584           4608
+         *              <- 1024B - >
+         *
+         * We have two COW operations which are 4k blocks.
+         * The IO is requested for 1024 Bytes which are spanned
+         * between two COW operations. We will split this IO
+         * into two parts:
+         *
+         * 1: IO of size 512B from offset 3584 bytes (COW OP-1)
+         * 2: IO of size 512B from offset 4096 bytes (COW OP-2)
+         */
+        return ReadUnalignedSector(sector, size, it);
+    }
+
+    int num_ops = DIV_ROUND_UP(size, BLOCK_SZ);
+    while (num_ops) {
+        if (!ProcessCowOp(it->second)) {
+            return -1;
+        }
+        num_ops -= 1;
+        it++;
+        // Update the buffer offset
+        bufsink_.UpdateBufferOffset(BLOCK_SZ);
+
+        SNAP_LOG(DEBUG) << "ReadData at sector: " << sector << " size: " << size;
+    }
+
+    // Reset the buffer offset
+    bufsink_.ResetBufferOffset();
+    return size;
+}
+
+/*
+ * dm-snap does prefetch reads while reading disk-exceptions.
+ * By default, prefetch value is set to 12; this means that
+ * dm-snap will issue 12 areas wherein each area is a 4k page
+ * of disk-exceptions.
+ *
+ * If during prefetch, if the chunk-id seen is beyond the
+ * actual number of metadata page, fill the buffer with zero.
+ * When dm-snap starts parsing the buffer, it will stop
+ * reading metadata page once the buffer content is zero.
+ */
+bool WorkerThread::ZerofillDiskExceptions(size_t read_size) {
+    size_t size = exceptions_per_area_ * sizeof(struct disk_exception);
+
+    if (read_size > size) {
+        return false;
+    }
+
+    void* buffer = bufsink_.GetPayloadBuffer(size);
+    CHECK(buffer != nullptr);
+
+    memset(buffer, 0, size);
+    return true;
+}
+
+/*
+ * A disk exception is a simple mapping of old_chunk to new_chunk.
+ * When dm-snapshot device is created, kernel requests these mapping.
+ *
+ * Each disk exception is of size 16 bytes. Thus a single 4k page can
+ * have:
+ *
+ * exceptions_per_area_ = 4096/16 = 256. This entire 4k page
+ * is considered a metadata page and it is represented by chunk ID.
+ *
+ * Convert the chunk ID to index into the vector which gives us
+ * the metadata page.
+ */
+bool WorkerThread::ReadDiskExceptions(chunk_t chunk, size_t read_size) {
+    uint32_t stride = exceptions_per_area_ + 1;
+    size_t size;
+    const std::vector<std::unique_ptr<uint8_t[]>>& vec = snapuserd_->GetMetadataVec();
+
+    // ChunkID to vector index
+    lldiv_t divresult = lldiv(chunk, stride);
+
+    if (divresult.quot < vec.size()) {
+        size = exceptions_per_area_ * sizeof(struct disk_exception);
+
+        CHECK(read_size == size);
+
+        void* buffer = bufsink_.GetPayloadBuffer(size);
+        CHECK(buffer != nullptr);
+
+        memcpy(buffer, vec[divresult.quot].get(), size);
+    } else {
+        return ZerofillDiskExceptions(read_size);
+    }
+
+    return true;
+}
+
+loff_t WorkerThread::GetMergeStartOffset(void* merged_buffer, void* unmerged_buffer,
+                                         int* unmerged_exceptions) {
+    loff_t offset = 0;
+    *unmerged_exceptions = 0;
+
+    while (*unmerged_exceptions <= exceptions_per_area_) {
+        struct disk_exception* merged_de =
+                reinterpret_cast<struct disk_exception*>((char*)merged_buffer + offset);
+        struct disk_exception* cow_de =
+                reinterpret_cast<struct disk_exception*>((char*)unmerged_buffer + offset);
+
+        // Unmerged op by the kernel
+        if (merged_de->old_chunk != 0 || merged_de->new_chunk != 0) {
+            CHECK(merged_de->old_chunk == cow_de->old_chunk);
+            CHECK(merged_de->new_chunk == cow_de->new_chunk);
+
+            offset += sizeof(struct disk_exception);
+            *unmerged_exceptions += 1;
+            continue;
+        }
+
+        break;
+    }
+
+    CHECK(!(*unmerged_exceptions == exceptions_per_area_));
+
+    SNAP_LOG(DEBUG) << "Unmerged_Exceptions: " << *unmerged_exceptions << " Offset: " << offset;
+    return offset;
+}
+
+int WorkerThread::GetNumberOfMergedOps(void* merged_buffer, void* unmerged_buffer, loff_t offset,
+                                       int unmerged_exceptions) {
+    int merged_ops_cur_iter = 0;
+    std::vector<std::pair<sector_t, const CowOperation*>>& chunk_vec = snapuserd_->GetChunkVec();
+
+    // Find the operations which are merged in this cycle.
+    while ((unmerged_exceptions + merged_ops_cur_iter) < exceptions_per_area_) {
+        struct disk_exception* merged_de =
+                reinterpret_cast<struct disk_exception*>((char*)merged_buffer + offset);
+        struct disk_exception* cow_de =
+                reinterpret_cast<struct disk_exception*>((char*)unmerged_buffer + offset);
+
+        CHECK(merged_de->new_chunk == 0);
+        CHECK(merged_de->old_chunk == 0);
+
+        if (cow_de->new_chunk != 0) {
+            merged_ops_cur_iter += 1;
+            offset += sizeof(struct disk_exception);
+            auto it = std::lower_bound(chunk_vec.begin(), chunk_vec.end(),
+                                       std::make_pair(ChunkToSector(cow_de->new_chunk), nullptr),
+                                       Snapuserd::compare);
+            CHECK(it != chunk_vec.end());
+            CHECK(it->first == ChunkToSector(cow_de->new_chunk));
+            const CowOperation* cow_op = it->second;
+
+            CHECK(cow_op != nullptr);
+
+            CHECK(cow_op->new_block == cow_de->old_chunk);
+            // zero out to indicate that operation is merged.
+            cow_de->old_chunk = 0;
+            cow_de->new_chunk = 0;
+        } else if (cow_de->old_chunk == 0) {
+            // Already merged op in previous iteration or
+            // This could also represent a partially filled area.
+            //
+            // If the op was merged in previous cycle, we don't have
+            // to count them.
+            CHECK(cow_de->new_chunk == 0);
+            break;
+        } else {
+            SNAP_LOG(ERROR) << "Error in merge operation. Found invalid metadata: "
+                            << " merged_de-old-chunk: " << merged_de->old_chunk
+                            << " merged_de-new-chunk: " << merged_de->new_chunk
+                            << " cow_de-old-chunk: " << cow_de->old_chunk
+                            << " cow_de-new-chunk: " << cow_de->new_chunk
+                            << " unmerged_exceptions: " << unmerged_exceptions
+                            << " merged_ops_cur_iter: " << merged_ops_cur_iter
+                            << " offset: " << offset;
+            return -1;
+        }
+    }
+    return merged_ops_cur_iter;
+}
+
+bool WorkerThread::ProcessMergeComplete(chunk_t chunk, void* buffer) {
+    uint32_t stride = exceptions_per_area_ + 1;
+    const std::vector<std::unique_ptr<uint8_t[]>>& vec = snapuserd_->GetMetadataVec();
+
+    // ChunkID to vector index
+    lldiv_t divresult = lldiv(chunk, stride);
+    CHECK(divresult.quot < vec.size());
+    SNAP_LOG(DEBUG) << "ProcessMergeComplete: chunk: " << chunk
+                    << " Metadata-Index: " << divresult.quot;
+
+    int unmerged_exceptions = 0;
+    loff_t offset = GetMergeStartOffset(buffer, vec[divresult.quot].get(), &unmerged_exceptions);
+
+    int merged_ops_cur_iter =
+            GetNumberOfMergedOps(buffer, vec[divresult.quot].get(), offset, unmerged_exceptions);
+
+    // There should be at least one operation merged in this cycle
+    CHECK(merged_ops_cur_iter > 0);
+    if (!snapuserd_->CommitMerge(merged_ops_cur_iter)) {
+        return false;
+    }
+
+    SNAP_LOG(DEBUG) << "Merge success: " << merged_ops_cur_iter << "chunk: " << chunk;
+    return true;
+}
+
+// Read Header from dm-user misc device. This gives
+// us the sector number for which IO is issued by dm-snapshot device
+bool WorkerThread::ReadDmUserHeader() {
+    if (!android::base::ReadFully(ctrl_fd_, bufsink_.GetBufPtr(), sizeof(struct dm_user_header))) {
+        if (errno != ENOTBLK) {
+            SNAP_PLOG(ERROR) << "Control-read failed";
+        }
+        return false;
+    }
+
+    return true;
+}
+
+// Send the payload/data back to dm-user misc device.
+bool WorkerThread::WriteDmUserPayload(size_t size) {
+    if (!android::base::WriteFully(ctrl_fd_, bufsink_.GetBufPtr(),
+                                   sizeof(struct dm_user_header) + size)) {
+        SNAP_PLOG(ERROR) << "Write to dm-user failed size: " << size;
+        return false;
+    }
+
+    return true;
+}
+
+bool WorkerThread::ReadDmUserPayload(void* buffer, size_t size) {
+    if (!android::base::ReadFully(ctrl_fd_, buffer, size)) {
+        SNAP_PLOG(ERROR) << "ReadDmUserPayload failed size: " << size;
+        return false;
+    }
+
+    return true;
+}
+
+bool WorkerThread::DmuserWriteRequest() {
+    struct dm_user_header* header = bufsink_.GetHeaderPtr();
+
+    // device mapper has the capability to allow
+    // targets to flush the cache when writes are completed. This
+    // is controlled by each target by a flag "flush_supported".
+    // This flag is set by dm-user. When flush is supported,
+    // a number of zero-length bio's will be submitted to
+    // the target for the purpose of flushing cache. It is the
+    // responsibility of the target driver - which is dm-user in this
+    // case, to remap these bio's to the underlying device. Since,
+    // there is no underlying device for dm-user, this zero length
+    // bio's gets routed to daemon.
+    //
+    // Flush operations are generated post merge by dm-snap by having
+    // REQ_PREFLUSH flag set. Snapuser daemon doesn't have anything
+    // to flush per se; hence, just respond back with a success message.
+    if (header->sector == 0) {
+        CHECK(header->len == 0);
+        header->type = DM_USER_RESP_SUCCESS;
+        if (!WriteDmUserPayload(0)) {
+            return false;
+        }
+        return true;
+    }
+
+    std::vector<std::pair<sector_t, const CowOperation*>>& chunk_vec = snapuserd_->GetChunkVec();
+    size_t remaining_size = header->len;
+    size_t read_size = std::min(PAYLOAD_SIZE, remaining_size);
+    CHECK(read_size == BLOCK_SZ) << "DmuserWriteRequest: read_size: " << read_size;
+
+    CHECK(header->sector > 0);
+    chunk_t chunk = SectorToChunk(header->sector);
+    auto it = std::lower_bound(chunk_vec.begin(), chunk_vec.end(),
+                               std::make_pair(header->sector, nullptr), Snapuserd::compare);
+
+    bool not_found = (it == chunk_vec.end() || it->first != header->sector);
+    CHECK(not_found);
+
+    void* buffer = bufsink_.GetPayloadBuffer(read_size);
+    CHECK(buffer != nullptr);
+    header->type = DM_USER_RESP_SUCCESS;
+
+    if (!ReadDmUserPayload(buffer, read_size)) {
+        SNAP_LOG(ERROR) << "ReadDmUserPayload failed for chunk id: " << chunk
+                        << "Sector: " << header->sector;
+        header->type = DM_USER_RESP_ERROR;
+    }
+
+    if (header->type == DM_USER_RESP_SUCCESS && !ProcessMergeComplete(chunk, buffer)) {
+        SNAP_LOG(ERROR) << "ProcessMergeComplete failed for chunk id: " << chunk
+                        << "Sector: " << header->sector;
+        header->type = DM_USER_RESP_ERROR;
+    } else {
+        SNAP_LOG(DEBUG) << "ProcessMergeComplete success for chunk id: " << chunk
+                        << "Sector: " << header->sector;
+    }
+
+    if (!WriteDmUserPayload(0)) {
+        return false;
+    }
+
+    return true;
+}
+
+bool WorkerThread::DmuserReadRequest() {
+    struct dm_user_header* header = bufsink_.GetHeaderPtr();
+    size_t remaining_size = header->len;
+    loff_t offset = 0;
+    sector_t sector = header->sector;
+    std::vector<std::pair<sector_t, const CowOperation*>>& chunk_vec = snapuserd_->GetChunkVec();
+    do {
+        size_t read_size = std::min(PAYLOAD_SIZE, remaining_size);
+
+        int ret = read_size;
+        header->type = DM_USER_RESP_SUCCESS;
+        chunk_t chunk = SectorToChunk(header->sector);
+
+        // Request to sector 0 is always for kernel
+        // representation of COW header. This IO should be only
+        // once during dm-snapshot device creation. We should
+        // never see multiple IO requests. Additionally this IO
+        // will always be a single 4k.
+        if (header->sector == 0) {
+            CHECK(read_size == BLOCK_SZ) << " Sector 0 read request of size: " << read_size;
+            ConstructKernelCowHeader();
+            SNAP_LOG(DEBUG) << "Kernel header constructed";
+        } else {
+            auto it = std::lower_bound(chunk_vec.begin(), chunk_vec.end(),
+                                       std::make_pair(header->sector, nullptr), Snapuserd::compare);
+            bool not_found = (it == chunk_vec.end() || it->first != header->sector);
+            if (!offset && (read_size == BLOCK_SZ) && not_found) {
+                if (!ReadDiskExceptions(chunk, read_size)) {
+                    SNAP_LOG(ERROR) << "ReadDiskExceptions failed for chunk id: " << chunk
+                                    << "Sector: " << header->sector;
+                    header->type = DM_USER_RESP_ERROR;
+                } else {
+                    SNAP_LOG(DEBUG) << "ReadDiskExceptions success for chunk id: " << chunk
+                                    << "Sector: " << header->sector;
+                }
+            } else {
+                chunk_t num_sectors_read = (offset >> SECTOR_SHIFT);
+                ret = ReadData(sector + num_sectors_read, read_size);
+                if (ret < 0) {
+                    SNAP_LOG(ERROR) << "ReadData failed for chunk id: " << chunk
+                                    << " Sector: " << (sector + num_sectors_read)
+                                    << " size: " << read_size << " header-len: " << header->len;
+                    header->type = DM_USER_RESP_ERROR;
+                } else {
+                    SNAP_LOG(DEBUG) << "ReadData success for chunk id: " << chunk
+                                    << "Sector: " << header->sector;
+                }
+            }
+        }
+
+        // Daemon will not be terminated if there is any error. We will
+        // just send the error back to dm-user.
+        if (!WriteDmUserPayload(ret)) {
+            return false;
+        }
+
+        remaining_size -= ret;
+        offset += ret;
+    } while (remaining_size > 0);
+
+    return true;
+}
+
+void WorkerThread::InitializeBufsink() {
+    // Allocate the buffer which is used to communicate between
+    // daemon and dm-user. The buffer comprises of header and a fixed payload.
+    // If the dm-user requests a big IO, the IO will be broken into chunks
+    // of PAYLOAD_SIZE.
+    size_t buf_size = sizeof(struct dm_user_header) + PAYLOAD_SIZE;
+    bufsink_.Initialize(buf_size);
+}
+
+bool WorkerThread::RunThread() {
+    InitializeBufsink();
+
+    if (!InitializeFds()) {
+        return false;
+    }
+
+    if (!InitReader()) {
+        return false;
+    }
+
+    // Start serving IO
+    while (true) {
+        if (!ProcessIORequest()) {
+            break;
+        }
+    }
+
+    CloseFds();
+    reader_->CloseCowFd();
+
+    return true;
+}
+
+bool WorkerThread::ProcessIORequest() {
+    struct dm_user_header* header = bufsink_.GetHeaderPtr();
+
+    if (!ReadDmUserHeader()) {
+        return false;
+    }
+
+    SNAP_LOG(DEBUG) << "msg->seq: " << std::hex << header->seq;
+    SNAP_LOG(DEBUG) << "msg->type: " << std::hex << header->type;
+    SNAP_LOG(DEBUG) << "msg->flags: " << std::hex << header->flags;
+    SNAP_LOG(DEBUG) << "msg->sector: " << std::hex << header->sector;
+    SNAP_LOG(DEBUG) << "msg->len: " << std::hex << header->len;
+
+    switch (header->type) {
+        case DM_USER_REQ_MAP_READ: {
+            if (!DmuserReadRequest()) {
+                return false;
+            }
+            break;
+        }
+
+        case DM_USER_REQ_MAP_WRITE: {
+            if (!DmuserWriteRequest()) {
+                return false;
+            }
+            break;
+        }
+    }
+
+    return true;
+}
+
+}  // namespace snapshot
+}  // namespace android
diff --git a/fs_mgr/libstorage_literals/Android.bp b/fs_mgr/libstorage_literals/Android.bp
index 635ca49..5b07168 100644
--- a/fs_mgr/libstorage_literals/Android.bp
+++ b/fs_mgr/libstorage_literals/Android.bp
@@ -8,4 +8,9 @@
     host_supported: true,
     recovery_available: true,
     export_include_dirs: ["."],
+    target: {
+        windows: {
+            enabled: true,
+        },
+    },
 }
diff --git a/fs_mgr/tests/fs_mgr_test.cpp b/fs_mgr/tests/fs_mgr_test.cpp
index bcd5180..5887641 100644
--- a/fs_mgr/tests/fs_mgr_test.cpp
+++ b/fs_mgr/tests/fs_mgr_test.cpp
@@ -127,7 +127,7 @@
         "androidboot.serialno = \"BLAHBLAHBLAH\"\n"
         "androidboot.slot_suffix = \"_a\"\n"
         "androidboot.hardware.platform = \"sdw813\"\n"
-        "androidboot.hardware = \"foo\"\n"
+        "hardware = \"foo\"\n"
         "androidboot.revision = \"EVT1.0\"\n"
         "androidboot.bootloader = \"burp-0.1-7521\"\n"
         "androidboot.hardware.sku = \"mary\"\n"
@@ -159,7 +159,7 @@
         {"androidboot.serialno", "BLAHBLAHBLAH"},
         {"androidboot.slot_suffix", "_a"},
         {"androidboot.hardware.platform", "sdw813"},
-        {"androidboot.hardware", "foo"},
+        {"hardware", "foo"},
         {"androidboot.revision", "EVT1.0"},
         {"androidboot.bootloader", "burp-0.1-7521"},
         {"androidboot.hardware.sku", "mary"},
diff --git a/gatekeeperd/Android.bp b/gatekeeperd/Android.bp
index 49e8085..95e814b 100644
--- a/gatekeeperd/Android.bp
+++ b/gatekeeperd/Android.bp
@@ -40,8 +40,6 @@
         "libbase",
         "libutils",
         "libcrypto",
-        "libkeystore_aidl",
-        "libkeystore_binder",
         "libhidlbase",
         "android.hardware.gatekeeper@1.0",
         "libgatekeeper_aidl",
diff --git a/gatekeeperd/gatekeeperd.cpp b/gatekeeperd/gatekeeperd.cpp
index f9c0cdd..8792c83 100644
--- a/gatekeeperd/gatekeeperd.cpp
+++ b/gatekeeperd/gatekeeperd.cpp
@@ -29,13 +29,11 @@
 #include <android-base/properties.h>
 #include <android/binder_ibinder.h>
 #include <android/binder_manager.h>
-#include <android/security/keystore/IKeystoreService.h>
 #include <binder/IPCThreadState.h>
 #include <binder/IServiceManager.h>
 #include <binder/PermissionCache.h>
 #include <gatekeeper/password_handle.h>  // for password_handle_t
 #include <hardware/hw_auth_token.h>
-#include <keystore/keystore_return_types.h>
 #include <libgsi/libgsi.h>
 #include <log/log.h>
 #include <utils/String16.h>
@@ -303,7 +301,7 @@
             if (gkResponse->payload().size() != 0) {
                 // try to connect to IKeystoreAuthorization AIDL service first.
                 AIBinder* authzAIBinder =
-                        AServiceManager_checkService("android.security.authorization");
+                        AServiceManager_getService("android.security.authorization");
                 ::ndk::SpAIBinder authzBinder(authzAIBinder);
                 auto authzService = IKeystoreAuthorization::fromBinder(authzBinder);
                 if (authzService) {
@@ -328,21 +326,6 @@
                         LOG(ERROR) << "Failure in sending AuthToken to AuthorizationService.";
                         return GK_ERROR;
                     }
-                }
-                sp<IServiceManager> sm = defaultServiceManager();
-
-                sp<IBinder> binder = sm->getService(String16("android.security.keystore"));
-                sp<security::keystore::IKeystoreService> service =
-                        interface_cast<security::keystore::IKeystoreService>(binder);
-
-                if (service) {
-                    int result = 0;
-                    auto binder_result = service->addAuthToken(gkResponse->payload(), &result);
-                    if (!binder_result.isOk() ||
-                        !keystore::KeyStoreServiceReturnCode(result).isOk()) {
-                        LOG(ERROR) << "Failure sending auth token to KeyStore: " << result;
-                        return GK_ERROR;
-                    }
                 } else {
                     LOG(ERROR) << "Cannot deliver auth token. Unable to communicate with "
                                   "Keystore.";
diff --git a/init/first_stage_init.cpp b/init/first_stage_init.cpp
index b2ab550..5cde167 100644
--- a/init/first_stage_init.cpp
+++ b/init/first_stage_init.cpp
@@ -337,12 +337,21 @@
         SwitchRoot("/first_stage_ramdisk");
     }
 
+    std::string force_debuggable("/force_debuggable");
+    std::string adb_debug_prop("/adb_debug.prop");
+    std::string userdebug_sepolicy("/userdebug_plat_sepolicy.cil");
+    if (IsRecoveryMode()) {
+        // Update these file paths since we didn't switch root
+        force_debuggable.insert(0, "/first_stage_ramdisk");
+        adb_debug_prop.insert(0, "/first_stage_ramdisk");
+        userdebug_sepolicy.insert(0, "/first_stage_ramdisk");
+    }
     // If this file is present, the second-stage init will use a userdebug sepolicy
     // and load adb_debug.prop to allow adb root, if the device is unlocked.
-    if (access("/force_debuggable", F_OK) == 0) {
+    if (access(force_debuggable.c_str(), F_OK) == 0) {
         std::error_code ec;  // to invoke the overloaded copy_file() that won't throw.
-        if (!fs::copy_file("/adb_debug.prop", kDebugRamdiskProp, ec) ||
-            !fs::copy_file("/userdebug_plat_sepolicy.cil", kDebugRamdiskSEPolicy, ec)) {
+        if (!fs::copy_file(adb_debug_prop, kDebugRamdiskProp, ec) ||
+            !fs::copy_file(userdebug_sepolicy, kDebugRamdiskSEPolicy, ec)) {
             LOG(ERROR) << "Failed to setup debug ramdisk";
         } else {
             // setenv for second-stage init to read above kDebugRamdisk* files.
diff --git a/init/init.cpp b/init/init.cpp
index 70d6809..7264b22 100644
--- a/init/init.cpp
+++ b/init/init.cpp
@@ -518,11 +518,9 @@
     if (!android::base::GetBoolProperty("ro.oem_unlock_supported", false)) {
         return;
     }
-    ImportKernelCmdline([](const std::string& key, const std::string& value) {
-        if (key == "androidboot.verifiedbootstate") {
-            SetProperty("ro.boot.flash.locked", value == "orange" ? "0" : "1");
-        }
-    });
+    SetProperty(
+            "ro.boot.flash.locked",
+            android::base::GetProperty("ro.boot.verifiedbootstate", "") == "orange" ? "0" : "1");
 }
 
 static Result<void> property_enable_triggers_action(const BuiltinArguments& args) {
diff --git a/init/property_service.cpp b/init/property_service.cpp
index b722702..73ef97a 100644
--- a/init/property_service.cpp
+++ b/init/property_service.cpp
@@ -44,6 +44,7 @@
 #include <mutex>
 #include <optional>
 #include <queue>
+#include <string_view>
 #include <thread>
 #include <vector>
 
@@ -1162,28 +1163,55 @@
     }
 }
 
+constexpr auto ANDROIDBOOT_PREFIX = "androidboot."sv;
+
+// emulator specific, should be removed once emulator is migrated to
+// bootconfig, see b/182291166.
+static std::string RemapEmulatorPropertyName(const std::string_view qemu_key) {
+    if (StartsWith(qemu_key, "dalvik."sv) || StartsWith(qemu_key, "opengles."sv) ||
+        StartsWith(qemu_key, "config."sv)) {
+        return std::string(qemu_key);
+    } else if (qemu_key == "uirenderer"sv) {
+        return "debug.hwui.renderer"s;
+    } else if (qemu_key == "media.ccodec"sv) {
+        return "debug.stagefright.ccodec"s;
+    } else {
+        return ""s;  // TBD
+    }
+}
+
 static void ProcessKernelCmdline() {
-    bool for_emulator = false;
     ImportKernelCmdline([&](const std::string& key, const std::string& value) {
-        if (key == "qemu") {
-            for_emulator = true;
-        } else if (StartsWith(key, "androidboot.")) {
-            InitPropertySet("ro.boot." + key.substr(12), value);
+        constexpr auto qemu_prefix = "qemu."sv;
+
+        if (StartsWith(key, ANDROIDBOOT_PREFIX)) {
+            InitPropertySet("ro.boot." + key.substr(ANDROIDBOOT_PREFIX.size()), value);
+        } else if (StartsWith(key, qemu_prefix)) {
+            InitPropertySet("ro.kernel." + key, value);  // emulator specific, deprecated
+
+            // emulator specific, should be retired once emulator migrates to
+            // androidboot.
+            const auto new_name =
+                    RemapEmulatorPropertyName(std::string_view(key).substr(qemu_prefix.size()));
+            if (!new_name.empty()) {
+                InitPropertySet("ro.boot." + new_name, value);
+            }
+        } else if (key == "qemu") {
+            // emulator specific, should be retired once emulator migrates to
+            // androidboot.
+            InitPropertySet("ro.boot." + key, value);
         }
     });
-
-    if (for_emulator) {
-        ImportKernelCmdline([&](const std::string& key, const std::string& value) {
-            // In the emulator, export any kernel option with the "ro.kernel." prefix.
-            InitPropertySet("ro.kernel." + key, value);
-        });
-    }
 }
 
 static void ProcessBootconfig() {
     ImportBootconfig([&](const std::string& key, const std::string& value) {
-        if (StartsWith(key, "androidboot.")) {
-            InitPropertySet("ro.boot." + key.substr(12), value);
+        if (StartsWith(key, ANDROIDBOOT_PREFIX)) {
+            InitPropertySet("ro.boot." + key.substr(ANDROIDBOOT_PREFIX.size()), value);
+        } else if (key == "hardware") {
+            // "hardware" in bootconfig replaces "androidboot.hardware" kernel
+            // cmdline parameter
+            InitPropertySet("ro.boot." + key, value);
         }
     });
 }
diff --git a/init/selinux.cpp b/init/selinux.cpp
index 0336936..62c4586 100644
--- a/init/selinux.cpp
+++ b/init/selinux.cpp
@@ -63,6 +63,7 @@
 #include <android-base/file.h>
 #include <android-base/logging.h>
 #include <android-base/parseint.h>
+#include <android-base/result.h>
 #include <android-base/strings.h>
 #include <android-base/unique_fd.h>
 #include <fs_avb/fs_avb.h>
@@ -92,7 +93,7 @@
 
 enum EnforcingStatus { SELINUX_PERMISSIVE, SELINUX_ENFORCING };
 
-EnforcingStatus StatusFromCmdline() {
+EnforcingStatus StatusFromProperty() {
     EnforcingStatus status = SELINUX_ENFORCING;
 
     ImportKernelCmdline([&](const std::string& key, const std::string& value) {
@@ -101,12 +102,20 @@
         }
     });
 
+    if (status == SELINUX_ENFORCING) {
+        ImportBootconfig([&](const std::string& key, const std::string& value) {
+            if (key == "androidboot.selinux" && value == "permissive") {
+                status = SELINUX_PERMISSIVE;
+            }
+        });
+    }
+
     return status;
 }
 
 bool IsEnforcing() {
     if (ALLOW_PERMISSIVE_SELINUX) {
-        return StatusFromCmdline() == SELINUX_ENFORCING;
+        return StatusFromProperty() == SELINUX_ENFORCING;
     }
     return true;
 }
@@ -214,8 +223,8 @@
     return true;
 }
 
-bool FindPrecompiledSplitPolicy(std::string* file) {
-    file->clear();
+Result<std::string> FindPrecompiledSplitPolicy() {
+    std::string precompiled_sepolicy;
     // If there is an odm partition, precompiled_sepolicy will be in
     // odm/etc/selinux. Otherwise it will be in vendor/etc/selinux.
     static constexpr const char vendor_precompiled_sepolicy[] =
@@ -223,62 +232,49 @@
     static constexpr const char odm_precompiled_sepolicy[] =
         "/odm/etc/selinux/precompiled_sepolicy";
     if (access(odm_precompiled_sepolicy, R_OK) == 0) {
-        *file = odm_precompiled_sepolicy;
+        precompiled_sepolicy = odm_precompiled_sepolicy;
     } else if (access(vendor_precompiled_sepolicy, R_OK) == 0) {
-        *file = vendor_precompiled_sepolicy;
+        precompiled_sepolicy = vendor_precompiled_sepolicy;
     } else {
-        PLOG(INFO) << "No precompiled sepolicy";
-        return false;
-    }
-    std::string actual_plat_id;
-    if (!ReadFirstLine("/system/etc/selinux/plat_sepolicy_and_mapping.sha256", &actual_plat_id)) {
-        PLOG(INFO) << "Failed to read "
-                      "/system/etc/selinux/plat_sepolicy_and_mapping.sha256";
-        return false;
-    }
-    std::string actual_system_ext_id;
-    if (!ReadFirstLine("/system_ext/etc/selinux/system_ext_sepolicy_and_mapping.sha256",
-                       &actual_system_ext_id)) {
-        PLOG(INFO) << "Failed to read "
-                      "/system_ext/etc/selinux/system_ext_sepolicy_and_mapping.sha256";
-        return false;
-    }
-    std::string actual_product_id;
-    if (!ReadFirstLine("/product/etc/selinux/product_sepolicy_and_mapping.sha256",
-                       &actual_product_id)) {
-        PLOG(INFO) << "Failed to read "
-                      "/product/etc/selinux/product_sepolicy_and_mapping.sha256";
-        return false;
+        return ErrnoError() << "No precompiled sepolicy at " << vendor_precompiled_sepolicy;
     }
 
-    std::string precompiled_plat_id;
-    std::string precompiled_plat_sha256 = *file + ".plat_sepolicy_and_mapping.sha256";
-    if (!ReadFirstLine(precompiled_plat_sha256.c_str(), &precompiled_plat_id)) {
-        PLOG(INFO) << "Failed to read " << precompiled_plat_sha256;
-        file->clear();
-        return false;
+    // Use precompiled sepolicy only when all corresponding hashes are equal.
+    // plat_sepolicy is always checked, while system_ext and product are checked only when they
+    // exist.
+    std::vector<std::pair<std::string, std::string>> sepolicy_hashes{
+            {"/system/etc/selinux/plat_sepolicy_and_mapping.sha256",
+             precompiled_sepolicy + ".plat_sepolicy_and_mapping.sha256"},
+    };
+
+    if (access("/system_ext/etc/selinux/system_ext_sepolicy.cil", F_OK) == 0) {
+        sepolicy_hashes.emplace_back(
+                "/system_ext/etc/selinux/system_ext_sepolicy_and_mapping.sha256",
+                precompiled_sepolicy + ".system_ext_sepolicy_and_mapping.sha256");
     }
-    std::string precompiled_system_ext_id;
-    std::string precompiled_system_ext_sha256 = *file + ".system_ext_sepolicy_and_mapping.sha256";
-    if (!ReadFirstLine(precompiled_system_ext_sha256.c_str(), &precompiled_system_ext_id)) {
-        PLOG(INFO) << "Failed to read " << precompiled_system_ext_sha256;
-        file->clear();
-        return false;
+
+    if (access("/product/etc/selinux/product_sepolicy.cil", F_OK) == 0) {
+        sepolicy_hashes.emplace_back("/product/etc/selinux/product_sepolicy_and_mapping.sha256",
+                                     precompiled_sepolicy + ".product_sepolicy_and_mapping.sha256");
     }
-    std::string precompiled_product_id;
-    std::string precompiled_product_sha256 = *file + ".product_sepolicy_and_mapping.sha256";
-    if (!ReadFirstLine(precompiled_product_sha256.c_str(), &precompiled_product_id)) {
-        PLOG(INFO) << "Failed to read " << precompiled_product_sha256;
-        file->clear();
-        return false;
+
+    for (const auto& [actual_id_path, precompiled_id_path] : sepolicy_hashes) {
+        std::string actual_id;
+        if (!ReadFirstLine(actual_id_path.c_str(), &actual_id)) {
+            return ErrnoError() << "Failed to read " << actual_id_path;
+        }
+
+        std::string precompiled_id;
+        if (!ReadFirstLine(precompiled_id_path.c_str(), &precompiled_id)) {
+            return ErrnoError() << "Failed to read " << precompiled_id_path;
+        }
+
+        if (actual_id.empty() || actual_id != precompiled_id) {
+            return Error() << actual_id_path << " and " << precompiled_id_path << " differ";
+        }
     }
-    if (actual_plat_id.empty() || actual_plat_id != precompiled_plat_id ||
-        actual_system_ext_id.empty() || actual_system_ext_id != precompiled_system_ext_id ||
-        actual_product_id.empty() || actual_product_id != precompiled_product_id) {
-        file->clear();
-        return false;
-    }
-    return true;
+
+    return precompiled_sepolicy;
 }
 
 bool GetVendorMappingVersion(std::string* plat_vers) {
@@ -325,15 +321,18 @@
 
     // Load precompiled policy from vendor image, if a matching policy is found there. The policy
     // must match the platform policy on the system image.
-    std::string precompiled_sepolicy_file;
     // use_userdebug_policy requires compiling sepolicy with userdebug_plat_sepolicy.cil.
     // Thus it cannot use the precompiled policy from vendor image.
-    if (!use_userdebug_policy && FindPrecompiledSplitPolicy(&precompiled_sepolicy_file)) {
-        unique_fd fd(open(precompiled_sepolicy_file.c_str(), O_RDONLY | O_CLOEXEC | O_BINARY));
-        if (fd != -1) {
-            policy_file->fd = std::move(fd);
-            policy_file->path = std::move(precompiled_sepolicy_file);
-            return true;
+    if (!use_userdebug_policy) {
+        if (auto res = FindPrecompiledSplitPolicy(); res.ok()) {
+            unique_fd fd(open(res->c_str(), O_RDONLY | O_CLOEXEC | O_BINARY));
+            if (fd != -1) {
+                policy_file->fd = std::move(fd);
+                policy_file->path = std::move(*res);
+                return true;
+            }
+        } else {
+            LOG(INFO) << res.error();
         }
     }
     // No suitable precompiled policy could be loaded
diff --git a/init/util.cpp b/init/util.cpp
index eab99d4..a40d104 100644
--- a/init/util.cpp
+++ b/init/util.cpp
@@ -376,6 +376,15 @@
             android_dt_dir = value;
         }
     });
+    // ..Or bootconfig
+    if (android_dt_dir == kDefaultAndroidDtDir) {
+        ImportBootconfig([&](const std::string& key, const std::string& value) {
+            if (key == "androidboot.android_dt_dir") {
+                android_dt_dir = value;
+            }
+        });
+    }
+
     LOG(INFO) << "Using Android DT directory " << android_dt_dir;
     return android_dt_dir;
 }
diff --git a/libcutils/Android.bp b/libcutils/Android.bp
index b38818a..0d9f2c7 100644
--- a/libcutils/Android.bp
+++ b/libcutils/Android.bp
@@ -170,14 +170,10 @@
         linux_bionic: {
             enabled: true,
         },
-        linux: {
-            srcs: [
-                "fs_config.cpp",
-            ],
-        },
         not_windows: {
             srcs: libcutils_nonwindows_sources + [
                 "ashmem-host.cpp",
+                "fs_config.cpp",
                 "trace-host.cpp",
             ],
         },
@@ -197,6 +193,7 @@
             srcs: libcutils_nonwindows_sources + [
                 "android_reboot.cpp",
                 "ashmem-dev.cpp",
+                "fs_config.cpp",
                 "klog.cpp",
                 "partition_utils.cpp",
                 "qtaguid.cpp",
diff --git a/libcutils/fs_config.cpp b/libcutils/fs_config.cpp
index 54eeeac..e9497a8 100644
--- a/libcutils/fs_config.cpp
+++ b/libcutils/fs_config.cpp
@@ -35,6 +35,7 @@
 #include <string>
 
 #include <android-base/strings.h>
+#include <cutils/fs.h>
 #include <log/log.h>
 #include <private/android_filesystem_config.h>
 
@@ -85,7 +86,7 @@
     { 00751, AID_ROOT,         AID_SHELL,        0, "system/bin" },
     { 00755, AID_ROOT,         AID_ROOT,         0, "system/etc/ppp" },
     { 00755, AID_ROOT,         AID_SHELL,        0, "system/vendor" },
-    { 00751, AID_ROOT,         AID_SHELL,        0, "system/xbin" },
+    { 00750, AID_ROOT,         AID_SHELL,        0, "system/xbin" },
     { 00751, AID_ROOT,         AID_SHELL,        0, "system/apex/*/bin" },
     { 00751, AID_ROOT,         AID_SHELL,        0, "system_ext/bin" },
     { 00751, AID_ROOT,         AID_SHELL,        0, "system_ext/apex/*/bin" },
diff --git a/libkeyutils/Android.bp b/libkeyutils/Android.bp
index 9848cd8..86f68fb 100644
--- a/libkeyutils/Android.bp
+++ b/libkeyutils/Android.bp
@@ -2,25 +2,10 @@
     default_applicable_licenses: ["system_core_libkeyutils_license"],
 }
 
-// Added automatically by a large-scale-change that took the approach of
-// 'apply every license found to every target'. While this makes sure we respect
-// every license restriction, it may not be entirely correct.
-//
-// e.g. GPL in an MIT project might only apply to the contrib/ directory.
-//
-// Please consider splitting the single license below into multiple licenses,
-// taking care not to lose any license_kind information, and overriding the
-// default license using the 'licenses: [...]' property on targets as needed.
-//
-// For unused files, consider creating a 'fileGroup' with "//visibility:private"
-// to attach the license to, and including a comment whether the files may be
-// used in the current project.
-// See: http://go/android-license-faq
 license {
     name: "system_core_libkeyutils_license",
     visibility: [":__subpackages__"],
     license_kinds: [
-        "SPDX-license-identifier-Apache-2.0",
         "SPDX-license-identifier-BSD",
     ],
     // large-scale-change unable to identify any license_text files
diff --git a/libprocessgroup/cgrouprc/include/android/cgrouprc.h b/libprocessgroup/cgrouprc/include/android/cgrouprc.h
index 9a79954..100d60e 100644
--- a/libprocessgroup/cgrouprc/include/android/cgrouprc.h
+++ b/libprocessgroup/cgrouprc/include/android/cgrouprc.h
@@ -68,6 +68,7 @@
  */
 #define CGROUPRC_CONTROLLER_FLAG_MOUNTED 0x1
 #define CGROUPRC_CONTROLLER_FLAG_NEEDS_ACTIVATION 0x2
+#define CGROUPRC_CONTROLLER_FLAG_OPTIONAL 0x4
 
 /**
  * Returns the flags bitmask of the given controller.
diff --git a/libprocessgroup/profiles/cgroups.json b/libprocessgroup/profiles/cgroups.json
index 962d2ba..0634220 100644
--- a/libprocessgroup/profiles/cgroups.json
+++ b/libprocessgroup/profiles/cgroups.json
@@ -26,7 +26,8 @@
       "Path": "/dev/memcg",
       "Mode": "0700",
       "UID": "root",
-      "GID": "system"
+      "GID": "system",
+      "Optional": true
     }
   ],
   "Cgroups2": {
diff --git a/libprocessgroup/profiles/cgroups.proto b/libprocessgroup/profiles/cgroups.proto
index 13adcae..f2de345 100644
--- a/libprocessgroup/profiles/cgroups.proto
+++ b/libprocessgroup/profiles/cgroups.proto
@@ -24,7 +24,7 @@
     Cgroups2 cgroups2 = 2 [json_name = "Cgroups2"];
 }
 
-// Next: 7
+// Next: 8
 message Cgroup {
     string controller = 1 [json_name = "Controller"];
     string path = 2 [json_name = "Path"];
@@ -35,6 +35,7 @@
 // when a boolean is specified as false, so leave unspecified in that case
 // https://developers.google.com/protocol-buffers/docs/proto3#default
     bool needs_activation = 6 [json_name = "NeedsActivation"];
+    bool is_optional = 7 [json_name = "Optional"];
 }
 
 // Next: 6
diff --git a/libprocessgroup/profiles/cgroups_30.json b/libprocessgroup/profiles/cgroups_30.json
index 17d4929..80a074b 100644
--- a/libprocessgroup/profiles/cgroups_30.json
+++ b/libprocessgroup/profiles/cgroups_30.json
@@ -5,7 +5,8 @@
       "Path": "/dev/stune",
       "Mode": "0755",
       "UID": "system",
-      "GID": "system"
+      "GID": "system",
+      "Optional": true
     }
   ]
 }
diff --git a/libprocessgroup/setup/cgroup_map_write.cpp b/libprocessgroup/setup/cgroup_map_write.cpp
index aa41acb..3121d24 100644
--- a/libprocessgroup/setup/cgroup_map_write.cpp
+++ b/libprocessgroup/setup/cgroup_map_write.cpp
@@ -161,6 +161,10 @@
         controller_flags |= CGROUPRC_CONTROLLER_FLAG_NEEDS_ACTIVATION;
     }
 
+    if (cgroup["Optional"].isBool() && cgroup["Optional"].asBool()) {
+        controller_flags |= CGROUPRC_CONTROLLER_FLAG_OPTIONAL;
+    }
+
     CgroupDescriptor descriptor(
             cgroups_version, name, path, std::strtoul(cgroup["Mode"].asString().c_str(), 0, 8),
             cgroup["UID"].asString(), cgroup["GID"].asString(), controller_flags);
@@ -267,8 +271,6 @@
                                        descriptor.gid())) {
                 LOG(ERROR) << "Failed to create directory for " << controller->name() << " cgroup";
                 result = -1;
-            } else {
-                LOG(ERROR) << "restored ownership for " << controller->name() << " cgroup";
             }
         } else {
             if (!Mkdir(controller->path(), descriptor.mode(), descriptor.uid(), descriptor.gid())) {
@@ -310,8 +312,15 @@
     }
 
     if (result < 0) {
-        PLOG(ERROR) << "Failed to mount " << controller->name() << " cgroup";
-        return false;
+        bool optional = controller->flags() & CGROUPRC_CONTROLLER_FLAG_OPTIONAL;
+
+        if (optional && errno == EINVAL) {
+            // Optional controllers are allowed to fail to mount if kernel does not support them
+            LOG(INFO) << "Optional " << controller->name() << " cgroup controller is not mounted";
+        } else {
+            PLOG(ERROR) << "Failed to mount " << controller->name() << " cgroup";
+            return false;
+        }
     }
 
     return true;
diff --git a/libstats/OWNERS b/libstats/OWNERS
index 7855774..d391679 100644
--- a/libstats/OWNERS
+++ b/libstats/OWNERS
@@ -1,6 +1,7 @@
-joeo@google.com
+jeffreyhuang@google.com
+jtnguyen@google.com
 muhammadq@google.com
-ruchirr@google.com
+sharaienko@google.com
 singhtejinder@google.com
 tsaichristine@google.com
 yaochen@google.com
diff --git a/libstats/pull_lazy/Android.bp b/libstats/pull_lazy/Android.bp
new file mode 100644
index 0000000..65dce26
--- /dev/null
+++ b/libstats/pull_lazy/Android.bp
@@ -0,0 +1,48 @@
+// Lazy loading version of libstatspull that can be used by code
+// that is running before the statsd APEX is mounted and
+// libstatspull.so is available.
+package {
+    default_applicable_licenses: ["Android-Apache-2.0"],
+}
+
+cc_library_static {
+    name: "libstatspull_lazy",
+    header_libs: [
+        "libstatspull_headers",
+        "libstatssocket_headers",
+    ],
+    export_header_lib_headers: [
+        "libstatspull_headers",
+    ],
+    apex_available: ["//apex_available:platform"],
+    srcs: ["libstatspull_lazy.cpp"],
+}
+
+cc_test {
+    name: "libstatspull_lazy_test",
+    srcs: [
+        "tests/libstatspull_lazy_test.cpp",
+    ],
+    static_libs: [
+        "libstatspull_lazy",
+        "libstatssocket_lazy",
+    ],
+    shared_libs: ["liblog"],
+    cflags: [
+        "-Wall",
+        "-Werror",
+    ],
+    test_suites: ["device-tests", "mts-statsd"],
+    test_config: "libstatspull_lazy_test.xml",
+    // TODO(b/153588990): Remove when the build system properly separates.
+    // 32bit and 64bit architectures.
+    compile_multilib: "both",
+    multilib: {
+        lib64: {
+            suffix: "64",
+        },
+        lib32: {
+            suffix: "32",
+        },
+    },
+}
diff --git a/libstats/pull_lazy/TEST_MAPPING b/libstats/pull_lazy/TEST_MAPPING
new file mode 100644
index 0000000..89b8c2a
--- /dev/null
+++ b/libstats/pull_lazy/TEST_MAPPING
@@ -0,0 +1,7 @@
+{
+  "presubmit" : [
+    {
+      "name" : "libstatspull_lazy_test"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/libstats/pull_lazy/libstatspull_lazy.cpp b/libstats/pull_lazy/libstatspull_lazy.cpp
new file mode 100644
index 0000000..b11fcee
--- /dev/null
+++ b/libstats/pull_lazy/libstatspull_lazy.cpp
@@ -0,0 +1,190 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#include "libstatspull_lazy.h"
+
+#include <mutex>
+
+#include <dlfcn.h>
+#include <stdatomic.h>
+
+#include "log/log.h"
+
+#include "stats_pull_atom_callback.h"
+
+// This file provides a lazy interface to libstatspull.so to address early boot dependencies.
+// Specifically bootanimation, surfaceflinger, and lmkd run before the statsd APEX is loaded and
+// libstatspull.so is in the statsd APEX.
+
+// Method pointers to libstatspull methods are held in an array which simplifies checking
+// all pointers are initialized.
+enum MethodIndex {
+    // PullAtomMetadata APIs in stats_pull_atom_callback.h.
+    k_AStatsManager_PullAtomMetadata_obtain,
+    k_AStatsManager_PullAtomMetadata_release,
+    k_AStatsManager_PullAtomMetadata_setCoolDownMillis,
+    k_AStatsManager_PullAtomMetadata_getCoolDownMillis,
+    k_AStatsManager_PullAtomMetadata_setTimeoutMillis,
+    k_AStatsManager_PullAtomMetadata_getTimeoutMillis,
+    k_AStatsManager_PullAtomMetadata_setAdditiveFields,
+    k_AStatsManager_PullAtomMetadata_getNumAdditiveFields,
+    k_AStatsManager_PullAtomMetadata_getAdditiveFields,
+
+    // AStatsEventList APIs in stats_pull_atom_callback.h
+    k_AStatsEventList_addStatsEvent,
+
+    // PullAtomCallback APIs in stats_pull_atom_callback.h
+    k_AStatsManager_setPullAtomCallback,
+    k_AStatsManager_clearPullAtomCallback,
+
+    // Marker for count of methods
+    k_MethodCount
+};
+
+// Table of methods pointers in libstatspull APIs.
+static void* g_Methods[k_MethodCount];
+
+//
+// Libstatspull lazy loading.
+//
+
+static atomic_bool gPreventLibstatspullLoading = false;  // Allows tests to block loading.
+
+void PreventLibstatspullLazyLoadingForTests() {
+    gPreventLibstatspullLoading.store(true);
+}
+
+static void* LoadLibstatspull(int dlopen_flags) {
+    if (gPreventLibstatspullLoading.load()) {
+        return nullptr;
+    }
+    return dlopen("libstatspull.so", dlopen_flags);
+}
+
+//
+// Initialization and symbol binding.
+
+static void BindSymbol(void* handle, const char* name, enum MethodIndex index) {
+    void* symbol = dlsym(handle, name);
+    LOG_ALWAYS_FATAL_IF(symbol == nullptr, "Failed to find symbol '%s' in libstatspull.so: %s",
+                        name, dlerror());
+    g_Methods[index] = symbol;
+}
+
+static void InitializeOnce() {
+    void* handle = LoadLibstatspull(RTLD_NOW);
+    LOG_ALWAYS_FATAL_IF(handle == nullptr, "Failed to load libstatspull.so: %s", dlerror());
+
+#undef BIND_SYMBOL
+#define BIND_SYMBOL(name) BindSymbol(handle, #name, k_##name);
+    // PullAtomMetadata APIs in stats_pull_atom_callback.h.
+    BIND_SYMBOL(AStatsManager_PullAtomMetadata_obtain);
+    BIND_SYMBOL(AStatsManager_PullAtomMetadata_release);
+    BIND_SYMBOL(AStatsManager_PullAtomMetadata_setCoolDownMillis);
+    BIND_SYMBOL(AStatsManager_PullAtomMetadata_getCoolDownMillis);
+    BIND_SYMBOL(AStatsManager_PullAtomMetadata_setTimeoutMillis);
+    BIND_SYMBOL(AStatsManager_PullAtomMetadata_getTimeoutMillis);
+    BIND_SYMBOL(AStatsManager_PullAtomMetadata_setAdditiveFields);
+    BIND_SYMBOL(AStatsManager_PullAtomMetadata_getNumAdditiveFields);
+    BIND_SYMBOL(AStatsManager_PullAtomMetadata_getAdditiveFields);
+
+    // AStatsEventList APIs in stats_pull_atom_callback.h
+    BIND_SYMBOL(AStatsEventList_addStatsEvent);
+
+    // PullAtomCallback APIs in stats_pull_atom_callback.h
+    BIND_SYMBOL(AStatsManager_setPullAtomCallback);
+    BIND_SYMBOL(AStatsManager_clearPullAtomCallback);
+
+#undef BIND_SYMBOL
+
+    // Check every symbol is bound.
+    for (int i = 0; i < k_MethodCount; ++i) {
+        LOG_ALWAYS_FATAL_IF(g_Methods[i] == nullptr,
+                            "Uninitialized method in libstatspull_lazy at index: %d", i);
+    }
+}
+
+static void EnsureInitialized() {
+    static std::once_flag initialize_flag;
+    std::call_once(initialize_flag, InitializeOnce);
+}
+
+#define INVOKE_METHOD(name, args...)                            \
+    do {                                                        \
+        EnsureInitialized();                                    \
+        void* method = g_Methods[k_##name];                     \
+        return reinterpret_cast<decltype(&name)>(method)(args); \
+    } while (0)
+
+//
+// Forwarding for methods in stats_pull_atom_callback.h.
+//
+
+AStatsManager_PullAtomMetadata* AStatsManager_PullAtomMetadata_obtain() {
+    INVOKE_METHOD(AStatsManager_PullAtomMetadata_obtain);
+}
+
+void AStatsManager_PullAtomMetadata_release(AStatsManager_PullAtomMetadata* metadata) {
+    INVOKE_METHOD(AStatsManager_PullAtomMetadata_release, metadata);
+}
+
+void AStatsManager_PullAtomMetadata_setCoolDownMillis(AStatsManager_PullAtomMetadata* metadata,
+                                                      int64_t cool_down_millis) {
+    INVOKE_METHOD(AStatsManager_PullAtomMetadata_setCoolDownMillis, metadata, cool_down_millis);
+}
+
+int64_t AStatsManager_PullAtomMetadata_getCoolDownMillis(AStatsManager_PullAtomMetadata* metadata) {
+    INVOKE_METHOD(AStatsManager_PullAtomMetadata_getCoolDownMillis, metadata);
+}
+
+void AStatsManager_PullAtomMetadata_setTimeoutMillis(AStatsManager_PullAtomMetadata* metadata,
+                                                     int64_t timeout_millis) {
+    INVOKE_METHOD(AStatsManager_PullAtomMetadata_setTimeoutMillis, metadata, timeout_millis);
+}
+
+int64_t AStatsManager_PullAtomMetadata_getTimeoutMillis(AStatsManager_PullAtomMetadata* metadata) {
+    INVOKE_METHOD(AStatsManager_PullAtomMetadata_getTimeoutMillis, metadata);
+}
+
+void AStatsManager_PullAtomMetadata_setAdditiveFields(AStatsManager_PullAtomMetadata* metadata,
+                                                      int32_t* additive_fields,
+                                                      int32_t num_fields) {
+    INVOKE_METHOD(AStatsManager_PullAtomMetadata_setAdditiveFields, metadata, additive_fields,
+                  num_fields);
+}
+
+int32_t AStatsManager_PullAtomMetadata_getNumAdditiveFields(
+        AStatsManager_PullAtomMetadata* metadata) {
+    INVOKE_METHOD(AStatsManager_PullAtomMetadata_getNumAdditiveFields, metadata);
+}
+
+void AStatsManager_PullAtomMetadata_getAdditiveFields(AStatsManager_PullAtomMetadata* metadata,
+                                                      int32_t* fields) {
+    INVOKE_METHOD(AStatsManager_PullAtomMetadata_getAdditiveFields, metadata, fields);
+}
+
+AStatsEvent* AStatsEventList_addStatsEvent(AStatsEventList* pull_data) {
+    INVOKE_METHOD(AStatsEventList_addStatsEvent, pull_data);
+}
+
+void AStatsManager_setPullAtomCallback(int32_t atom_tag, AStatsManager_PullAtomMetadata* metadata,
+                                       AStatsManager_PullAtomCallback callback, void* cookie) {
+    INVOKE_METHOD(AStatsManager_setPullAtomCallback, atom_tag, metadata, callback, cookie);
+}
+
+void AStatsManager_clearPullAtomCallback(int32_t atom_tag) {
+    INVOKE_METHOD(AStatsManager_clearPullAtomCallback, atom_tag);
+}
diff --git a/libstats/pull_lazy/libstatspull_lazy.h b/libstats/pull_lazy/libstatspull_lazy.h
new file mode 100644
index 0000000..2edddc7
--- /dev/null
+++ b/libstats/pull_lazy/libstatspull_lazy.h
@@ -0,0 +1,19 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#pragma once
+
+extern "C" void PreventLibstatspullLazyLoadingForTests();
\ No newline at end of file
diff --git a/libstats/pull_lazy/libstatspull_lazy_test.xml b/libstats/pull_lazy/libstatspull_lazy_test.xml
new file mode 100644
index 0000000..1b619af
--- /dev/null
+++ b/libstats/pull_lazy/libstatspull_lazy_test.xml
@@ -0,0 +1,37 @@
+<?xml version="1.0" encoding="utf-8"?>
+<!-- Copyright (C) 2021 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.
+-->
+<configuration description="Runs libstatspull_lazy_test.">
+    <option name="test-suite-tag" value="apct" />
+    <option name="test-suite-tag" value="apct-native" />
+    <option name="test-suite-tag" value="mts" />
+
+    <target_preparer class="com.android.tradefed.targetprep.RootTargetPreparer"/>
+
+    <target_preparer class="com.android.compatibility.common.tradefed.targetprep.FilePusher">
+        <option name="cleanup" value="true" />
+        <option name="push" value="libstatspull_lazy_test->/data/local/tmp/libstatspull_lazy_test" />
+        <option name="append-bitness" value="true" />
+    </target_preparer>
+
+    <test class="com.android.tradefed.testtype.GTest" >
+        <option name="native-test-device-path" value="/data/local/tmp" />
+        <option name="module-name" value="libstatspull_lazy_test" />
+    </test>
+
+    <object type="module_controller" class="com.android.tradefed.testtype.suite.module.MainlineTestModuleController">
+        <option name="mainline-module-package-name" value="com.google.android.os.statsd" />
+    </object>
+</configuration>
\ No newline at end of file
diff --git a/libstats/pull_lazy/tests/libstatspull_lazy_test.cpp b/libstats/pull_lazy/tests/libstatspull_lazy_test.cpp
new file mode 100644
index 0000000..41f82d0
--- /dev/null
+++ b/libstats/pull_lazy/tests/libstatspull_lazy_test.cpp
@@ -0,0 +1,58 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#include "../libstatspull_lazy.h"
+
+#include <gtest/gtest.h>
+
+#include "stats_pull_atom_callback.h"
+//#include "stats_event.h"
+
+// The tests here are just for the case when libstatspull.so cannot be loaded by
+// libstatspull_lazy.
+class LibstatspullLazyTest : public ::testing::Test {
+  protected:
+    virtual void SetUp() {
+        ::testing::Test::SetUp();
+        PreventLibstatspullLazyLoadingForTests();
+    }
+};
+
+static const char* kLoadFailed = "Failed to load libstatspull.so";
+
+TEST_F(LibstatspullLazyTest, NoLibstatspullForPullAtomMetadata) {
+    AStatsManager_PullAtomMetadata* metadata = NULL;
+    EXPECT_DEATH(AStatsManager_PullAtomMetadata_obtain(), kLoadFailed);
+    EXPECT_DEATH(AStatsManager_PullAtomMetadata_release(metadata), kLoadFailed);
+    EXPECT_DEATH(AStatsManager_PullAtomMetadata_setCoolDownMillis(metadata, 0), kLoadFailed);
+    EXPECT_DEATH(AStatsManager_PullAtomMetadata_getCoolDownMillis(metadata), kLoadFailed);
+    EXPECT_DEATH(AStatsManager_PullAtomMetadata_setTimeoutMillis(metadata, 0), kLoadFailed);
+    EXPECT_DEATH(AStatsManager_PullAtomMetadata_getTimeoutMillis(metadata), kLoadFailed);
+    EXPECT_DEATH(AStatsManager_PullAtomMetadata_setAdditiveFields(metadata, NULL, 0), kLoadFailed);
+    EXPECT_DEATH(AStatsManager_PullAtomMetadata_getNumAdditiveFields(metadata), kLoadFailed);
+    EXPECT_DEATH(AStatsManager_PullAtomMetadata_getAdditiveFields(metadata, NULL), kLoadFailed);
+}
+
+TEST_F(LibstatspullLazyTest, NoLibstatspullForAStatsEventList) {
+    AStatsEventList* event_list = NULL;
+    EXPECT_DEATH(AStatsEventList_addStatsEvent(event_list), kLoadFailed);
+}
+
+TEST_F(LibstatspullLazyTest, NoLibstatspullForPullAtomCallback) {
+    AStatsManager_PullAtomCallback callback = NULL;
+    EXPECT_DEATH(AStatsManager_setPullAtomCallback(0, NULL, callback, NULL), kLoadFailed);
+    EXPECT_DEATH(AStatsManager_clearPullAtomCallback(0), kLoadFailed);
+}
\ No newline at end of file
diff --git a/libstats/socket_lazy/Android.bp b/libstats/socket_lazy/Android.bp
new file mode 100644
index 0000000..b2cd7b2
--- /dev/null
+++ b/libstats/socket_lazy/Android.bp
@@ -0,0 +1,44 @@
+// Lazy loading version of libstatssocket that can be used by code
+// that is running before the statsd APEX is mounted and
+// libstatssocket.so is available.
+package {
+    default_applicable_licenses: ["Android-Apache-2.0"],
+}
+
+cc_library_static {
+    name: "libstatssocket_lazy",
+    header_libs: [
+        "libstatssocket_headers",
+    ],
+    export_header_lib_headers: [
+        "libstatssocket_headers",
+    ],
+    apex_available: ["//apex_available:platform"],
+    srcs: ["libstatssocket_lazy.cpp"],
+}
+
+cc_test {
+    name: "libstatssocket_lazy_test",
+    srcs: [
+        "tests/libstatssocket_lazy_test.cpp",
+    ],
+    static_libs: ["libstatssocket_lazy"],
+    shared_libs: ["liblog"],
+    cflags: [
+        "-Wall",
+        "-Werror",
+    ],
+    test_suites: ["device-tests", "mts-statsd"],
+    test_config: "libstatssocket_lazy_test.xml",
+    // TODO(b/153588990): Remove when the build system properly separates.
+    // 32bit and 64bit architectures.
+    compile_multilib: "both",
+    multilib: {
+        lib64: {
+            suffix: "64",
+        },
+        lib32: {
+            suffix: "32",
+        },
+    },
+}
diff --git a/libstats/socket_lazy/TEST_MAPPING b/libstats/socket_lazy/TEST_MAPPING
new file mode 100644
index 0000000..13afc00
--- /dev/null
+++ b/libstats/socket_lazy/TEST_MAPPING
@@ -0,0 +1,7 @@
+{
+  "presubmit" : [
+    {
+      "name" : "libstatssocket_lazy_test"
+    }
+  ]
+}
\ No newline at end of file
diff --git a/libstats/socket_lazy/libstatssocket_lazy.cpp b/libstats/socket_lazy/libstatssocket_lazy.cpp
new file mode 100644
index 0000000..dd93eeb
--- /dev/null
+++ b/libstats/socket_lazy/libstatssocket_lazy.cpp
@@ -0,0 +1,201 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#include "libstatssocket_lazy.h"
+
+#include <mutex>
+
+#include <dlfcn.h>
+#include <stdatomic.h>
+
+#include "log/log.h"
+
+#include "stats_event.h"
+#include "stats_socket.h"
+
+// This file provides a lazy interface to libstatssocket.so to address early boot dependencies.
+// Specifically bootanimation, surfaceflinger, and lmkd run before the statsd APEX is loaded and
+// libstatssocket.so is in the statsd APEX.
+
+// Method pointers to libstatssocket methods are held in an array which simplifies checking
+// all pointers are initialized.
+enum MethodIndex {
+    // Stats Event APIs in stats_event.h.
+    k_AStatsEvent_obtain,
+    k_AStatsEvent_build,
+    k_AStatsEvent_write,
+    k_AStatsEvent_release,
+    k_AStatsEvent_setAtomId,
+    k_AStatsEvent_writeInt32,
+    k_AStatsEvent_writeInt64,
+    k_AStatsEvent_writeFloat,
+    k_AStatsEvent_writeBool,
+    k_AStatsEvent_writeByteArray,
+    k_AStatsEvent_writeString,
+    k_AStatsEvent_writeAttributionChain,
+    k_AStatsEvent_addBoolAnnotation,
+    k_AStatsEvent_addInt32Annotation,
+
+    // Stats Socket APIs in stats_socket.h.
+    k_AStatsSocket_close,
+
+    // Marker for count of methods
+    k_MethodCount
+};
+
+// Table of methods pointers in libstatssocket APIs.
+static void* g_Methods[k_MethodCount];
+
+//
+// Libstatssocket lazy loading.
+//
+
+static atomic_bool gPreventLibstatssocketLoading = false;  // Allows tests to block loading.
+
+void PreventLibstatssocketLazyLoadingForTests() {
+    gPreventLibstatssocketLoading.store(true);
+}
+
+static void* LoadLibstatssocket(int dlopen_flags) {
+    if (gPreventLibstatssocketLoading.load()) {
+        return nullptr;
+    }
+    return dlopen("libstatssocket.so", dlopen_flags);
+}
+
+//
+// Initialization and symbol binding.
+
+static void BindSymbol(void* handle, const char* name, enum MethodIndex index) {
+    void* symbol = dlsym(handle, name);
+    LOG_ALWAYS_FATAL_IF(symbol == nullptr, "Failed to find symbol '%s' in libstatssocket.so: %s",
+                        name, dlerror());
+    g_Methods[index] = symbol;
+}
+
+static void InitializeOnce() {
+    void* handle = LoadLibstatssocket(RTLD_NOW);
+    LOG_ALWAYS_FATAL_IF(handle == nullptr, "Failed to load libstatssocket.so: %s", dlerror());
+
+#undef BIND_SYMBOL
+#define BIND_SYMBOL(name) BindSymbol(handle, #name, k_##name);
+    // Methods in stats_event.h.
+    BIND_SYMBOL(AStatsEvent_obtain);
+    BIND_SYMBOL(AStatsEvent_build);
+    BIND_SYMBOL(AStatsEvent_write);
+    BIND_SYMBOL(AStatsEvent_release);
+    BIND_SYMBOL(AStatsEvent_setAtomId);
+    BIND_SYMBOL(AStatsEvent_writeInt32);
+    BIND_SYMBOL(AStatsEvent_writeInt64);
+    BIND_SYMBOL(AStatsEvent_writeFloat);
+    BIND_SYMBOL(AStatsEvent_writeBool);
+    BIND_SYMBOL(AStatsEvent_writeByteArray);
+    BIND_SYMBOL(AStatsEvent_writeString);
+    BIND_SYMBOL(AStatsEvent_writeAttributionChain);
+    BIND_SYMBOL(AStatsEvent_addBoolAnnotation);
+    BIND_SYMBOL(AStatsEvent_addInt32Annotation);
+
+    // Methods in stats_socket.h.
+    BIND_SYMBOL(AStatsSocket_close);
+#undef BIND_SYMBOL
+
+    // Check every symbol is bound.
+    for (int i = 0; i < k_MethodCount; ++i) {
+        LOG_ALWAYS_FATAL_IF(g_Methods[i] == nullptr,
+                            "Uninitialized method in libstatssocket_lazy at index: %d", i);
+    }
+}
+
+static void EnsureInitialized() {
+    static std::once_flag initialize_flag;
+    std::call_once(initialize_flag, InitializeOnce);
+}
+
+#define INVOKE_METHOD(name, args...)                            \
+    do {                                                        \
+        EnsureInitialized();                                    \
+        void* method = g_Methods[k_##name];                     \
+        return reinterpret_cast<decltype(&name)>(method)(args); \
+    } while (0)
+
+//
+// Forwarding for methods in stats_event.h.
+//
+
+AStatsEvent* AStatsEvent_obtain() {
+    INVOKE_METHOD(AStatsEvent_obtain);
+}
+
+void AStatsEvent_build(AStatsEvent* event) {
+    INVOKE_METHOD(AStatsEvent_build, event);
+}
+
+int AStatsEvent_write(AStatsEvent* event) {
+    INVOKE_METHOD(AStatsEvent_write, event);
+}
+
+void AStatsEvent_release(AStatsEvent* event) {
+    INVOKE_METHOD(AStatsEvent_release, event);
+}
+
+void AStatsEvent_setAtomId(AStatsEvent* event, uint32_t atomId) {
+    INVOKE_METHOD(AStatsEvent_setAtomId, event, atomId);
+}
+
+void AStatsEvent_writeInt32(AStatsEvent* event, int32_t value) {
+    INVOKE_METHOD(AStatsEvent_writeInt32, event, value);
+}
+
+void AStatsEvent_writeInt64(AStatsEvent* event, int64_t value) {
+    INVOKE_METHOD(AStatsEvent_writeInt64, event, value);
+}
+
+void AStatsEvent_writeFloat(AStatsEvent* event, float value) {
+    INVOKE_METHOD(AStatsEvent_writeFloat, event, value);
+}
+
+void AStatsEvent_writeBool(AStatsEvent* event, bool value) {
+    INVOKE_METHOD(AStatsEvent_writeBool, event, value);
+}
+
+void AStatsEvent_writeByteArray(AStatsEvent* event, const uint8_t* buf, size_t numBytes) {
+    INVOKE_METHOD(AStatsEvent_writeByteArray, event, buf, numBytes);
+}
+
+void AStatsEvent_writeString(AStatsEvent* event, const char* value) {
+    INVOKE_METHOD(AStatsEvent_writeString, event, value);
+}
+
+void AStatsEvent_writeAttributionChain(AStatsEvent* event, const uint32_t* uids,
+                                       const char* const* tags, uint8_t numNodes) {
+    INVOKE_METHOD(AStatsEvent_writeAttributionChain, event, uids, tags, numNodes);
+}
+
+void AStatsEvent_addBoolAnnotation(AStatsEvent* event, uint8_t annotationId, bool value) {
+    INVOKE_METHOD(AStatsEvent_addBoolAnnotation, event, annotationId, value);
+}
+
+void AStatsEvent_addInt32Annotation(AStatsEvent* event, uint8_t annotationId, int32_t value) {
+    INVOKE_METHOD(AStatsEvent_addInt32Annotation, event, annotationId, value);
+}
+
+//
+// Forwarding for methods in stats_socket.h.
+//
+
+void AStatsSocket_close() {
+    INVOKE_METHOD(AStatsSocket_close);
+}
\ No newline at end of file
diff --git a/libstats/socket_lazy/libstatssocket_lazy.h b/libstats/socket_lazy/libstatssocket_lazy.h
new file mode 100644
index 0000000..3ff87cb
--- /dev/null
+++ b/libstats/socket_lazy/libstatssocket_lazy.h
@@ -0,0 +1,19 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#pragma once
+
+extern "C" void PreventLibstatssocketLazyLoadingForTests();
\ No newline at end of file
diff --git a/libstats/socket_lazy/libstatssocket_lazy_test.xml b/libstats/socket_lazy/libstatssocket_lazy_test.xml
new file mode 100644
index 0000000..ca6339b
--- /dev/null
+++ b/libstats/socket_lazy/libstatssocket_lazy_test.xml
@@ -0,0 +1,37 @@
+<?xml version="1.0" encoding="utf-8"?>
+<!-- Copyright (C) 2021 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.
+-->
+<configuration description="Runs libstatssocket_lazy_test.">
+    <option name="test-suite-tag" value="apct" />
+    <option name="test-suite-tag" value="apct-native" />
+    <option name="test-suite-tag" value="mts" />
+
+    <target_preparer class="com.android.tradefed.targetprep.RootTargetPreparer"/>
+
+    <target_preparer class="com.android.compatibility.common.tradefed.targetprep.FilePusher">
+        <option name="cleanup" value="true" />
+        <option name="push" value="libstatssocket_lazy_test->/data/local/tmp/libstatssocket_lazy_test" />
+        <option name="append-bitness" value="true" />
+    </target_preparer>
+
+    <test class="com.android.tradefed.testtype.GTest" >
+        <option name="native-test-device-path" value="/data/local/tmp" />
+        <option name="module-name" value="libstatssocket_lazy_test" />
+    </test>
+
+    <object type="module_controller" class="com.android.tradefed.testtype.suite.module.MainlineTestModuleController">
+        <option name="mainline-module-package-name" value="com.google.android.os.statsd" />
+    </object>
+</configuration>
\ No newline at end of file
diff --git a/libstats/socket_lazy/tests/libstatssocket_lazy_test.cpp b/libstats/socket_lazy/tests/libstatssocket_lazy_test.cpp
new file mode 100644
index 0000000..fe13598
--- /dev/null
+++ b/libstats/socket_lazy/tests/libstatssocket_lazy_test.cpp
@@ -0,0 +1,58 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#include "../libstatssocket_lazy.h"
+
+#include <gtest/gtest.h>
+
+#include "stats_event.h"
+#include "stats_socket.h"
+
+// The tests here are just for the case when libstatssocket.so cannot be loaded by
+// libstatssocket_lazy.
+class LibstatssocketLazyTest : public ::testing::Test {
+  protected:
+    virtual void SetUp() {
+        ::testing::Test::SetUp();
+        PreventLibstatssocketLazyLoadingForTests();
+    }
+};
+
+static const char* kLoadFailed = "Failed to load libstatssocket.so";
+
+TEST_F(LibstatssocketLazyTest, NoLibstatssocketForStatsEvent) {
+    AStatsEvent* event = NULL;
+    EXPECT_DEATH(AStatsEvent_obtain(), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_build(event), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_write(event), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_release(event), kLoadFailed);
+
+    EXPECT_DEATH(AStatsEvent_setAtomId(event, 0), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_writeInt32(event, 0), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_writeInt64(event, 0), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_writeFloat(event, 0), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_writeBool(event, false), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_writeByteArray(event, NULL, 0), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_writeString(event, NULL), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_writeAttributionChain(event, NULL, NULL, 0), kLoadFailed);
+
+    EXPECT_DEATH(AStatsEvent_addBoolAnnotation(event, 0, false), kLoadFailed);
+    EXPECT_DEATH(AStatsEvent_addInt32Annotation(event, 0, 0), kLoadFailed);
+}
+
+TEST_F(LibstatssocketLazyTest, NoLibstatssocketForStatsSocket) {
+    EXPECT_DEATH(AStatsSocket_close(), kLoadFailed);
+}
\ No newline at end of file
diff --git a/llkd/README.md b/llkd/README.md
index 6f92f14..9bcf806 100644
--- a/llkd/README.md
+++ b/llkd/README.md
@@ -207,7 +207,7 @@
 
 The `llkd` does not monitor the specified subset of processes for live lock stack
 signatures. Default is process names
-`init,lmkd.llkd,llkd,keystore,ueventd,apexd,logd`. Prevents the sepolicy
+`init,lmkd.llkd,llkd,keystore,keystore2,ueventd,apexd,logd`. Prevents the sepolicy
 violation associated with processes that block `ptrace` (as these can't be
 checked). **Active only on userdebug and eng builds**. For details on build
 types, refer to [Building Android](/setup/build/building#choose-a-target).
diff --git a/llkd/include/llkd.h b/llkd/include/llkd.h
index 4b20a56..0822a3e 100644
--- a/llkd/include/llkd.h
+++ b/llkd/include/llkd.h
@@ -60,7 +60,7 @@
 #define LLK_IGNORELIST_UID_PROPERTY     "ro.llk.ignorelist.uid"
 #define LLK_IGNORELIST_UID_DEFAULT      ""
 #define LLK_IGNORELIST_STACK_PROPERTY   "ro.llk.ignorelist.process.stack"
-#define LLK_IGNORELIST_STACK_DEFAULT    "init,lmkd.llkd,llkd,keystore,ueventd,apexd"
+#define LLK_IGNORELIST_STACK_DEFAULT    "init,lmkd.llkd,llkd,keystore,keystore2,ueventd,apexd"
 /* clang-format on */
 
 __END_DECLS
diff --git a/llkd/libllkd.cpp b/llkd/libllkd.cpp
index 9f3e218..c4c58ee 100644
--- a/llkd/libllkd.cpp
+++ b/llkd/libllkd.cpp
@@ -115,8 +115,8 @@
 // list of uids, and uid names, to skip, default nothing
 std::unordered_set<std::string> llkIgnorelistUid;
 #ifdef __PTRACE_ENABLED__
-// list of names to skip stack checking. "init", "lmkd", "llkd", "keystore" or
-// "logd" (if not userdebug).
+// list of names to skip stack checking. "init", "lmkd", "llkd", "keystore",
+// "keystore2", or "logd" (if not userdebug).
 std::unordered_set<std::string> llkIgnorelistStack;
 #endif
 
@@ -962,7 +962,8 @@
     //
     // This alarm is effectively the live lock detection of llkd, as
     // we understandably can not monitor ourselves otherwise.
-    ::alarm(duration_cast<seconds>(llkTimeoutMs * 2 * android::base::TimeoutMultiplier()).count());
+    ::alarm(duration_cast<seconds>(llkTimeoutMs * 2 * android::base::HwTimeoutMultiplier())
+                    .count());
 
     // kernel jiffy precision fastest acquisition
     static timespec last;
diff --git a/libkeyutils/mini_keyctl/Android.bp b/mini_keyctl/Android.bp
similarity index 100%
rename from libkeyutils/mini_keyctl/Android.bp
rename to mini_keyctl/Android.bp
diff --git a/libkeyutils/mini_keyctl/mini_keyctl.cpp b/mini_keyctl/mini_keyctl.cpp
similarity index 100%
rename from libkeyutils/mini_keyctl/mini_keyctl.cpp
rename to mini_keyctl/mini_keyctl.cpp
diff --git a/libkeyutils/mini_keyctl/mini_keyctl_utils.cpp b/mini_keyctl/mini_keyctl_utils.cpp
similarity index 100%
rename from libkeyutils/mini_keyctl/mini_keyctl_utils.cpp
rename to mini_keyctl/mini_keyctl_utils.cpp
diff --git a/libkeyutils/mini_keyctl/mini_keyctl_utils.h b/mini_keyctl/mini_keyctl_utils.h
similarity index 100%
rename from libkeyutils/mini_keyctl/mini_keyctl_utils.h
rename to mini_keyctl/mini_keyctl_utils.h
diff --git a/rootdir/etc/linker.config.json b/rootdir/etc/linker.config.json
index 2faf608..83cb6ff 100644
--- a/rootdir/etc/linker.config.json
+++ b/rootdir/etc/linker.config.json
@@ -6,6 +6,7 @@
     "libnativebridge.so",
     "libnativehelper.so",
     "libnativeloader.so",
+    "libsigchain.so",
     "libandroidicu.so",
     "libicu.so",
     // TODO(b/122876336): Remove libpac.so once it's migrated to Webview
@@ -26,4 +27,4 @@
     "libadb_pairing_connection.so",
     "libadb_pairing_server.so"
   ]
-}
\ No newline at end of file
+}
diff --git a/rootdir/init.rc b/rootdir/init.rc
index 0e1e98b..a04ae9f 100644
--- a/rootdir/init.rc
+++ b/rootdir/init.rc
@@ -630,6 +630,9 @@
     write /sys/kernel/tracing/instances/bootreceiver/events/error_report/error_report_end/enable 1
 
 on post-fs-data
+    # Boot level 30 - at this point daemons like apexd and odsign run
+    setprop keystore.boot_level 30
+
     mark_post_data
 
     # Start checkpoint before we touch data
@@ -881,9 +884,12 @@
     wait_for_prop apexd.status activated
     perform_apex_config
 
-    # Export *CLASSPATH variables from /etc/classpath
-    # TODO(b/180105615): export from the generated file instead.
-    load_exports /etc/classpath
+    # Define and export *CLASSPATH variables
+    mkdir /data/system/environ 0700 system system
+    # Must start before 'odsign', as odsign depends on *CLASSPATH variables
+    exec_start derive_classpath
+    load_exports /data/system/environ/classpath
+    rm /data/system/environ/classpath
 
     # Special-case /data/media/obb per b/64566063
     mkdir /data/media 0770 media_rw media_rw encryption=None
@@ -899,7 +905,12 @@
 
     # Start the on-device signing daemon, and wait for it to finish, to ensure
     # ART artifacts are generated if needed.
-    exec_start odsign
+    # Must start after 'derive_classpath' to have *CLASSPATH variables set.
+    start odsign
+
+    # Before we can lock keys and proceed to the next boot stage, wait for
+    # odsign to be done with the key
+    wait_for_prop odsign.key.done 1
 
     # After apexes are mounted, tell keymaster early boot has ended, so it will
     # stop allowing use of early-boot keys
@@ -908,6 +919,8 @@
     # Lock the fs-verity keyring, so no more keys can be added
     exec -- /system/bin/fsverity_init --lock
 
+    setprop keystore.boot_level 40
+
     # Allow apexd to snapshot and restore device encrypted apex data in the case
     # of a rollback. This should be done immediately after DE_user data keys
     # are loaded. APEXes should not access this data until this has been
@@ -939,6 +952,7 @@
 # It is recommended to put unnecessary data/ initialization from post-fs-data
 # to start-zygote in device's init.rc to unblock zygote start.
 on zygote-start && property:ro.crypto.state=unencrypted
+    wait_for_prop odsign.verification.done 1
     # A/B update verifier that marks a successful boot.
     exec_start update_verifier_nonencrypted
     start statsd
@@ -947,6 +961,7 @@
     start zygote_secondary
 
 on zygote-start && property:ro.crypto.state=unsupported
+    wait_for_prop odsign.verification.done 1
     # A/B update verifier that marks a successful boot.
     exec_start update_verifier_nonencrypted
     start statsd
@@ -955,6 +970,7 @@
     start zygote_secondary
 
 on zygote-start && property:ro.crypto.state=encrypted && property:ro.crypto.type=file
+    wait_for_prop odsign.verification.done 1
     # A/B update verifier that marks a successful boot.
     exec_start update_verifier_nonencrypted
     start statsd
@@ -1069,7 +1085,7 @@
     chown root radio /proc/cmdline
 
     # Define default initial receive window size in segments.
-    setprop net.tcp.default_init_rwnd 60
+    setprop net.tcp_def_init_rwnd 60
 
     # Start standard binderized HAL daemons
     class_start hal
@@ -1129,6 +1145,11 @@
 on property:sys.sysctl.extra_free_kbytes=*
     write /proc/sys/vm/extra_free_kbytes ${sys.sysctl.extra_free_kbytes}
 
+# Allow users to drop caches
+on property:perf.drop_caches=3
+    write /proc/sys/vm/drop_caches 3
+    setprop perf.drop_caches 0
+
 # "tcp_default_init_rwnd" Is too long!
 on property:net.tcp_def_init_rwnd=*
     write /proc/sys/net/ipv4/tcp_default_init_rwnd ${net.tcp_def_init_rwnd}
diff --git a/trusty/fuzz/Android.bp b/trusty/fuzz/Android.bp
index d147767..5d0ff79 100644
--- a/trusty/fuzz/Android.bp
+++ b/trusty/fuzz/Android.bp
@@ -30,7 +30,6 @@
         "-Werror",
     ],
     fuzz_config: {
-        fuzz_on_haiku_device: false,
         fuzz_on_haiku_host: false,
     },
 }
diff --git a/trusty/fuzz/test/Android.bp b/trusty/fuzz/test/Android.bp
index 7d74913..e0bca55 100644
--- a/trusty/fuzz/test/Android.bp
+++ b/trusty/fuzz/test/Android.bp
@@ -24,5 +24,8 @@
         "-DTRUSTY_APP_PORT=\"com.android.trusty.sancov.test.srv\"",
         "-DTRUSTY_APP_UUID=\"77f68803-c514-43ba-bdce-3254531c3d24\"",
         "-DTRUSTY_APP_FILENAME=\"srv.syms.elf\"",
-    ]
+    ],
+    fuzz_config: {
+        fuzz_on_haiku_device: false,
+    },
 }
diff --git a/trusty/fuzz/tipc_fuzzer.cpp b/trusty/fuzz/tipc_fuzzer.cpp
index 24b0f98..3258944 100644
--- a/trusty/fuzz/tipc_fuzzer.cpp
+++ b/trusty/fuzz/tipc_fuzzer.cpp
@@ -51,13 +51,21 @@
         exit(-1);
     }
 
+    /* Make sure lazy-loaded TAs have started and connected to coverage service. */
+    TrustyApp ta(TIPC_DEV, TRUSTY_APP_PORT);
+    auto ret = ta.Connect();
+    if (!ret.ok()) {
+        std::cerr << ret.error() << std::endl;
+        exit(-1);
+    }
+
     record = std::make_unique<CoverageRecord>(TIPC_DEV, &module_uuid, TRUSTY_APP_FILENAME);
     if (!record) {
         std::cerr << "Failed to allocate coverage record" << std::endl;
         exit(-1);
     }
 
-    auto ret = record->Open();
+    ret = record->Open();
     if (!ret.ok()) {
         std::cerr << ret.error() << std::endl;
         exit(-1);
diff --git a/trusty/utils/acvp/Android.bp b/trusty/utils/acvp/Android.bp
new file mode 100644
index 0000000..b851e39
--- /dev/null
+++ b/trusty/utils/acvp/Android.bp
@@ -0,0 +1,40 @@
+// Copyright (C) 2021 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.
+
+package {
+    default_applicable_licenses: ["Android-Apache-2.0"],
+}
+
+cc_binary {
+    name: "trusty_acvp_modulewrapper",
+    vendor: true,
+
+    srcs: [
+        "trusty_modulewrapper.cpp",
+    ],
+    static_libs: [
+        "libacvp_modulewrapper",
+    ],
+    shared_libs: [
+        "libbase",
+        "libc",
+        "libdmabufheap",
+        "liblog",
+        "libtrusty",
+        "libssl",
+    ],
+    cflags: [
+        "-Wall",
+        "-Werror",
+    ],
+}
diff --git a/trusty/utils/acvp/acvp_ipc.h b/trusty/utils/acvp/acvp_ipc.h
new file mode 100644
index 0000000..8b48ae3
--- /dev/null
+++ b/trusty/utils/acvp/acvp_ipc.h
@@ -0,0 +1,77 @@
+/*
+ * Copyright (C) 2021 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.
+ */
+
+#pragma once
+
+#include <stdint.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+#define ACVP_PORT "com.android.trusty.acvp"
+
+/*
+ * Maximum number of arguments
+ */
+#define ACVP_MAX_NUM_ARGUMENTS 8
+
+/*
+ * Maximum length of an algorithm name
+ */
+#define ACVP_MAX_NAME_LENGTH 30
+
+/*
+ * Maximum length of an ACVP request message
+ */
+#define ACVP_MAX_MESSAGE_LENGTH sizeof(struct acvp_req)
+
+/*
+ * Minimum length of the shared memory buffer
+ *
+ * This must be at least as long as the longest reply from the ACVP service
+ * (currently the reply from getConfig()).
+ */
+#define ACVP_MIN_SHARED_MEMORY 16384
+
+/**
+ * acvp_req - Request for the Trusty ACVP app
+ * @num_args: Number of acvp_arg structures following this struct
+ * @buffer_size: Total size of shared memory buffer
+ * @lengths: Length of each argument in the shared memory buffer
+ *
+ * @num_args copies of the acvp_arg struct follow this structure.
+ */
+struct acvp_req {
+    uint32_t num_args;
+    uint32_t buffer_size;
+    uint32_t lengths[ACVP_MAX_NUM_ARGUMENTS];
+};
+
+/**
+ * acvp_resp - Response to a ACVP request
+ *
+ * @num_spans: Number of response sections
+ * @lengths: Length of each response section
+ */
+struct acvp_resp {
+    uint32_t num_spans;
+    uint32_t lengths[ACVP_MAX_NUM_ARGUMENTS];
+};
+
+#ifdef __cplusplus
+}  // extern "C"
+#endif
diff --git a/trusty/utils/acvp/trusty_modulewrapper.cpp b/trusty/utils/acvp/trusty_modulewrapper.cpp
new file mode 100644
index 0000000..70ffb52
--- /dev/null
+++ b/trusty/utils/acvp/trusty_modulewrapper.cpp
@@ -0,0 +1,235 @@
+/*
+ * Copyright 2021, 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.
+ */
+
+#define LOG_TAG "TrustyAcvpModulewrapper"
+
+#include <BufferAllocator/BufferAllocator.h>
+#include <android-base/file.h>
+#include <android-base/result.h>
+#include <android-base/unique_fd.h>
+#include <errno.h>
+#include <log/log.h>
+#include <modulewrapper.h>
+#include <openssl/span.h>
+#include <stdint.h>
+#include <stdlib.h>
+#include <sys/mman.h>
+#include <trusty/tipc.h>
+#include <unistd.h>
+#include <iostream>
+
+#include "acvp_ipc.h"
+
+constexpr const char kTrustyDeviceName[] = "/dev/trusty-ipc-dev0";
+
+using android::base::ErrnoError;
+using android::base::Error;
+using android::base::Result;
+using android::base::unique_fd;
+using android::base::WriteFully;
+
+static inline size_t AlignUpToPage(size_t size) {
+    return (size + (PAGE_SIZE - 1)) & ~(PAGE_SIZE - 1);
+}
+
+namespace {
+
+class ModuleWrapper {
+  private:
+    static const char* kAcvpPort_;
+    static const char* kTrustyDeviceName_;
+
+  public:
+    ModuleWrapper();
+    ~ModuleWrapper();
+
+    Result<void> SendMessage(bssl::Span<const bssl::Span<const uint8_t>>);
+
+    Result<void> ForwardResponse();
+
+  private:
+    // Connection to the Trusty ACVP service
+    int tipc_fd_ = -1;
+
+    // Shared memory DMA buf
+    unique_fd dmabuf_fd_;
+
+    // Size of shared memory mapping
+    size_t shm_size_ = 0;
+
+    // Shared memory mapping
+    uint8_t* shm_buffer_ = nullptr;
+};
+
+}  // namespace
+
+const char* ModuleWrapper::kAcvpPort_ = ACVP_PORT;
+const char* ModuleWrapper::kTrustyDeviceName_ = kTrustyDeviceName;
+
+ModuleWrapper::ModuleWrapper() {
+    tipc_fd_ = tipc_connect(kTrustyDeviceName_, kAcvpPort_);
+    if (tipc_fd_ < 0) {
+        fprintf(stderr, "Failed to connect to Trusty ACVP test app: %s\n", strerror(-tipc_fd_));
+    }
+}
+
+ModuleWrapper::~ModuleWrapper() {
+    if (tipc_fd_ >= 0) {
+        tipc_close(tipc_fd_);
+    }
+
+    if (shm_buffer_) {
+        munmap(shm_buffer_, shm_size_);
+    }
+}
+
+Result<void> ModuleWrapper::SendMessage(bssl::Span<const bssl::Span<const uint8_t>> args) {
+    assert(args.size() < ACVP_MAX_NUM_ARGUMENTS);
+    assert(args[0].size() < ACVP_MAX_NAME_LENGTH);
+
+    struct acvp_req request;
+    request.num_args = args.size();
+
+    size_t total_args_size = 0;
+    for (auto arg : args) {
+        total_args_size += arg.size();
+    }
+
+    shm_size_ = ACVP_MIN_SHARED_MEMORY;
+    if (total_args_size > shm_size_) {
+        shm_size_ = AlignUpToPage(total_args_size);
+    }
+    request.buffer_size = shm_size_;
+
+    struct iovec iov = {
+            .iov_base = &request,
+            .iov_len = sizeof(struct acvp_req),
+    };
+
+    BufferAllocator alloc;
+    dmabuf_fd_.reset(alloc.Alloc(kDmabufSystemHeapName, shm_size_));
+    if (!dmabuf_fd_.ok()) {
+        return ErrnoError() << "Error creating dmabuf";
+    }
+
+    shm_buffer_ = (uint8_t*)mmap(0, shm_size_, PROT_READ | PROT_WRITE, MAP_SHARED, dmabuf_fd_, 0);
+    if (shm_buffer_ == MAP_FAILED) {
+        return ErrnoError() << "Failed to map shared memory dmabuf";
+    }
+
+    size_t cur_offset = 0;
+    for (int i = 0; i < args.size(); ++i) {
+        request.lengths[i] = args[i].size();
+        memcpy(shm_buffer_ + cur_offset, args[i].data(), args[i].size());
+        cur_offset += args[i].size();
+    }
+
+    struct trusty_shm shm = {
+            .fd = dmabuf_fd_.get(),
+            .transfer = TRUSTY_SHARE,
+    };
+
+    int rc = tipc_send(tipc_fd_, &iov, 1, &shm, 1);
+    if (rc != sizeof(struct acvp_req)) {
+        return ErrnoError() << "Failed to send request to Trusty ACVP service";
+    }
+
+    return {};
+}
+
+Result<void> ModuleWrapper::ForwardResponse() {
+    struct acvp_resp resp;
+    int bytes_read = read(tipc_fd_, &resp, sizeof(struct acvp_resp));
+    if (bytes_read < 0) {
+        return ErrnoError() << "Failed to read response from Trusty ACVP service";
+    }
+
+    if (bytes_read != sizeof(struct acvp_resp)) {
+        return Error() << "Trusty ACVP response overflowed expected size";
+    }
+
+    size_t total_args_size = 0;
+    for (size_t i = 0; i < resp.num_spans; i++) {
+        total_args_size += resp.lengths[i];
+    }
+
+    iovec iovs[2];
+    iovs[0].iov_base = &resp;
+    iovs[0].iov_len = sizeof(uint32_t) * (1 + resp.num_spans);
+
+    iovs[1].iov_base = shm_buffer_;
+    iovs[1].iov_len = total_args_size;
+
+    size_t iov_done = 0;
+    while (iov_done < 2) {
+        ssize_t r;
+        do {
+            r = writev(STDOUT_FILENO, &iovs[iov_done], 2 - iov_done);
+        } while (r == -1 && errno == EINTR);
+
+        if (r <= 0) {
+            return Error() << "Failed to write ACVP response to standard out";
+        }
+
+        size_t written = r;
+        for (size_t i = iov_done; i < 2 && written > 0; i++) {
+            iovec& iov = iovs[i];
+
+            size_t done = written;
+            if (done > iov.iov_len) {
+                done = iov.iov_len;
+            }
+
+            iov.iov_base = reinterpret_cast<uint8_t*>(iov.iov_base) + done;
+            iov.iov_len -= done;
+            written -= done;
+
+            if (iov.iov_len == 0) {
+                iov_done++;
+            }
+        }
+
+        assert(written == 0);
+    }
+
+    return {};
+}
+
+int main() {
+    for (;;) {
+        auto buffer = bssl::acvp::RequestBuffer::New();
+        auto args = bssl::acvp::ParseArgsFromFd(STDIN_FILENO, buffer.get());
+        if (args.empty()) {
+            ALOGE("Could not parse arguments\n");
+            return EXIT_FAILURE;
+        }
+
+        ModuleWrapper wrapper;
+        auto res = wrapper.SendMessage(args);
+        if (!res.ok()) {
+            std::cerr << res.error() << std::endl;
+            return EXIT_FAILURE;
+        }
+
+        res = wrapper.ForwardResponse();
+        if (!res.ok()) {
+            std::cerr << res.error() << std::endl;
+            return EXIT_FAILURE;
+        }
+    }
+
+    return EXIT_SUCCESS;
+};