Merge "Deprecate /factory"
diff --git a/debuggerd/Android.bp b/debuggerd/Android.bp
index 6391acc..fd62392 100644
--- a/debuggerd/Android.bp
+++ b/debuggerd/Android.bp
@@ -225,9 +225,6 @@
         debuggable: {
             cflags: ["-DROOT_POSSIBLE"],
         },
-        experimental_mte: {
-            cflags: ["-DANDROID_EXPERIMENTAL_MTE"],
-        },
     },
 }
 
@@ -297,12 +294,6 @@
     },
 
     test_suites: ["device-tests"],
-
-    product_variables: {
-        experimental_mte: {
-            cflags: ["-DANDROID_EXPERIMENTAL_MTE"],
-        },
-    },
 }
 
 cc_benchmark {
@@ -353,12 +344,6 @@
     apex_available: [
         "com.android.runtime",
     ],
-
-    product_variables: {
-        experimental_mte: {
-            cflags: ["-DANDROID_EXPERIMENTAL_MTE"],
-        },
-    },
 }
 
 cc_binary {
diff --git a/debuggerd/crash_dump.cpp b/debuggerd/crash_dump.cpp
index b3e81b0..4f60005 100644
--- a/debuggerd/crash_dump.cpp
+++ b/debuggerd/crash_dump.cpp
@@ -40,7 +40,6 @@
 #include <android-base/stringprintf.h>
 #include <android-base/strings.h>
 #include <android-base/unique_fd.h>
-#include <bionic/mte_kernel.h>
 #include <bionic/reserved_signals.h>
 #include <cutils/sockets.h>
 #include <log/log.h>
@@ -484,7 +483,6 @@
         continue;
       }
 
-#ifdef ANDROID_EXPERIMENTAL_MTE
       struct iovec iov = {
           &info.tagged_addr_ctrl,
           sizeof(info.tagged_addr_ctrl),
@@ -493,7 +491,6 @@
                  reinterpret_cast<void*>(&iov)) == -1) {
         info.tagged_addr_ctrl = -1;
       }
-#endif
 
       if (thread == g_target_thread) {
         // Read the thread's registers along with the rest of the crash info out of the pipe.
diff --git a/debuggerd/debuggerd_test.cpp b/debuggerd/debuggerd_test.cpp
index e5af425..7938a61 100644
--- a/debuggerd/debuggerd_test.cpp
+++ b/debuggerd/debuggerd_test.cpp
@@ -383,7 +383,7 @@
 #endif
 }
 
-#if defined(__aarch64__) && defined(ANDROID_EXPERIMENTAL_MTE)
+#if defined(__aarch64__)
 static void SetTagCheckingLevelSync() {
   HeapTaggingLevel heap_tagging_level = M_HEAP_TAGGING_LEVEL_SYNC;
   if (!android_mallopt(M_SET_HEAP_TAGGING_LEVEL, &heap_tagging_level, sizeof(heap_tagging_level))) {
@@ -393,7 +393,7 @@
 #endif
 
 TEST_F(CrasherTest, mte_uaf) {
-#if defined(__aarch64__) && defined(ANDROID_EXPERIMENTAL_MTE)
+#if defined(__aarch64__)
   if (!mte_supported()) {
     GTEST_SKIP() << "Requires MTE";
   }
@@ -425,12 +425,12 @@
   ASSERT_MATCH(result, R"(deallocated by thread .*
       #00 pc)");
 #else
-  GTEST_SKIP() << "Requires aarch64 + ANDROID_EXPERIMENTAL_MTE";
+  GTEST_SKIP() << "Requires aarch64";
 #endif
 }
 
 TEST_F(CrasherTest, mte_overflow) {
-#if defined(__aarch64__) && defined(ANDROID_EXPERIMENTAL_MTE)
+#if defined(__aarch64__)
   if (!mte_supported()) {
     GTEST_SKIP() << "Requires MTE";
   }
@@ -459,12 +459,12 @@
 allocated by thread .*
       #00 pc)");
 #else
-  GTEST_SKIP() << "Requires aarch64 + ANDROID_EXPERIMENTAL_MTE";
+  GTEST_SKIP() << "Requires aarch64";
 #endif
 }
 
 TEST_F(CrasherTest, mte_underflow) {
-#if defined(__aarch64__) && defined(ANDROID_EXPERIMENTAL_MTE)
+#if defined(__aarch64__)
   if (!mte_supported()) {
     GTEST_SKIP() << "Requires MTE";
   }
@@ -493,12 +493,12 @@
 allocated by thread .*
       #00 pc)");
 #else
-  GTEST_SKIP() << "Requires aarch64 + ANDROID_EXPERIMENTAL_MTE";
+  GTEST_SKIP() << "Requires aarch64";
 #endif
 }
 
 TEST_F(CrasherTest, mte_multiple_causes) {
-#if defined(__aarch64__) && defined(ANDROID_EXPERIMENTAL_MTE)
+#if defined(__aarch64__)
   if (!mte_supported()) {
     GTEST_SKIP() << "Requires MTE";
   }
@@ -547,11 +547,11 @@
   // overflows), so we can't match explicitly for an underflow message.
   ASSERT_MATCH(result, R"(Cause: \[MTE\]: Buffer Overflow, 0 bytes right of a 16-byte allocation)");
 #else
-  GTEST_SKIP() << "Requires aarch64 + ANDROID_EXPERIMENTAL_MTE";
+  GTEST_SKIP() << "Requires aarch64";
 #endif
 }
 
-#if defined(__aarch64__) && defined(ANDROID_EXPERIMENTAL_MTE)
+#if defined(__aarch64__)
 static uintptr_t CreateTagMapping() {
   uintptr_t mapping =
       reinterpret_cast<uintptr_t>(mmap(nullptr, getpagesize(), PROT_READ | PROT_WRITE | PROT_MTE,
@@ -568,7 +568,7 @@
 #endif
 
 TEST_F(CrasherTest, mte_tag_dump) {
-#if defined(__aarch64__) && defined(ANDROID_EXPERIMENTAL_MTE)
+#if defined(__aarch64__)
   if (!mte_supported()) {
     GTEST_SKIP() << "Requires MTE";
   }
@@ -596,7 +596,7 @@
     01.............0 0000000000000000 0000000000000000  ................
     00.............0)");
 #else
-  GTEST_SKIP() << "Requires aarch64 + ANDROID_EXPERIMENTAL_MTE";
+  GTEST_SKIP() << "Requires aarch64";
 #endif
 }
 
diff --git a/debuggerd/libdebuggerd/utility.cpp b/debuggerd/libdebuggerd/utility.cpp
index 2f1b693..7826efc 100644
--- a/debuggerd/libdebuggerd/utility.cpp
+++ b/debuggerd/libdebuggerd/utility.cpp
@@ -35,7 +35,6 @@
 #include <android-base/stringprintf.h>
 #include <android-base/strings.h>
 #include <android-base/unique_fd.h>
-#include <bionic/mte_kernel.h>
 #include <bionic/reserved_signals.h>
 #include <debuggerd/handler.h>
 #include <log/log.h>
diff --git a/fastboot/fs.cpp b/fastboot/fs.cpp
index 8addcb6..458a7a1 100644
--- a/fastboot/fs.cpp
+++ b/fastboot/fs.cpp
@@ -168,6 +168,19 @@
     return exec_cmd(e2fsdroid_args[0], e2fsdroid_args.data(), nullptr);
 }
 
+enum {
+    // clang-format off
+    FSCK_SUCCESS                 = 0,
+    FSCK_ERROR_CORRECTED         = 1 << 0,
+    FSCK_SYSTEM_SHOULD_REBOOT    = 1 << 1,
+    FSCK_ERRORS_LEFT_UNCORRECTED = 1 << 2,
+    FSCK_OPERATIONAL_ERROR       = 1 << 3,
+    FSCK_USAGE_OR_SYNTAX_ERROR   = 1 << 4,
+    FSCK_USER_CANCELLED          = 1 << 5,
+    FSCK_SHARED_LIB_ERROR        = 1 << 7,
+    // clang-format on
+};
+
 static int generate_f2fs_image(const char* fileName, long long partSize,
                                const std::string& initial_dir, unsigned /* unused */,
                                unsigned /* unused */, const unsigned fsOptions) {
@@ -216,7 +229,11 @@
     std::vector<const char*> sload_args = {sload_path.c_str(), "-S",
                                        "-f", initial_dir.c_str(), fileName, nullptr};
 
-    return exec_cmd(sload_args[0], sload_args.data(), nullptr);
+    ret = exec_cmd(sload_args[0], sload_args.data(), nullptr);
+    if (ret != 0 && ret != FSCK_ERROR_CORRECTED) {
+        return -1;
+    }
+    return 0;
 }
 
 static const struct fs_generator {
diff --git a/fs_mgr/README.overlayfs.md b/fs_mgr/README.overlayfs.md
index ca782b9..94b2f8c 100644
--- a/fs_mgr/README.overlayfs.md
+++ b/fs_mgr/README.overlayfs.md
@@ -8,8 +8,8 @@
 system partition as read-write and then add or modify any number of files
 without reflashing the system image, which is efficient for a development cycle.
 
-Limited memory systems use read-only types of file systems or logical resizable
-Android partitions (LRAPs). These file systems land system partition images
+Limited memory systems use read-only types of file systems or dynamic
+Android partitions (DAPs). These file systems land system partition images
 right-sized, and have been deduped at the block level to compress the content.
 This means that a remount either isn’t possible, or isn't useful because of
 space limitations or support logistics.
diff --git a/fs_mgr/fs_mgr.cpp b/fs_mgr/fs_mgr.cpp
index 4c9fd9b..2876094 100644
--- a/fs_mgr/fs_mgr.cpp
+++ b/fs_mgr/fs_mgr.cpp
@@ -1324,6 +1324,7 @@
     int error_count = 0;
     CheckpointManager checkpoint_manager;
     AvbUniquePtr avb_handle(nullptr);
+    bool wiped = false;
 
     bool userdata_mounted = false;
     if (fstab->empty()) {
@@ -1457,7 +1458,8 @@
                 encryptable = status;
                 if (status == FS_MGR_MNTALL_DEV_NEEDS_METADATA_ENCRYPTION) {
                     if (!call_vdc({"cryptfs", "encryptFstab", attempted_entry.blk_device,
-                                   attempted_entry.mount_point},
+                                   attempted_entry.mount_point, wiped ? "true" : "false",
+                                   attempted_entry.fs_type},
                                   nullptr)) {
                         LERROR << "Encryption failed";
                         set_type_property(encryptable);
@@ -1474,7 +1476,7 @@
         }
 
         // Mounting failed, understand why and retry.
-        bool wiped = partition_wiped(current_entry.blk_device.c_str());
+        wiped = partition_wiped(current_entry.blk_device.c_str());
         bool crypt_footer = false;
         if (mount_errno != EBUSY && mount_errno != EACCES &&
             current_entry.fs_mgr_flags.formattable && wiped) {
@@ -1499,6 +1501,27 @@
             } else if (current_entry.is_encryptable() && current_entry.key_loc == KEY_IN_FOOTER) {
                 crypt_footer = true;
             }
+
+            // EncryptInplace will be used when vdc gives an error or needs to format partitions
+            // other than /data
+            if (should_use_metadata_encryption(current_entry) &&
+                current_entry.mount_point == "/data") {
+
+                // vdc->Format requires "ro.crypto.type" to set an encryption flag
+                encryptable = FS_MGR_MNTALL_DEV_IS_METADATA_ENCRYPTED;
+                set_type_property(encryptable);
+
+                if (!call_vdc({"cryptfs", "encryptFstab", current_entry.blk_device,
+                               current_entry.mount_point, "true" /* shouldFormat */,
+                               current_entry.fs_type},
+                              nullptr)) {
+                    LERROR << "Encryption failed";
+                } else {
+                    userdata_mounted = true;
+                    continue;
+                }
+            }
+
             if (fs_mgr_do_format(current_entry, crypt_footer) == 0) {
                 // Let's replay the mount actions.
                 i = top_idx - 1;
diff --git a/fs_mgr/libsnapshot/Android.bp b/fs_mgr/libsnapshot/Android.bp
index f0adbea..d36a7f0 100644
--- a/fs_mgr/libsnapshot/Android.bp
+++ b/fs_mgr/libsnapshot/Android.bp
@@ -441,6 +441,7 @@
     system_shared_libs: [],
     ramdisk_available: true,
     vendor_ramdisk_available: true,
+    recovery_available: true,
 }
 
 cc_test {
diff --git a/fs_mgr/libsnapshot/cow_api_test.cpp b/fs_mgr/libsnapshot/cow_api_test.cpp
index 35020f4..defe8d4 100644
--- a/fs_mgr/libsnapshot/cow_api_test.cpp
+++ b/fs_mgr/libsnapshot/cow_api_test.cpp
@@ -171,6 +171,70 @@
     ASSERT_TRUE(iter->Done());
 }
 
+TEST_F(CowTest, ClusterCompressGz) {
+    CowOptions options;
+    options.compression = "gz";
+    options.cluster_ops = 2;
+    CowWriter writer(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()));
+
+    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);
+
+    CowReader reader;
+    ASSERT_TRUE(reader.Parse(cow_->fd));
+
+    auto iter = reader.GetOpIter();
+    ASSERT_NE(iter, nullptr);
+    ASSERT_FALSE(iter->Done());
+    auto op = &iter->Get();
+
+    StringSink sink;
+
+    ASSERT_EQ(op->type, kCowReplaceOp);
+    ASSERT_EQ(op->compression, kCowCompressGz);
+    ASSERT_EQ(op->data_length, 56);  // compressed!
+    ASSERT_EQ(op->new_block, 50);
+    ASSERT_TRUE(reader.ReadData(*op, &sink));
+    ASSERT_EQ(sink.stream(), data);
+
+    iter->Next();
+    ASSERT_FALSE(iter->Done());
+    op = &iter->Get();
+
+    ASSERT_EQ(op->type, kCowClusterOp);
+
+    iter->Next();
+    ASSERT_FALSE(iter->Done());
+    op = &iter->Get();
+
+    sink.Reset();
+    ASSERT_EQ(op->compression, kCowCompressGz);
+    ASSERT_EQ(op->data_length, 41);  // compressed!
+    ASSERT_EQ(op->new_block, 51);
+    ASSERT_TRUE(reader.ReadData(*op, &sink));
+    ASSERT_EQ(sink.stream(), data2);
+
+    iter->Next();
+    ASSERT_FALSE(iter->Done());
+    op = &iter->Get();
+
+    ASSERT_EQ(op->type, kCowClusterOp);
+
+    iter->Next();
+    ASSERT_TRUE(iter->Done());
+}
+
 TEST_F(CowTest, CompressTwoBlocks) {
     CowOptions options;
     options.compression = "gz";
diff --git a/fs_mgr/libsnapshot/cow_snapuserd_test.cpp b/fs_mgr/libsnapshot/cow_snapuserd_test.cpp
index ed67a1c..4cc0fd3 100644
--- a/fs_mgr/libsnapshot/cow_snapuserd_test.cpp
+++ b/fs_mgr/libsnapshot/cow_snapuserd_test.cpp
@@ -239,16 +239,12 @@
     ASSERT_TRUE(rnd_fd > 0);
 
     std::unique_ptr<uint8_t[]> random_buffer_1_ = std::make_unique<uint8_t[]>(size_);
-    std::unique_ptr<uint8_t[]> random_buffer_2_ = std::make_unique<uint8_t[]>(size_);
 
     // Fill random data
     for (size_t j = 0; j < (size_ / 1_MiB); j++) {
         ASSERT_EQ(ReadFullyAtOffset(rnd_fd, (char*)random_buffer_1_.get() + offset, 1_MiB, 0),
                   true);
 
-        ASSERT_EQ(ReadFullyAtOffset(rnd_fd, (char*)random_buffer_2_.get() + offset, 1_MiB, 0),
-                  true);
-
         offset += 1_MiB;
     }
 
@@ -280,7 +276,7 @@
 
     size_t blk_random2_replace_start = blk_zero_copy_end;
 
-    ASSERT_TRUE(writer.AddRawBlocks(blk_random2_replace_start, random_buffer_2_.get(), size_));
+    ASSERT_TRUE(writer.AddRawBlocks(blk_random2_replace_start, random_buffer_1_.get(), size_));
 
     // Flush operations
     ASSERT_TRUE(writer.Finalize());
@@ -292,7 +288,7 @@
     ASSERT_EQ(android::base::ReadFullyAtOffset(base_fd_, orig_buffer_.get(), size_, size_), true);
     memcpy((char*)orig_buffer_.get() + size_, random_buffer_1_.get(), size_);
     memcpy((char*)orig_buffer_.get() + (size_ * 2), (void*)zero_buffer.c_str(), size_);
-    memcpy((char*)orig_buffer_.get() + (size_ * 3), random_buffer_2_.get(), size_);
+    memcpy((char*)orig_buffer_.get() + (size_ * 3), random_buffer_1_.get(), size_);
 }
 
 void CowSnapuserdTest::InitCowDevice() {
diff --git a/fs_mgr/libsnapshot/cow_writer.cpp b/fs_mgr/libsnapshot/cow_writer.cpp
index 8535252..c1a5f32 100644
--- a/fs_mgr/libsnapshot/cow_writer.cpp
+++ b/fs_mgr/libsnapshot/cow_writer.cpp
@@ -433,11 +433,6 @@
 }
 
 bool CowWriter::WriteOperation(const CowOperation& op, const void* data, size_t size) {
-    // If there isn't room for this op and the cluster end op, end the current cluster
-    if (cluster_size_ && op.type != kCowClusterOp &&
-        cluster_size_ < current_cluster_size_ + 2 * sizeof(op)) {
-        if (!EmitCluster()) return false;
-    }
     if (lseek(fd_.get(), next_op_pos_, SEEK_SET) < 0) {
         PLOG(ERROR) << "lseek failed for writing operation.";
         return false;
@@ -449,6 +444,11 @@
         if (!WriteRawData(data, size)) return false;
     }
     AddOperation(op);
+    // If there isn't room for another op and the cluster end op, end the current cluster
+    if (cluster_size_ && op.type != kCowClusterOp &&
+        cluster_size_ < current_cluster_size_ + 2 * sizeof(op)) {
+        if (!EmitCluster()) return false;
+    }
     return true;
 }
 
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h b/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
index f8d3cdc..397ff2e 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
@@ -306,13 +306,17 @@
     // Helper function for second stage init to restorecon on the rollback indicator.
     static std::string GetGlobalRollbackIndicatorPath();
 
-    // Initiate the transition from first-stage to second-stage snapuserd. This
-    // process involves re-creating the dm-user table entries for each device,
-    // so that they connect to the new daemon. Once all new tables have been
-    // activated, we ask the first-stage daemon to cleanly exit.
-    //
-    // The caller must pass a function which starts snapuserd.
-    bool PerformSecondStageTransition();
+    // Detach dm-user devices from the current snapuserd, and populate
+    // |snapuserd_argv| with the necessary arguments to restart snapuserd
+    // and reattach them.
+    bool DetachSnapuserdForSelinux(std::vector<std::string>* snapuserd_argv);
+
+    // Perform the transition from the selinux stage of snapuserd into the
+    // second-stage of snapuserd. This process involves re-creating the dm-user
+    // table entries for each device, so that they connect to the new daemon.
+    // Once all new tables have been activated, we ask the first-stage daemon
+    // to cleanly exit.
+    bool PerformSecondStageInitTransition();
 
     // ISnapshotManager overrides.
     bool BeginUpdate() override;
@@ -626,6 +630,9 @@
     // The reverse of MapPartitionWithSnapshot.
     bool UnmapPartitionWithSnapshot(LockedFile* lock, const std::string& target_partition_name);
 
+    // Unmap a dm-user device through snapuserd.
+    bool UnmapDmUserDevice(const std::string& snapshot_name);
+
     // If there isn't a previous update, return true. |needs_merge| is set to false.
     // If there is a previous update but the device has not boot into it, tries to cancel the
     //   update and delete any snapshots. Return true if successful. |needs_merge| is set to false.
@@ -693,6 +700,19 @@
     // returns true.
     bool WaitForDevice(const std::string& device, std::chrono::milliseconds timeout_ms);
 
+    enum class InitTransition { SELINUX_DETACH, SECOND_STAGE };
+
+    // Initiate the transition from first-stage to second-stage snapuserd. This
+    // process involves re-creating the dm-user table entries for each device,
+    // so that they connect to the new daemon. Once all new tables have been
+    // activated, we ask the first-stage daemon to cleanly exit.
+    //
+    // If the mode is SELINUX_DETACH, snapuserd_argv must be non-null and will
+    // be populated with a list of snapuserd arguments to pass to execve(). It
+    // is otherwise ignored.
+    bool PerformInitTransition(InitTransition transition,
+                               std::vector<std::string>* snapuserd_argv = nullptr);
+
     std::string gsid_dir_;
     std::string metadata_dir_;
     std::unique_ptr<IDeviceInfo> device_;
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_client.h b/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_client.h
index aa9ba6e..1dab361 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_client.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_client.h
@@ -32,8 +32,6 @@
 
 static constexpr char kSnapuserdSocket[] = "snapuserd";
 
-static constexpr char kSnapuserdFirstStagePidVar[] = "FIRST_STAGE_SNAPUSERD_PID";
-
 // Ensure that the second-stage daemon for snapuserd is running.
 bool EnsureSnapuserdStarted();
 
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_kernel.h b/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_kernel.h
index e8dbe6e..7941e68 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_kernel.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapuserd_kernel.h
@@ -50,6 +50,8 @@
 static constexpr uint32_t BLOCK_SIZE = 4096;
 static constexpr uint32_t BLOCK_SHIFT = (__builtin_ffs(BLOCK_SIZE) - 1);
 
+#define DIV_ROUND_UP(n, d) (((n) + (d)-1) / (d))
+
 // This structure represents the kernel COW header.
 // All the below fields should be in Little Endian format.
 struct disk_header {
diff --git a/fs_mgr/libsnapshot/inspect_cow.cpp b/fs_mgr/libsnapshot/inspect_cow.cpp
index 5ad61f3..453b5c6 100644
--- a/fs_mgr/libsnapshot/inspect_cow.cpp
+++ b/fs_mgr/libsnapshot/inspect_cow.cpp
@@ -36,10 +36,30 @@
 }
 
 static void usage(void) {
-    LOG(ERROR) << "Usage: inspect_cow [-s] <COW_FILE>";
+    LOG(ERROR) << "Usage: inspect_cow [-sd] <COW_FILE>";
+    LOG(ERROR) << "\t -s Run Silent";
+    LOG(ERROR) << "\t -d Attempt to decompress\n";
 }
 
-static bool Inspect(const std::string& path, bool silent) {
+// Sink that always appends to the end of a string.
+class StringSink : public IByteSink {
+  public:
+    void* GetBuffer(size_t requested, size_t* actual) override {
+        size_t old_size = stream_.size();
+        stream_.resize(old_size + requested, '\0');
+        *actual = requested;
+        return stream_.data() + old_size;
+    }
+    bool ReturnData(void*, size_t) override { return true; }
+    void Reset() { stream_.clear(); }
+
+    std::string& stream() { return stream_; }
+
+  private:
+    std::string stream_;
+};
+
+static bool Inspect(const std::string& path, bool silent, bool decompress) {
     android::base::unique_fd fd(open(path.c_str(), O_RDONLY));
     if (fd < 0) {
         PLOG(ERROR) << "open failed: " << path;
@@ -76,15 +96,25 @@
     }
 
     auto iter = reader.GetOpIter();
+    StringSink sink;
+    bool success = true;
     while (!iter->Done()) {
         const CowOperation& op = iter->Get();
 
         if (!silent) std::cout << op << "\n";
 
+        if (decompress && op.type == kCowReplaceOp && op.compression != kCowCompressNone) {
+            if (!reader.ReadData(op, &sink)) {
+                std::cerr << "Failed to decompress for :" << op << "\n";
+                success = false;
+            }
+            sink.Reset();
+        }
+
         iter->Next();
     }
 
-    return true;
+    return success;
 }
 
 }  // namespace snapshot
@@ -93,11 +123,15 @@
 int main(int argc, char** argv) {
     int ch;
     bool silent = false;
-    while ((ch = getopt(argc, argv, "s")) != -1) {
+    bool decompress = false;
+    while ((ch = getopt(argc, argv, "sd")) != -1) {
         switch (ch) {
             case 's':
                 silent = true;
                 break;
+            case 'd':
+                decompress = true;
+                break;
             default:
                 android::snapshot::usage();
         }
@@ -109,7 +143,7 @@
         return 1;
     }
 
-    if (!android::snapshot::Inspect(argv[optind], silent)) {
+    if (!android::snapshot::Inspect(argv[optind], silent, decompress)) {
         return 1;
     }
     return 0;
diff --git a/fs_mgr/libsnapshot/snapshot.cpp b/fs_mgr/libsnapshot/snapshot.cpp
index 04c3ccc..f0646fc 100644
--- a/fs_mgr/libsnapshot/snapshot.cpp
+++ b/fs_mgr/libsnapshot/snapshot.cpp
@@ -1247,6 +1247,10 @@
         return false;
     }
 
+    if (status.compression_enabled()) {
+        UnmapDmUserDevice(name);
+    }
+
     // Cleanup the base device as well, since it is no longer used. This does
     // not block cleanup.
     auto base_name = GetBaseDeviceName(name);
@@ -1291,12 +1295,13 @@
     return RemoveAllUpdateState(lock, before_cancel);
 }
 
-bool SnapshotManager::PerformSecondStageTransition() {
-    LOG(INFO) << "Performing second-stage transition for snapuserd.";
+bool SnapshotManager::PerformInitTransition(InitTransition transition,
+                                            std::vector<std::string>* snapuserd_argv) {
+    LOG(INFO) << "Performing transition for snapuserd.";
 
     // Don't use EnsuerSnapuserdConnected() because this is called from init,
     // and attempting to do so will deadlock.
-    if (!snapuserd_client_) {
+    if (!snapuserd_client_ && transition != InitTransition::SELINUX_DETACH) {
         snapuserd_client_ = SnapuserdClient::Connect(kSnapuserdSocket, 10s);
         if (!snapuserd_client_) {
             LOG(ERROR) << "Unable to connect to snapuserd";
@@ -1343,6 +1348,9 @@
         }
 
         auto misc_name = user_cow_name;
+        if (transition == InitTransition::SELINUX_DETACH) {
+            misc_name += "-selinux";
+        }
 
         DmTable table;
         table.Emplace<DmTargetUser>(0, target.spec.length, misc_name);
@@ -1378,6 +1386,17 @@
             continue;
         }
 
+        if (transition == InitTransition::SELINUX_DETACH) {
+            auto message = misc_name + "," + cow_image_device + "," + backing_device;
+            snapuserd_argv->emplace_back(std::move(message));
+
+            // Do not attempt to connect to the new snapuserd yet, it hasn't
+            // been started. We do however want to wait for the misc device
+            // to have been created.
+            ok_cows++;
+            continue;
+        }
+
         uint64_t base_sectors =
                 snapuserd_client_->InitDmUserCow(misc_name, cow_image_device, backing_device);
         if (base_sectors == 0) {
@@ -2048,27 +2067,8 @@
 
     auto& dm = DeviceMapper::Instance();
 
-    auto dm_user_name = GetDmUserCowName(name);
-    if (IsCompressionEnabled() && dm.GetState(dm_user_name) != DmDeviceState::INVALID) {
-        if (!EnsureSnapuserdConnected()) {
-            return false;
-        }
-        if (!dm.DeleteDeviceIfExists(dm_user_name)) {
-            LOG(ERROR) << "Cannot unmap " << dm_user_name;
-            return false;
-        }
-
-        if (!snapuserd_client_->WaitForDeviceDelete(dm_user_name)) {
-            LOG(ERROR) << "Failed to wait for " << dm_user_name << " control device to delete";
-            return false;
-        }
-
-        // Ensure the control device is gone so we don't run into ABA problems.
-        auto control_device = "/dev/dm-user/" + dm_user_name;
-        if (!android::fs_mgr::WaitForFileDeleted(control_device, 10s)) {
-            LOG(ERROR) << "Timed out waiting for " << control_device << " to unlink";
-            return false;
-        }
+    if (IsCompressionEnabled() && !UnmapDmUserDevice(name)) {
+        return false;
     }
 
     auto cow_name = GetCowName(name);
@@ -2085,6 +2085,37 @@
     return true;
 }
 
+bool SnapshotManager::UnmapDmUserDevice(const std::string& snapshot_name) {
+    auto& dm = DeviceMapper::Instance();
+
+    if (!EnsureSnapuserdConnected()) {
+        return false;
+    }
+
+    auto dm_user_name = GetDmUserCowName(snapshot_name);
+    if (dm.GetState(dm_user_name) == DmDeviceState::INVALID) {
+        return true;
+    }
+
+    if (!dm.DeleteDeviceIfExists(dm_user_name)) {
+        LOG(ERROR) << "Cannot unmap " << dm_user_name;
+        return false;
+    }
+
+    if (!snapuserd_client_->WaitForDeviceDelete(dm_user_name)) {
+        LOG(ERROR) << "Failed to wait for " << dm_user_name << " control device to delete";
+        return false;
+    }
+
+    // Ensure the control device is gone so we don't run into ABA problems.
+    auto control_device = "/dev/dm-user/" + dm_user_name;
+    if (!android::fs_mgr::WaitForFileDeleted(control_device, 10s)) {
+        LOG(ERROR) << "Timed out waiting for " << control_device << " to unlink";
+        return false;
+    }
+    return true;
+}
+
 bool SnapshotManager::MapAllSnapshots(const std::chrono::milliseconds& timeout_ms) {
     auto lock = LockExclusive();
     if (!lock) return false;
@@ -3311,5 +3342,13 @@
     return status.state() != UpdateState::None && status.compression_enabled();
 }
 
+bool SnapshotManager::DetachSnapuserdForSelinux(std::vector<std::string>* snapuserd_argv) {
+    return PerformInitTransition(InitTransition::SELINUX_DETACH, snapuserd_argv);
+}
+
+bool SnapshotManager::PerformSecondStageInitTransition() {
+    return PerformInitTransition(InitTransition::SECOND_STAGE);
+}
+
 }  // namespace snapshot
 }  // namespace android
diff --git a/fs_mgr/libsnapshot/snapshot_test.cpp b/fs_mgr/libsnapshot/snapshot_test.cpp
index c6a1786..bb44425 100644
--- a/fs_mgr/libsnapshot/snapshot_test.cpp
+++ b/fs_mgr/libsnapshot/snapshot_test.cpp
@@ -146,6 +146,7 @@
                 "base-device",
                 "test_partition_b",
                 "test_partition_b-base",
+                "test_partition_b-base",
         };
         for (const auto& partition : partitions) {
             DeleteDevice(partition);
@@ -180,12 +181,22 @@
     }
 
     // If |path| is non-null, the partition will be mapped after creation.
-    bool CreatePartition(const std::string& name, uint64_t size, std::string* path = nullptr) {
+    bool CreatePartition(const std::string& name, uint64_t size, std::string* path = nullptr,
+                         const std::optional<std::string> group = {}) {
         TestPartitionOpener opener(fake_super);
         auto builder = MetadataBuilder::New(opener, "super", 0);
         if (!builder) return false;
 
-        auto partition = builder->AddPartition(name, 0);
+        std::string partition_group = std::string(android::fs_mgr::kDefaultGroup);
+        if (group) {
+            partition_group = *group;
+        }
+        return CreatePartition(builder.get(), name, size, path, partition_group);
+    }
+
+    bool CreatePartition(MetadataBuilder* builder, const std::string& name, uint64_t size,
+                         std::string* path, const std::string& group) {
+        auto partition = builder->AddPartition(name, group, 0);
         if (!partition) return false;
         if (!builder->ResizePartition(partition, size)) {
             return false;
@@ -194,6 +205,8 @@
         // Update the source slot.
         auto metadata = builder->Export();
         if (!metadata) return false;
+
+        TestPartitionOpener opener(fake_super);
         if (!UpdatePartitionTable(opener, "super", *metadata.get(), 0)) {
             return false;
         }
@@ -210,39 +223,54 @@
         return CreateLogicalPartition(params, path);
     }
 
-    bool MapUpdatePartitions() {
+    AssertionResult MapUpdateSnapshot(const std::string& name,
+                                      std::unique_ptr<ICowWriter>* writer) {
         TestPartitionOpener opener(fake_super);
-        auto builder = MetadataBuilder::NewForUpdate(opener, "super", 0, 1);
-        if (!builder) return false;
+        CreateLogicalPartitionParams params{
+                .block_device = fake_super,
+                .metadata_slot = 1,
+                .partition_name = name,
+                .timeout_ms = 10s,
+                .partition_opener = &opener,
+        };
 
-        auto metadata = builder->Export();
-        if (!metadata) return false;
-
-        // Update the destination slot, mark it as updated.
-        if (!UpdatePartitionTable(opener, "super", *metadata.get(), 1)) {
-            return false;
+        auto result = sm->OpenSnapshotWriter(params, {});
+        if (!result) {
+            return AssertionFailure() << "Cannot open snapshot for writing: " << name;
+        }
+        if (!result->Initialize()) {
+            return AssertionFailure() << "Cannot initialize snapshot for writing: " << name;
         }
 
-        for (const auto& partition : metadata->partitions) {
-            CreateLogicalPartitionParams params = {
-                    .block_device = fake_super,
-                    .metadata = metadata.get(),
-                    .partition = &partition,
-                    .force_writable = true,
-                    .timeout_ms = 10s,
-                    .device_name = GetPartitionName(partition) + "-base",
-            };
-            std::string ignore_path;
-            if (!CreateLogicalPartition(params, &ignore_path)) {
-                return false;
-            }
+        if (writer) {
+            *writer = std::move(result);
         }
-        return true;
+        return AssertionSuccess();
+    }
+
+    AssertionResult MapUpdateSnapshot(const std::string& name, std::string* path) {
+        TestPartitionOpener opener(fake_super);
+        CreateLogicalPartitionParams params{
+                .block_device = fake_super,
+                .metadata_slot = 1,
+                .partition_name = name,
+                .timeout_ms = 10s,
+                .partition_opener = &opener,
+        };
+
+        auto result = sm->MapUpdateSnapshot(params, path);
+        if (!result) {
+            return AssertionFailure() << "Cannot open snapshot for writing: " << name;
+        }
+        return AssertionSuccess();
     }
 
     AssertionResult DeleteSnapshotDevice(const std::string& snapshot) {
         AssertionResult res = AssertionSuccess();
         if (!(res = DeleteDevice(snapshot))) return res;
+        if (!sm->UnmapDmUserDevice(snapshot)) {
+            return AssertionFailure() << "Cannot delete dm-user device for " << snapshot;
+        }
         if (!(res = DeleteDevice(snapshot + "-inner"))) return res;
         if (!(res = DeleteDevice(snapshot + "-cow"))) return res;
         if (!image_manager_->UnmapImageIfExists(snapshot + "-cow-img")) {
@@ -289,37 +317,55 @@
 
     // Prepare A/B slot for a partition named "test_partition".
     AssertionResult PrepareOneSnapshot(uint64_t device_size,
-                                       std::string* out_snap_device = nullptr) {
-        std::string base_device, cow_device, snap_device;
-        if (!CreatePartition("test_partition_a", device_size)) {
-            return AssertionFailure();
+                                       std::unique_ptr<ICowWriter>* writer = nullptr) {
+        lock_ = nullptr;
+
+        DeltaArchiveManifest manifest;
+
+        auto group = manifest.mutable_dynamic_partition_metadata()->add_groups();
+        group->set_name("group");
+        group->set_size(device_size * 2);
+        group->add_partition_names("test_partition");
+
+        auto pu = manifest.add_partitions();
+        pu->set_partition_name("test_partition");
+        pu->set_estimate_cow_size(device_size);
+        SetSize(pu, device_size);
+
+        auto extent = pu->add_operations()->add_dst_extents();
+        extent->set_start_block(0);
+        if (device_size) {
+            extent->set_num_blocks(device_size / manifest.block_size());
         }
-        if (!MapUpdatePartitions()) {
-            return AssertionFailure();
+
+        TestPartitionOpener opener(fake_super);
+        auto builder = MetadataBuilder::New(opener, "super", 0);
+        if (!builder) {
+            return AssertionFailure() << "Failed to open MetadataBuilder";
         }
-        if (!dm_.GetDmDevicePathByName("test_partition_b-base", &base_device)) {
-            return AssertionFailure();
+        builder->AddGroup("group_a", 16_GiB);
+        builder->AddGroup("group_b", 16_GiB);
+        if (!CreatePartition(builder.get(), "test_partition_a", device_size, nullptr, "group_a")) {
+            return AssertionFailure() << "Failed create test_partition_a";
         }
-        SnapshotStatus status;
-        status.set_name("test_partition_b");
-        status.set_device_size(device_size);
-        status.set_snapshot_size(device_size);
-        status.set_cow_file_size(device_size);
-        if (!sm->CreateSnapshot(lock_.get(), &status)) {
-            return AssertionFailure();
+
+        if (!sm->CreateUpdateSnapshots(manifest)) {
+            return AssertionFailure() << "Failed to create update snapshots";
         }
-        if (!CreateCowImage("test_partition_b")) {
-            return AssertionFailure();
+
+        if (writer) {
+            auto res = MapUpdateSnapshot("test_partition_b", writer);
+            if (!res) {
+                return res;
+            }
+        } else if (!IsCompressionEnabled()) {
+            std::string ignore;
+            if (!MapUpdateSnapshot("test_partition_b", &ignore)) {
+                return AssertionFailure() << "Failed to map test_partition_b";
+            }
         }
-        if (!MapCowImage("test_partition_b", 10s, &cow_device)) {
-            return AssertionFailure();
-        }
-        if (!sm->MapSnapshot(lock_.get(), "test_partition_b", base_device, cow_device, 10s,
-                             &snap_device)) {
-            return AssertionFailure();
-        }
-        if (out_snap_device) {
-            *out_snap_device = std::move(snap_device);
+        if (!AcquireLock()) {
+            return AssertionFailure() << "Failed to acquire lock";
         }
         return AssertionSuccess();
     }
@@ -328,20 +374,37 @@
     AssertionResult SimulateReboot() {
         lock_ = nullptr;
         if (!sm->FinishedSnapshotWrites(false)) {
-            return AssertionFailure();
+            return AssertionFailure() << "Failed to finish snapshot writes";
         }
-        if (!dm_.DeleteDevice("test_partition_b")) {
-            return AssertionFailure();
+        if (!sm->UnmapUpdateSnapshot("test_partition_b")) {
+            return AssertionFailure() << "Failed to unmap COW for test_partition_b";
         }
-        if (!DestroyLogicalPartition("test_partition_b-base")) {
-            return AssertionFailure();
+        if (!dm_.DeleteDeviceIfExists("test_partition_b")) {
+            return AssertionFailure() << "Failed to delete test_partition_b";
         }
-        if (!sm->UnmapCowImage("test_partition_b")) {
-            return AssertionFailure();
+        if (!dm_.DeleteDeviceIfExists("test_partition_b-base")) {
+            return AssertionFailure() << "Failed to destroy test_partition_b-base";
         }
         return AssertionSuccess();
     }
 
+    std::unique_ptr<SnapshotManager> NewManagerForFirstStageMount(
+            const std::string& slot_suffix = "_a") {
+        auto info = new TestDeviceInfo(fake_super, slot_suffix);
+        return NewManagerForFirstStageMount(info);
+    }
+
+    std::unique_ptr<SnapshotManager> NewManagerForFirstStageMount(TestDeviceInfo* info) {
+        auto init = SnapshotManager::NewForFirstStageMount(info);
+        if (!init) {
+            return nullptr;
+        }
+        init->SetUeventRegenCallback([](const std::string& device) -> bool {
+            return android::fs_mgr::WaitForFile(device, snapshot_timeout_);
+        });
+        return init;
+    }
+
     static constexpr std::chrono::milliseconds snapshot_timeout_ = 5s;
     DeviceMapper& dm_;
     std::unique_ptr<SnapshotManager::LockedFile> lock_;
@@ -439,8 +502,7 @@
 TEST_F(SnapshotTest, CleanFirstStageMount) {
     // If there's no update in progress, there should be no first-stage mount
     // needed.
-    TestDeviceInfo* info = new TestDeviceInfo(fake_super);
-    auto sm = SnapshotManager::NewForFirstStageMount(info);
+    auto sm = NewManagerForFirstStageMount();
     ASSERT_NE(sm, nullptr);
     ASSERT_FALSE(sm->NeedSnapshotsInFirstStageMount());
 }
@@ -449,8 +511,7 @@
     ASSERT_TRUE(sm->FinishedSnapshotWrites(false));
 
     // We didn't change the slot, so we shouldn't need snapshots.
-    TestDeviceInfo* info = new TestDeviceInfo(fake_super);
-    auto sm = SnapshotManager::NewForFirstStageMount(info);
+    auto sm = NewManagerForFirstStageMount();
     ASSERT_NE(sm, nullptr);
     ASSERT_FALSE(sm->NeedSnapshotsInFirstStageMount());
 
@@ -462,32 +523,30 @@
     ASSERT_TRUE(AcquireLock());
 
     static const uint64_t kDeviceSize = 1024 * 1024;
-    std::string snap_device;
-    ASSERT_TRUE(PrepareOneSnapshot(kDeviceSize, &snap_device));
 
-    std::string test_string = "This is a test string.";
-    {
-        unique_fd fd(open(snap_device.c_str(), O_RDWR | O_CLOEXEC | O_SYNC));
-        ASSERT_GE(fd, 0);
-        ASSERT_TRUE(android::base::WriteFully(fd, test_string.data(), test_string.size()));
-    }
-
-    // Note: we know there is no inner/outer dm device since we didn't request
-    // a linear segment.
-    DeviceMapper::TargetInfo target;
-    ASSERT_TRUE(sm->IsSnapshotDevice("test_partition_b", &target));
-    ASSERT_EQ(DeviceMapper::GetTargetType(target.spec), "snapshot");
+    std::unique_ptr<ICowWriter> writer;
+    ASSERT_TRUE(PrepareOneSnapshot(kDeviceSize, &writer));
 
     // Release the lock.
     lock_ = nullptr;
 
+    std::string test_string = "This is a test string.";
+    test_string.resize(writer->options().block_size);
+    ASSERT_TRUE(writer->AddRawBlocks(0, test_string.data(), test_string.size()));
+    ASSERT_TRUE(writer->Finalize());
+    writer = nullptr;
+
     // Done updating.
     ASSERT_TRUE(sm->FinishedSnapshotWrites(false));
 
+    ASSERT_TRUE(sm->UnmapUpdateSnapshot("test_partition_b"));
+
     test_device->set_slot_suffix("_b");
+    ASSERT_TRUE(sm->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
     ASSERT_TRUE(sm->InitiateMerge());
 
     // The device should have been switched to a snapshot-merge target.
+    DeviceMapper::TargetInfo target;
     ASSERT_TRUE(sm->IsSnapshotDevice("test_partition_b", &target));
     ASSERT_EQ(DeviceMapper::GetTargetType(target.spec), "snapshot-merge");
 
@@ -503,7 +562,7 @@
     // Test that we can read back the string we wrote to the snapshot. Note
     // that the base device is gone now. |snap_device| contains the correct
     // partition.
-    unique_fd fd(open(snap_device.c_str(), O_RDONLY | O_CLOEXEC));
+    unique_fd fd(open("/dev/block/mapper/test_partition_b", O_RDONLY | O_CLOEXEC));
     ASSERT_GE(fd, 0);
 
     std::string buffer(test_string.size(), '\0');
@@ -518,7 +577,7 @@
     ASSERT_TRUE(PrepareOneSnapshot(kDeviceSize));
     ASSERT_TRUE(SimulateReboot());
 
-    auto init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_b"));
+    auto init = NewManagerForFirstStageMount("_b");
     ASSERT_NE(init, nullptr);
     ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -547,7 +606,7 @@
     FormatFakeSuper();
     ASSERT_TRUE(CreatePartition("test_partition_b", kDeviceSize));
 
-    auto init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_b"));
+    auto init = NewManagerForFirstStageMount("_b");
     ASSERT_NE(init, nullptr);
     ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -574,7 +633,7 @@
     ASSERT_TRUE(PrepareOneSnapshot(kDeviceSize));
     ASSERT_TRUE(SimulateReboot());
 
-    auto init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_b"));
+    auto init = NewManagerForFirstStageMount("_b");
     ASSERT_NE(init, nullptr);
     ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -899,47 +958,7 @@
         return AssertionSuccess();
     }
 
-    AssertionResult MapUpdateSnapshot(const std::string& name,
-                                      std::unique_ptr<ICowWriter>* writer) {
-        CreateLogicalPartitionParams params{
-                .block_device = fake_super,
-                .metadata_slot = 1,
-                .partition_name = name,
-                .timeout_ms = 10s,
-                .partition_opener = opener_.get(),
-        };
-
-        auto result = sm->OpenSnapshotWriter(params, {});
-        if (!result) {
-            return AssertionFailure() << "Cannot open snapshot for writing: " << name;
-        }
-        if (!result->Initialize()) {
-            return AssertionFailure() << "Cannot initialize snapshot for writing: " << name;
-        }
-
-        if (writer) {
-            *writer = std::move(result);
-        }
-        return AssertionSuccess();
-    }
-
-    AssertionResult MapUpdateSnapshot(const std::string& name, std::string* path) {
-        CreateLogicalPartitionParams params{
-                .block_device = fake_super,
-                .metadata_slot = 1,
-                .partition_name = name,
-                .timeout_ms = 10s,
-                .partition_opener = opener_.get(),
-        };
-
-        auto result = sm->MapUpdateSnapshot(params, path);
-        if (!result) {
-            return AssertionFailure() << "Cannot open snapshot for writing: " << name;
-        }
-        return AssertionSuccess();
-    }
-
-    AssertionResult MapUpdateSnapshot(const std::string& name) {
+    AssertionResult MapOneUpdateSnapshot(const std::string& name) {
         if (IsCompressionEnabled()) {
             std::unique_ptr<ICowWriter> writer;
             return MapUpdateSnapshot(name, &writer);
@@ -983,7 +1002,7 @@
     AssertionResult MapUpdateSnapshots(const std::vector<std::string>& names = {"sys_b", "vnd_b",
                                                                                 "prd_b"}) {
         for (const auto& name : names) {
-            auto res = MapUpdateSnapshot(name);
+            auto res = MapOneUpdateSnapshot(name);
             if (!res) {
                 return res;
             }
@@ -1070,7 +1089,7 @@
     ASSERT_TRUE(UnmapAll());
 
     // After reboot, init does first stage mount.
-    auto init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_b"));
+    auto init = NewManagerForFirstStageMount("_b");
     ASSERT_NE(init, nullptr);
     ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -1202,7 +1221,7 @@
     ASSERT_TRUE(UnmapAll());
 
     // After reboot, init does first stage mount.
-    auto init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_b"));
+    auto init = NewManagerForFirstStageMount("_b");
     ASSERT_NE(init, nullptr);
     ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -1214,7 +1233,7 @@
 
     // Simulate shutting down the device again.
     ASSERT_TRUE(UnmapAll());
-    init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_a"));
+    init = NewManagerForFirstStageMount("_a");
     ASSERT_NE(init, nullptr);
     ASSERT_FALSE(init->NeedSnapshotsInFirstStageMount());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -1251,7 +1270,7 @@
     ASSERT_TRUE(UnmapAll());
 
     // After reboot, init does first stage mount.
-    auto init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_b"));
+    auto init = NewManagerForFirstStageMount("_b");
     ASSERT_NE(init, nullptr);
     ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -1387,8 +1406,8 @@
     ASSERT_TRUE(UnmapAll());
 
     // After reboot, init does first stage mount.
-    // Normally we should use NewForFirstStageMount, but if so, "gsid.mapped_image.sys_b-cow-img"
-    // won't be set.
+    // Normally we should use NewManagerForFirstStageMount, but if so,
+    // "gsid.mapped_image.sys_b-cow-img" won't be set.
     auto init = SnapshotManager::New(new TestDeviceInfo(fake_super, "_b"));
     ASSERT_NE(init, nullptr);
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -1495,7 +1514,7 @@
     ASSERT_TRUE(UnmapAll());
 
     // After reboot, init does first stage mount.
-    auto init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_b"));
+    auto init = NewManagerForFirstStageMount("_b");
     ASSERT_NE(init, nullptr);
     ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -1509,7 +1528,7 @@
     // Simulate a reboot into recovery.
     auto test_device = std::make_unique<TestDeviceInfo>(fake_super, "_b");
     test_device->set_recovery(true);
-    new_sm = SnapshotManager::NewForFirstStageMount(test_device.release());
+    new_sm = NewManagerForFirstStageMount(test_device.release());
 
     ASSERT_TRUE(new_sm->HandleImminentDataWipe());
     ASSERT_EQ(new_sm->GetUpdateState(), UpdateState::None);
@@ -1527,7 +1546,7 @@
     ASSERT_TRUE(UnmapAll());
 
     // After reboot, init does first stage mount.
-    auto init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_b"));
+    auto init = NewManagerForFirstStageMount("_b");
     ASSERT_NE(init, nullptr);
     ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -1541,7 +1560,7 @@
     // Simulate a reboot into recovery.
     auto test_device = std::make_unique<TestDeviceInfo>(fake_super, "_b");
     test_device->set_recovery(true);
-    new_sm = SnapshotManager::NewForFirstStageMount(test_device.release());
+    new_sm = NewManagerForFirstStageMount(test_device.release());
 
     ASSERT_TRUE(new_sm->FinishMergeInRecovery());
 
@@ -1551,12 +1570,12 @@
 
     // Finish the merge in a normal boot.
     test_device = std::make_unique<TestDeviceInfo>(fake_super, "_b");
-    init = SnapshotManager::NewForFirstStageMount(test_device.release());
+    init = NewManagerForFirstStageMount(test_device.release());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
     init = nullptr;
 
     test_device = std::make_unique<TestDeviceInfo>(fake_super, "_b");
-    new_sm = SnapshotManager::NewForFirstStageMount(test_device.release());
+    new_sm = NewManagerForFirstStageMount(test_device.release());
     ASSERT_EQ(new_sm->ProcessUpdateState(), UpdateState::MergeCompleted);
     ASSERT_EQ(new_sm->ProcessUpdateState(), UpdateState::None);
 }
@@ -1575,7 +1594,7 @@
     // Simulate a reboot into recovery.
     auto test_device = new TestDeviceInfo(fake_super, "_b");
     test_device->set_recovery(true);
-    auto new_sm = SnapshotManager::NewForFirstStageMount(test_device);
+    auto new_sm = NewManagerForFirstStageMount(test_device);
 
     ASSERT_TRUE(new_sm->HandleImminentDataWipe());
     // Manually mount metadata so that we can call GetUpdateState() below.
@@ -1600,7 +1619,7 @@
     // Simulate a rollback, with reboot into recovery.
     auto test_device = new TestDeviceInfo(fake_super, "_a");
     test_device->set_recovery(true);
-    auto new_sm = SnapshotManager::NewForFirstStageMount(test_device);
+    auto new_sm = NewManagerForFirstStageMount(test_device);
 
     ASSERT_TRUE(new_sm->HandleImminentDataWipe());
     EXPECT_EQ(new_sm->GetUpdateState(), UpdateState::None);
@@ -1628,7 +1647,7 @@
     // Simulate a reboot into recovery.
     auto test_device = new TestDeviceInfo(fake_super, "_b");
     test_device->set_recovery(true);
-    auto new_sm = SnapshotManager::NewForFirstStageMount(test_device);
+    auto new_sm = NewManagerForFirstStageMount(test_device);
 
     ASSERT_TRUE(new_sm->HandleImminentDataWipe());
     // Manually mount metadata so that we can call GetUpdateState() below.
@@ -1639,7 +1658,7 @@
 
     // Now reboot into new slot.
     test_device = new TestDeviceInfo(fake_super, "_b");
-    auto init = SnapshotManager::NewForFirstStageMount(test_device);
+    auto init = NewManagerForFirstStageMount(test_device);
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
     // Verify that we are on the downgraded build.
     for (const auto& name : {"sys_b", "vnd_b", "prd_b"}) {
@@ -1685,7 +1704,7 @@
     ASSERT_TRUE(UnmapAll());
 
     // After reboot, init does first stage mount.
-    auto init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_b"));
+    auto init = NewManagerForFirstStageMount("_b");
     ASSERT_NE(init, nullptr);
     ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
     ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
@@ -1773,7 +1792,7 @@
     ASSERT_TRUE(sm->FinishedSnapshotWrites(false));
     ASSERT_TRUE(UnmapAll());
 
-    auto init = SnapshotManager::NewForFirstStageMount(new TestDeviceInfo(fake_super, "_b"));
+    auto init = NewManagerForFirstStageMount("_b");
     ASSERT_NE(init, nullptr);
 
     ASSERT_TRUE(init->EnsureSnapuserdConnected());
@@ -1785,7 +1804,7 @@
     ASSERT_EQ(access("/dev/dm-user/sys_b-user-cow-init", F_OK), 0);
     ASSERT_EQ(access("/dev/dm-user/sys_b-user-cow", F_OK), -1);
 
-    ASSERT_TRUE(init->PerformSecondStageTransition());
+    ASSERT_TRUE(init->PerformInitTransition(SnapshotManager::InitTransition::SECOND_STAGE));
 
     // The control device should have been renamed.
     ASSERT_TRUE(android::fs_mgr::WaitForFileDeleted("/dev/dm-user/sys_b-user-cow-init", 10s));
@@ -1887,8 +1906,7 @@
     ASSERT_TRUE(UnmapAll());
 
     // Simulate reboot. After reboot, init does first stage mount.
-    auto init = SnapshotManager::NewForFirstStageMount(
-            new TestDeviceInfo(fake_super, flashed_slot_suffix));
+    auto init = NewManagerForFirstStageMount(flashed_slot_suffix);
     ASSERT_NE(init, nullptr);
 
     if (flashed_slot && after_merge) {
diff --git a/fs_mgr/libsnapshot/snapuserd.cpp b/fs_mgr/libsnapshot/snapuserd.cpp
index 4f94806..573b6a5 100644
--- a/fs_mgr/libsnapshot/snapuserd.cpp
+++ b/fs_mgr/libsnapshot/snapuserd.cpp
@@ -129,88 +129,126 @@
     return true;
 }
 
-/*
- * Read the data of size bytes from a given chunk.
- *
- * Kernel can potentially merge the blocks if the
- * successive chunks are contiguous. For chunk size of 8,
- * there can be 256 disk exceptions; and if
- * all 256 disk exceptions are contiguous, kernel can merge
- * them into a single IO.
- *
- * Since each chunk in the disk exception
- * mapping represents a 4k block, kernel can potentially
- * issue 256*4k = 1M IO in one shot.
- *
- * Even though kernel assumes that the blocks are
- * contiguous, we need to split the 1M IO into 4k chunks
- * as each operation represents 4k and it can either be:
- *
- * 1: Replace operation
- * 2: Copy operation
- * 3: Zero operation
- *
- */
-bool Snapuserd::ReadData(chunk_t chunk, size_t size) {
-    size_t read_size = size;
-    bool ret = true;
-    chunk_t chunk_key = chunk;
+bool Snapuserd::ProcessCowOp(const CowOperation* cow_op) {
+    CHECK(cow_op != nullptr);
 
-    if (!((read_size & (BLOCK_SIZE - 1)) == 0)) {
-        SNAP_LOG(ERROR) << "ReadData - unaligned read_size: " << read_size;
-        return false;
+    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 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";
+        return -1;
     }
 
-    while (read_size > 0) {
-        const CowOperation* cow_op = chunk_map_[chunk_key];
-        CHECK(cow_op != nullptr);
+    int num_sectors_skip = sector - it->first;
 
-        switch (cow_op->type) {
-            case kCowReplaceOp: {
-                ret = ProcessReplaceOp(cow_op);
-                break;
-            }
+    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]));
 
-            case kCowZeroOp: {
-                ret = ProcessZeroOp();
-                break;
-            }
+        memmove(msg->payload.buf, (char*)msg->payload.buf + skip_sector_size,
+                (BLOCK_SIZE - skip_sector_size));
+    }
 
-            case kCowCopyOp: {
-                ret = ProcessCopyOp(cow_op);
-                break;
-            }
+    bufsink_.ResetBufferOffset();
+    return std::min(size, (BLOCK_SIZE - skip_sector_size));
+}
 
-            default: {
-                SNAP_LOG(ERROR) << "Unknown operation-type found: " << cow_op->type;
-                ret = false;
-                break;
-            }
+/*
+ * 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 (!ret) {
-            SNAP_LOG(ERROR) << "ReadData failed for operation: " << cow_op->type;
-            return false;
-        }
+        /*
+         * 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_SIZE);
+    while (num_ops) {
+        if (!ProcessCowOp(it->second)) {
+            return -1;
+        }
+        num_ops -= 1;
+        it++;
         // Update the buffer offset
         bufsink_.UpdateBufferOffset(BLOCK_SIZE);
 
-        read_size -= BLOCK_SIZE;
-
-        // Start iterating the chunk incrementally; Since while
-        // constructing the metadata, we know that the chunk IDs
-        // are contiguous
-        chunk_key += 1;
-
-        if (cow_op->type == kCowCopyOp) {
-            CHECK(read_size == 0);
-        }
+        SNAP_LOG(DEBUG) << "ReadData at sector: " << sector << " size: " << size;
     }
 
     // Reset the buffer offset
     bufsink_.ResetBufferOffset();
-    return ret;
+    return size;
 }
 
 /*
@@ -261,9 +299,7 @@
     if (divresult.quot < vec_.size()) {
         size = exceptions_per_area_ * sizeof(struct disk_exception);
 
-        if (read_size > size) {
-            return false;
-        }
+        CHECK(read_size == size);
 
         void* buffer = bufsink_.GetPayloadBuffer(size);
         CHECK(buffer != nullptr);
@@ -329,7 +365,7 @@
         if (cow_de->new_chunk != 0) {
             merged_ops_cur_iter += 1;
             offset += sizeof(struct disk_exception);
-            const CowOperation* cow_op = chunk_map_[cow_de->new_chunk];
+            const CowOperation* cow_op = chunk_map_[ChunkToSector(cow_de->new_chunk)];
             CHECK(cow_op != nullptr);
             CHECK(cow_op->new_block == cow_de->old_chunk);
             if (cow_op->type == kCowCopyOp) {
@@ -563,7 +599,7 @@
         SNAP_LOG(DEBUG) << "Old-chunk: " << de->old_chunk << "New-chunk: " << de->new_chunk;
 
         // Store operation pointer.
-        chunk_map_[data_chunk_id] = cow_op;
+        chunk_map_[ChunkToSector(data_chunk_id)] = cow_op;
         num_ops += 1;
         offset += sizeof(struct disk_exception);
         cowop_riter_->Next();
@@ -680,6 +716,126 @@
     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_SIZE);
+
+    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_SIZE);
+            ConstructKernelCowHeader();
+            SNAP_LOG(DEBUG) << "Kernel header constructed";
+        } else {
+            if (!offset && (read_size == BLOCK_SIZE) &&
+                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: " << header->sector;
+                    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();
 
@@ -698,122 +854,16 @@
 
     switch (header->type) {
         case DM_USER_REQ_MAP_READ: {
-            size_t remaining_size = header->len;
-            loff_t offset = 0;
-            do {
-                size_t read_size = std::min(PAYLOAD_SIZE, remaining_size);
-                header->type = DM_USER_RESP_SUCCESS;
-
-                // 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_SIZE);
-                    ConstructKernelCowHeader();
-                    SNAP_LOG(DEBUG) << "Kernel header constructed";
-                } else {
-                    // Convert the sector number to a chunk ID.
-                    //
-                    // Check if the chunk ID represents a metadata
-                    // page. If the chunk ID is not found in the
-                    // vector, then it points to a metadata page.
-                    chunk_t chunk = SectorToChunk(header->sector);
-
-                    if (chunk_map_.find(chunk) == 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 {
-                        SNAP_LOG(DEBUG) << "ReadData: chunk: " << chunk << " len: " << header->len
-                                        << " read_size: " << read_size << " offset: " << offset;
-                        chunk_t num_chunks_read = (offset >> BLOCK_SHIFT);
-                        if (!ReadData(chunk + num_chunks_read, read_size)) {
-                            SNAP_LOG(ERROR) << "ReadData failed for chunk id: " << chunk
-                                            << "Sector: " << header->sector;
-                            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(read_size)) {
-                    return false;
-                }
-
-                remaining_size -= read_size;
-                offset += read_size;
-            } while (remaining_size);
-
+            if (!DmuserReadRequest()) {
+                return false;
+            }
             break;
         }
 
         case DM_USER_REQ_MAP_WRITE: {
-            // 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;
-                }
-                break;
-            }
-
-            size_t remaining_size = header->len;
-            size_t read_size = std::min(PAYLOAD_SIZE, remaining_size);
-            CHECK(read_size == BLOCK_SIZE);
-
-            CHECK(header->sector > 0);
-            chunk_t chunk = SectorToChunk(header->sector);
-            CHECK(chunk_map_.find(chunk) == 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)) {
+            if (!DmuserWriteRequest()) {
                 return false;
             }
-
             break;
         }
     }
diff --git a/fs_mgr/libsnapshot/snapuserd.h b/fs_mgr/libsnapshot/snapuserd.h
index eec6d45..c01fee3 100644
--- a/fs_mgr/libsnapshot/snapuserd.h
+++ b/fs_mgr/libsnapshot/snapuserd.h
@@ -22,9 +22,9 @@
 #include <cstring>
 #include <iostream>
 #include <limits>
+#include <map>
 #include <string>
 #include <thread>
-#include <unordered_map>
 #include <vector>
 
 #include <android-base/file.h>
@@ -72,6 +72,9 @@
     bool IsAttached() const { return ctrl_fd_ >= 0; }
 
   private:
+    bool DmuserReadRequest();
+    bool DmuserWriteRequest();
+
     bool ReadDmUserHeader();
     bool ReadDmUserPayload(void* buffer, size_t size);
     bool WriteDmUserPayload(size_t size);
@@ -79,10 +82,13 @@
     bool ReadMetadata();
     bool ZerofillDiskExceptions(size_t read_size);
     bool ReadDiskExceptions(chunk_t chunk, size_t size);
-    bool ReadData(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 ProcessCowOp(const CowOperation* cow_op);
     bool ProcessReplaceOp(const CowOperation* cow_op);
     bool ProcessCopyOp(const CowOperation* cow_op);
     bool ProcessZeroOp();
@@ -94,6 +100,7 @@
     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; }
+    bool IsBlockAligned(int read_size) { return ((read_size & (BLOCK_SIZE - 1)) == 0); }
 
     std::string cow_device_;
     std::string backing_store_device_;
@@ -116,9 +123,15 @@
     // mapping of old-chunk to new-chunk
     std::vector<std::unique_ptr<uint8_t[]>> vec_;
 
-    // Key - Chunk ID
+    // Key - Sector
     // Value - cow operation
-    std::unordered_map<chunk_t, const CowOperation*> chunk_map_;
+    //
+    // 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_;
 
     bool metadata_read_done_ = false;
     BufferSink bufsink_;
diff --git a/init/Android.bp b/init/Android.bp
index 19ba21b..cd295cf 100644
--- a/init/Android.bp
+++ b/init/Android.bp
@@ -60,6 +60,7 @@
     "selabel.cpp",
     "selinux.cpp",
     "sigchld_handler.cpp",
+    "snapuserd_transition.cpp",
     "switch_root.cpp",
     "uevent_listener.cpp",
     "ueventd.cpp",
diff --git a/init/Android.mk b/init/Android.mk
index c881e2f..561d641 100644
--- a/init/Android.mk
+++ b/init/Android.mk
@@ -57,6 +57,8 @@
     reboot_utils.cpp \
     selabel.cpp \
     selinux.cpp \
+    service_utils.cpp \
+    snapuserd_transition.cpp \
     switch_root.cpp \
     uevent_listener.cpp \
     util.cpp \
diff --git a/init/block_dev_initializer.h b/init/block_dev_initializer.h
index 79fe4ec..ec39ce0 100644
--- a/init/block_dev_initializer.h
+++ b/init/block_dev_initializer.h
@@ -12,6 +12,8 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#pragma once
+
 #include <memory>
 #include <set>
 #include <string>
diff --git a/init/first_stage_init.cpp b/init/first_stage_init.cpp
index 0949fc5..6954c03 100644
--- a/init/first_stage_init.cpp
+++ b/init/first_stage_init.cpp
@@ -42,6 +42,7 @@
 #include "first_stage_mount.h"
 #include "reboot_utils.h"
 #include "second_stage_resources.h"
+#include "snapuserd_transition.h"
 #include "switch_root.h"
 #include "util.h"
 
@@ -90,6 +91,12 @@
                     }
                 }
             }
+        } else if (de->d_type == DT_REG) {
+            // Do not free snapuserd if we will need the ramdisk copy during the
+            // selinux transition.
+            if (de->d_name == "snapuserd"s && IsFirstStageSnapuserdRunning()) {
+                continue;
+            }
         }
         unlinkat(dfd, de->d_name, is_dir ? AT_REMOVEDIR : 0);
     }
diff --git a/init/first_stage_mount.cpp b/init/first_stage_mount.cpp
index a0511cc..7c46281 100644
--- a/init/first_stage_mount.cpp
+++ b/init/first_stage_mount.cpp
@@ -44,6 +44,7 @@
 
 #include "block_dev_initializer.h"
 #include "devices.h"
+#include "snapuserd_transition.h"
 #include "switch_root.h"
 #include "uevent.h"
 #include "uevent_listener.h"
@@ -87,6 +88,7 @@
   protected:
     bool InitRequiredDevices(std::set<std::string> devices);
     bool CreateLogicalPartitions();
+    bool CreateSnapshotPartitions(android::snapshot::SnapshotManager* sm);
     bool MountPartition(const Fstab::iterator& begin, bool erase_same_mounts,
                         Fstab::iterator* end = nullptr);
 
@@ -109,6 +111,7 @@
 
     bool need_dm_verity_;
     bool dsu_not_on_userdata_ = false;
+    bool use_snapuserd_ = false;
 
     Fstab fstab_;
     // The super path is only set after InitDevices, and is invalid before.
@@ -338,21 +341,7 @@
             return false;
         }
         if (sm->NeedSnapshotsInFirstStageMount()) {
-            // When COW images are present for snapshots, they are stored on
-            // the data partition.
-            if (!InitRequiredDevices({"userdata"})) {
-                return false;
-            }
-            sm->SetUeventRegenCallback([this](const std::string& device) -> bool {
-                if (android::base::StartsWith(device, "/dev/block/dm-")) {
-                    return block_dev_init_.InitDmDevice(device);
-                }
-                if (android::base::StartsWith(device, "/dev/dm-user/")) {
-                    return block_dev_init_.InitDmUser(android::base::Basename(device));
-                }
-                return block_dev_init_.InitDevices({device});
-            });
-            return sm->CreateLogicalAndSnapshotPartitions(super_path_);
+            return CreateSnapshotPartitions(sm.get());
         }
     }
 
@@ -367,6 +356,37 @@
     return android::fs_mgr::CreateLogicalPartitions(*metadata.get(), super_path_);
 }
 
+bool FirstStageMount::CreateSnapshotPartitions(SnapshotManager* sm) {
+    // When COW images are present for snapshots, they are stored on
+    // the data partition.
+    if (!InitRequiredDevices({"userdata"})) {
+        return false;
+    }
+
+    use_snapuserd_ = sm->IsSnapuserdRequired();
+    if (use_snapuserd_) {
+        LaunchFirstStageSnapuserd();
+    }
+
+    sm->SetUeventRegenCallback([this](const std::string& device) -> bool {
+        if (android::base::StartsWith(device, "/dev/block/dm-")) {
+            return block_dev_init_.InitDmDevice(device);
+        }
+        if (android::base::StartsWith(device, "/dev/dm-user/")) {
+            return block_dev_init_.InitDmUser(android::base::Basename(device));
+        }
+        return block_dev_init_.InitDevices({device});
+    });
+    if (!sm->CreateLogicalAndSnapshotPartitions(super_path_)) {
+        return false;
+    }
+
+    if (use_snapuserd_) {
+        CleanupSnapuserdSocket();
+    }
+    return true;
+}
+
 bool FirstStageMount::MountPartition(const Fstab::iterator& begin, bool erase_same_mounts,
                                      Fstab::iterator* end) {
     // Sets end to begin + 1, so we can just return on failure below.
@@ -466,6 +486,10 @@
 
     if (system_partition == fstab_.end()) return true;
 
+    if (use_snapuserd_) {
+        SaveRamdiskPathToSnapuserd();
+    }
+
     if (MountPartition(system_partition, false /* erase_same_mounts */)) {
         if (dsu_not_on_userdata_ && fs_mgr_verity_is_check_at_most_once(*system_partition)) {
             LOG(ERROR) << "check_most_at_once forbidden on external media";
diff --git a/init/init.cpp b/init/init.cpp
index d1998a5..d6753f5 100644
--- a/init/init.cpp
+++ b/init/init.cpp
@@ -79,6 +79,7 @@
 #include "service.h"
 #include "service_parser.h"
 #include "sigchld_handler.h"
+#include "snapuserd_transition.h"
 #include "subcontext.h"
 #include "system/core/init/property_service.pb.h"
 #include "util.h"
@@ -741,10 +742,15 @@
         return {};
     }
     svc->Start();
+    svc->SetShutdownCritical();
 
-    if (!sm->PerformSecondStageTransition()) {
+    if (!sm->PerformSecondStageInitTransition()) {
         LOG(FATAL) << "Failed to transition snapuserd to second-stage";
     }
+
+    if (auto pid = GetSnapuserdFirstStagePid()) {
+        KillFirstStageSnapuserd(pid.value());
+    }
     return {};
 }
 
diff --git a/init/mount_namespace.cpp b/init/mount_namespace.cpp
index 59cc140..ec48cde 100644
--- a/init/mount_namespace.cpp
+++ b/init/mount_namespace.cpp
@@ -115,22 +115,29 @@
         return {};
     }
     dirent* entry;
+    std::vector<std::string> entries;
+
     while ((entry = readdir(dir.get())) != nullptr) {
         if (entry->d_name[0] == '.') continue;
         if (entry->d_type == DT_DIR) {
-            const std::string apex_path = from_dir + "/" + entry->d_name;
-            const auto apex_manifest = GetApexManifest(apex_path);
-            if (!apex_manifest.ok()) {
-                LOG(ERROR) << apex_path << " is not an APEX directory: " << apex_manifest.error();
-                continue;
-            }
-            const std::string mount_path = to_dir + "/" + apex_manifest->name();
-            if (auto result = MountDir(apex_path, mount_path); !result.ok()) {
-                return result;
-            }
-            on_activate(apex_path, *apex_manifest);
+            entries.push_back(entry->d_name);
         }
     }
+
+    std::sort(entries.begin(), entries.end());
+    for (const auto& name : entries) {
+        const std::string apex_path = from_dir + "/" + name;
+        const auto apex_manifest = GetApexManifest(apex_path);
+        if (!apex_manifest.ok()) {
+            LOG(ERROR) << apex_path << " is not an APEX directory: " << apex_manifest.error();
+            continue;
+        }
+        const std::string mount_path = to_dir + "/" + apex_manifest->name();
+        if (auto result = MountDir(apex_path, mount_path); !result.ok()) {
+            return result;
+        }
+        on_activate(apex_path, *apex_manifest);
+    }
     return {};
 }
 
diff --git a/init/reboot.cpp b/init/reboot.cpp
index 4699eca..e3aaa38 100644
--- a/init/reboot.cpp
+++ b/init/reboot.cpp
@@ -655,6 +655,7 @@
 
         if (do_shutdown_animation) {
             SetProperty("service.bootanim.exit", "0");
+            SetProperty("service.bootanim.progress", "0");
             // Could be in the middle of animation. Stop and start so that it can pick
             // up the right mode.
             boot_anim->Stop();
diff --git a/init/selinux.cpp b/init/selinux.cpp
index f03ca6b..0336936 100644
--- a/init/selinux.cpp
+++ b/init/selinux.cpp
@@ -45,7 +45,7 @@
 // 2) If these hashes do not match, then either /system or /system_ext or /product (or some of them)
 //    have been updated out of sync with /vendor (or /odm if it is present) and the init needs to
 //    compile the SEPolicy.  /system contains the SEPolicy compiler, secilc, and it is used by the
-//    LoadSplitPolicy() function below to compile the SEPolicy to a temp directory and load it.
+//    OpenSplitPolicy() function below to compile the SEPolicy to a temp directory and load it.
 //    That function contains even more documentation with the specific implementation details of how
 //    the SEPolicy is compiled if needed.
 
@@ -74,6 +74,7 @@
 #include "block_dev_initializer.h"
 #include "debug_ramdisk.h"
 #include "reboot_utils.h"
+#include "snapuserd_transition.h"
 #include "util.h"
 
 using namespace std::string_literals;
@@ -298,7 +299,12 @@
     return access(plat_policy_cil_file, R_OK) != -1;
 }
 
-bool LoadSplitPolicy() {
+struct PolicyFile {
+    unique_fd fd;
+    std::string path;
+};
+
+bool OpenSplitPolicy(PolicyFile* policy_file) {
     // IMPLEMENTATION NOTE: Split policy consists of three CIL files:
     // * platform -- policy needed due to logic contained in the system image,
     // * non-platform -- policy needed due to logic contained in the vendor image,
@@ -325,10 +331,8 @@
     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) {
-            if (selinux_android_load_policy_from_fd(fd, precompiled_sepolicy_file.c_str()) < 0) {
-                LOG(ERROR) << "Failed to load SELinux policy from " << precompiled_sepolicy_file;
-                return false;
-            }
+            policy_file->fd = std::move(fd);
+            policy_file->path = std::move(precompiled_sepolicy_file);
             return true;
         }
     }
@@ -446,34 +450,39 @@
     }
     unlink(compiled_sepolicy);
 
-    LOG(INFO) << "Loading compiled SELinux policy";
-    if (selinux_android_load_policy_from_fd(compiled_sepolicy_fd, compiled_sepolicy) < 0) {
-        LOG(ERROR) << "Failed to load SELinux policy from " << compiled_sepolicy;
-        return false;
-    }
-
+    policy_file->fd = std::move(compiled_sepolicy_fd);
+    policy_file->path = compiled_sepolicy;
     return true;
 }
 
-bool LoadMonolithicPolicy() {
-    LOG(VERBOSE) << "Loading SELinux policy from monolithic file";
-    if (selinux_android_load_policy() < 0) {
-        PLOG(ERROR) << "Failed to load monolithic SELinux policy";
+bool OpenMonolithicPolicy(PolicyFile* policy_file) {
+    static constexpr char kSepolicyFile[] = "/sepolicy";
+
+    LOG(VERBOSE) << "Opening SELinux policy from monolithic file";
+    policy_file->fd.reset(open(kSepolicyFile, O_RDONLY | O_CLOEXEC | O_NOFOLLOW));
+    if (policy_file->fd < 0) {
+        PLOG(ERROR) << "Failed to open monolithic SELinux policy";
         return false;
     }
+    policy_file->path = kSepolicyFile;
     return true;
 }
 
-bool LoadPolicy() {
-    return IsSplitPolicyDevice() ? LoadSplitPolicy() : LoadMonolithicPolicy();
-}
+void ReadPolicy(std::string* policy) {
+    PolicyFile policy_file;
 
-void SelinuxInitialize() {
-    LOG(INFO) << "Loading SELinux policy";
-    if (!LoadPolicy()) {
-        LOG(FATAL) << "Unable to load SELinux policy";
+    bool ok = IsSplitPolicyDevice() ? OpenSplitPolicy(&policy_file)
+                                    : OpenMonolithicPolicy(&policy_file);
+    if (!ok) {
+        LOG(FATAL) << "Unable to open SELinux policy";
     }
 
+    if (!android::base::ReadFdToString(policy_file.fd, policy)) {
+        PLOG(FATAL) << "Failed to read policy file: " << policy_file.path;
+    }
+}
+
+void SelinuxSetEnforcement() {
     bool kernel_enforcing = (security_getenforce() == 1);
     bool is_enforcing = IsEnforcing();
     if (kernel_enforcing != is_enforcing) {
@@ -670,6 +679,30 @@
     }
 }
 
+static void LoadSelinuxPolicy(std::string& policy) {
+    LOG(INFO) << "Loading SELinux policy";
+
+    set_selinuxmnt("/sys/fs/selinux");
+    if (security_load_policy(policy.data(), policy.size()) < 0) {
+        PLOG(FATAL) << "SELinux:  Could not load policy";
+    }
+}
+
+// The SELinux setup process is carefully orchestrated around snapuserd. Policy
+// must be loaded off dynamic partitions, and during an OTA, those partitions
+// cannot be read without snapuserd. But, with kernel-privileged snapuserd
+// running, loading the policy will immediately trigger audits.
+//
+// We use a five-step process to address this:
+//  (1) Read the policy into a string, with snapuserd running.
+//  (2) Rewrite the snapshot device-mapper tables, to generate new dm-user
+//      devices and to flush I/O.
+//  (3) Kill snapuserd, which no longer has any dm-user devices to attach to.
+//  (4) Load the sepolicy and issue critical restorecons in /dev, carefully
+//      avoiding anything that would read from /system.
+//  (5) Re-launch snapuserd and attach it to the dm-user devices from step (2).
+//
+// After this sequence, it is safe to enable enforcing mode and continue booting.
 int SetupSelinux(char** argv) {
     SetStdioToDevNull(argv);
     InitKernelLogging(argv);
@@ -682,9 +715,31 @@
 
     MountMissingSystemPartitions();
 
-    // Set up SELinux, loading the SELinux policy.
     SelinuxSetupKernelLogging();
-    SelinuxInitialize();
+
+    LOG(INFO) << "Opening SELinux policy";
+
+    // Read the policy before potentially killing snapuserd.
+    std::string policy;
+    ReadPolicy(&policy);
+
+    auto snapuserd_helper = SnapuserdSelinuxHelper::CreateIfNeeded();
+    if (snapuserd_helper) {
+        // Kill the old snapused to avoid audit messages. After this we cannot
+        // read from /system (or other dynamic partitions) until we call
+        // FinishTransition().
+        snapuserd_helper->StartTransition();
+    }
+
+    LoadSelinuxPolicy(policy);
+
+    if (snapuserd_helper) {
+        // Before enforcing, finish the pending snapuserd transition.
+        snapuserd_helper->FinishTransition();
+        snapuserd_helper = nullptr;
+    }
+
+    SelinuxSetEnforcement();
 
     // We're in the kernel domain and want to transition to the init domain.  File systems that
     // store SELabels in their xattrs, such as ext4 do not need an explicit restorecon here,
diff --git a/init/snapuserd_transition.cpp b/init/snapuserd_transition.cpp
new file mode 100644
index 0000000..19b5c57
--- /dev/null
+++ b/init/snapuserd_transition.cpp
@@ -0,0 +1,305 @@
+/*
+ * 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_transition.h"
+
+#include <sys/mman.h>
+#include <sys/socket.h>
+#include <sys/syscall.h>
+#include <sys/xattr.h>
+#include <unistd.h>
+
+#include <filesystem>
+#include <string>
+
+#include <android-base/file.h>
+#include <android-base/logging.h>
+#include <android-base/parseint.h>
+#include <android-base/strings.h>
+#include <android-base/unique_fd.h>
+#include <cutils/sockets.h>
+#include <libsnapshot/snapshot.h>
+#include <libsnapshot/snapuserd_client.h>
+#include <private/android_filesystem_config.h>
+#include <selinux/android.h>
+
+#include "block_dev_initializer.h"
+#include "service_utils.h"
+#include "util.h"
+
+namespace android {
+namespace init {
+
+using namespace std::string_literals;
+
+using android::base::unique_fd;
+using android::snapshot::SnapshotManager;
+using android::snapshot::SnapuserdClient;
+
+static constexpr char kSnapuserdPath[] = "/system/bin/snapuserd";
+static constexpr char kSnapuserdFirstStagePidVar[] = "FIRST_STAGE_SNAPUSERD_PID";
+static constexpr char kSnapuserdFirstStageFdVar[] = "FIRST_STAGE_SNAPUSERD_FD";
+static constexpr char kSnapuserdLabel[] = "u:object_r:snapuserd_exec:s0";
+static constexpr char kSnapuserdSocketLabel[] = "u:object_r:snapuserd_socket:s0";
+
+void LaunchFirstStageSnapuserd() {
+    SocketDescriptor socket_desc;
+    socket_desc.name = android::snapshot::kSnapuserdSocket;
+    socket_desc.type = SOCK_STREAM;
+    socket_desc.perm = 0660;
+    socket_desc.uid = AID_SYSTEM;
+    socket_desc.gid = AID_SYSTEM;
+
+    // We specify a label here even though it technically is not needed. During
+    // first_stage_mount there is no sepolicy loaded. Once sepolicy is loaded,
+    // we bypass the socket entirely.
+    auto socket = socket_desc.Create(kSnapuserdSocketLabel);
+    if (!socket.ok()) {
+        LOG(FATAL) << "Could not create snapuserd socket: " << socket.error();
+    }
+
+    pid_t pid = fork();
+    if (pid < 0) {
+        PLOG(FATAL) << "Cannot launch snapuserd; fork failed";
+    }
+    if (pid == 0) {
+        socket->Publish();
+        char arg0[] = "/system/bin/snapuserd";
+        char* const argv[] = {arg0, nullptr};
+        if (execv(arg0, argv) < 0) {
+            PLOG(FATAL) << "Cannot launch snapuserd; execv failed";
+        }
+        _exit(127);
+    }
+
+    setenv(kSnapuserdFirstStagePidVar, std::to_string(pid).c_str(), 1);
+
+    LOG(INFO) << "Relaunched snapuserd with pid: " << pid;
+}
+
+std::optional<pid_t> GetSnapuserdFirstStagePid() {
+    const char* pid_str = getenv(kSnapuserdFirstStagePidVar);
+    if (!pid_str) {
+        return {};
+    }
+
+    int pid = 0;
+    if (!android::base::ParseInt(pid_str, &pid)) {
+        LOG(FATAL) << "Could not parse pid in environment, " << kSnapuserdFirstStagePidVar << "="
+                   << pid_str;
+    }
+    return {pid};
+}
+
+static void RelabelLink(const std::string& link) {
+    selinux_android_restorecon(link.c_str(), 0);
+
+    std::string path;
+    if (android::base::Readlink(link, &path)) {
+        selinux_android_restorecon(path.c_str(), 0);
+    }
+}
+
+static void RelabelDeviceMapper() {
+    selinux_android_restorecon("/dev/device-mapper", 0);
+
+    std::error_code ec;
+    for (auto& iter : std::filesystem::directory_iterator("/dev/block", ec)) {
+        const auto& path = iter.path();
+        if (android::base::StartsWith(path.string(), "/dev/block/dm-")) {
+            selinux_android_restorecon(path.string().c_str(), 0);
+        }
+    }
+}
+
+static std::optional<int> GetRamdiskSnapuserdFd() {
+    const char* fd_str = getenv(kSnapuserdFirstStageFdVar);
+    if (!fd_str) {
+        return {};
+    }
+
+    int fd;
+    if (!android::base::ParseInt(fd_str, &fd)) {
+        LOG(FATAL) << "Could not parse fd in environment, " << kSnapuserdFirstStageFdVar << "="
+                   << fd_str;
+    }
+    return {fd};
+}
+
+void RestoreconRamdiskSnapuserd(int fd) {
+    if (fsetxattr(fd, XATTR_NAME_SELINUX, kSnapuserdLabel, strlen(kSnapuserdLabel) + 1, 0) < 0) {
+        PLOG(FATAL) << "fsetxattr snapuserd failed";
+    }
+}
+
+SnapuserdSelinuxHelper::SnapuserdSelinuxHelper(std::unique_ptr<SnapshotManager>&& sm, pid_t old_pid)
+    : sm_(std::move(sm)), old_pid_(old_pid) {
+    // Only dm-user device names change during transitions, so the other
+    // devices are expected to be present.
+    sm_->SetUeventRegenCallback([this](const std::string& device) -> bool {
+        if (android::base::StartsWith(device, "/dev/dm-user/")) {
+            return block_dev_init_.InitDmUser(android::base::Basename(device));
+        }
+        return true;
+    });
+}
+
+void SnapuserdSelinuxHelper::StartTransition() {
+    LOG(INFO) << "Starting SELinux transition of snapuserd";
+
+    // The restorecon path reads from /system etc, so make sure any reads have
+    // been cached before proceeding.
+    auto handle = selinux_android_file_context_handle();
+    if (!handle) {
+        LOG(FATAL) << "Could not create SELinux file context handle";
+    }
+    selinux_android_set_sehandle(handle);
+
+    // We cannot access /system after the transition, so make sure init is
+    // pinned in memory.
+    if (mlockall(MCL_CURRENT) < 0) {
+        LOG(FATAL) << "mlockall failed";
+    }
+
+    argv_.emplace_back("snapuserd");
+    argv_.emplace_back("-no_socket");
+    if (!sm_->DetachSnapuserdForSelinux(&argv_)) {
+        LOG(FATAL) << "Could not perform selinux transition";
+    }
+
+    // Make sure the process is gone so we don't have any selinux audits.
+    KillFirstStageSnapuserd(old_pid_);
+}
+
+void SnapuserdSelinuxHelper::FinishTransition() {
+    RelabelLink("/dev/block/by-name/super");
+    RelabelDeviceMapper();
+
+    selinux_android_restorecon("/dev/null", 0);
+    selinux_android_restorecon("/dev/urandom", 0);
+    selinux_android_restorecon("/dev/kmsg", 0);
+    selinux_android_restorecon("/dev/dm-user", SELINUX_ANDROID_RESTORECON_RECURSE);
+
+    RelaunchFirstStageSnapuserd();
+
+    if (munlockall() < 0) {
+        PLOG(ERROR) << "munlockall failed";
+    }
+}
+
+void SnapuserdSelinuxHelper::RelaunchFirstStageSnapuserd() {
+    auto fd = GetRamdiskSnapuserdFd();
+    if (!fd) {
+        LOG(FATAL) << "Environment variable " << kSnapuserdFirstStageFdVar << " was not set!";
+    }
+    unsetenv(kSnapuserdFirstStageFdVar);
+
+    RestoreconRamdiskSnapuserd(fd.value());
+
+    pid_t pid = fork();
+    if (pid < 0) {
+        PLOG(FATAL) << "Fork to relaunch snapuserd failed";
+    }
+    if (pid > 0) {
+        // We don't need the descriptor anymore, and it should be closed to
+        // avoid leaking into subprocesses.
+        close(fd.value());
+
+        setenv(kSnapuserdFirstStagePidVar, std::to_string(pid).c_str(), 1);
+
+        LOG(INFO) << "Relaunched snapuserd with pid: " << pid;
+        return;
+    }
+
+    // Make sure the descriptor is gone after we exec.
+    if (fcntl(fd.value(), F_SETFD, FD_CLOEXEC) < 0) {
+        PLOG(FATAL) << "fcntl FD_CLOEXEC failed for snapuserd fd";
+    }
+
+    std::vector<char*> argv;
+    for (auto& arg : argv_) {
+        argv.emplace_back(arg.data());
+    }
+    argv.emplace_back(nullptr);
+
+    int rv = syscall(SYS_execveat, fd.value(), "", reinterpret_cast<char* const*>(argv.data()),
+                     nullptr, AT_EMPTY_PATH);
+    if (rv < 0) {
+        PLOG(FATAL) << "Failed to execveat() snapuserd";
+    }
+}
+
+std::unique_ptr<SnapuserdSelinuxHelper> SnapuserdSelinuxHelper::CreateIfNeeded() {
+    if (IsRecoveryMode()) {
+        return nullptr;
+    }
+
+    auto old_pid = GetSnapuserdFirstStagePid();
+    if (!old_pid) {
+        return nullptr;
+    }
+
+    auto sm = SnapshotManager::NewForFirstStageMount();
+    if (!sm) {
+        LOG(FATAL) << "Unable to create SnapshotManager";
+    }
+    return std::make_unique<SnapuserdSelinuxHelper>(std::move(sm), old_pid.value());
+}
+
+void KillFirstStageSnapuserd(pid_t pid) {
+    if (kill(pid, SIGTERM) < 0 && errno != ESRCH) {
+        LOG(ERROR) << "Kill snapuserd pid failed: " << pid;
+    } else {
+        LOG(INFO) << "Sent SIGTERM to snapuserd process " << pid;
+    }
+}
+
+void CleanupSnapuserdSocket() {
+    auto socket_path = ANDROID_SOCKET_DIR "/"s + android::snapshot::kSnapuserdSocket;
+    if (access(socket_path.c_str(), F_OK) != 0) {
+        return;
+    }
+
+    // Tell the daemon to stop accepting connections and to gracefully exit
+    // once all outstanding handlers have terminated.
+    if (auto client = SnapuserdClient::Connect(android::snapshot::kSnapuserdSocket, 3s)) {
+        client->DetachSnapuserd();
+    }
+
+    // Unlink the socket so we can create it again in second-stage.
+    if (unlink(socket_path.c_str()) < 0) {
+        PLOG(FATAL) << "unlink " << socket_path << " failed";
+    }
+}
+
+void SaveRamdiskPathToSnapuserd() {
+    int fd = open(kSnapuserdPath, O_PATH);
+    if (fd < 0) {
+        PLOG(FATAL) << "Unable to open snapuserd: " << kSnapuserdPath;
+    }
+
+    auto value = std::to_string(fd);
+    if (setenv(kSnapuserdFirstStageFdVar, value.c_str(), 1) < 0) {
+        PLOG(FATAL) << "setenv failed: " << kSnapuserdFirstStageFdVar << "=" << value;
+    }
+}
+
+bool IsFirstStageSnapuserdRunning() {
+    return GetSnapuserdFirstStagePid().has_value();
+}
+
+}  // namespace init
+}  // namespace android
diff --git a/init/snapuserd_transition.h b/init/snapuserd_transition.h
new file mode 100644
index 0000000..a5ab652
--- /dev/null
+++ b/init/snapuserd_transition.h
@@ -0,0 +1,87 @@
+/*
+ * 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.
+ */
+
+#pragma once
+
+#include <sys/types.h>
+
+#include <optional>
+#include <string>
+#include <vector>
+
+#include <libsnapshot/snapshot.h>
+
+#include "block_dev_initializer.h"
+
+namespace android {
+namespace init {
+
+// Fork and exec a new copy of snapuserd.
+void LaunchFirstStageSnapuserd();
+
+class SnapuserdSelinuxHelper final {
+    using SnapshotManager = android::snapshot::SnapshotManager;
+
+  public:
+    SnapuserdSelinuxHelper(std::unique_ptr<SnapshotManager>&& sm, pid_t old_pid);
+
+    void StartTransition();
+    void FinishTransition();
+
+    // Return a helper for facilitating the selinux transition of snapuserd.
+    // If snapuserd is not in use, null is returned. StartTransition() should
+    // be called after reading policy. FinishTransition() should be called
+    // after loading policy. In between, no reads of /system or other dynamic
+    // partitions are possible.
+    static std::unique_ptr<SnapuserdSelinuxHelper> CreateIfNeeded();
+
+  private:
+    void RelaunchFirstStageSnapuserd();
+    void ExecSnapuserd();
+
+    std::unique_ptr<SnapshotManager> sm_;
+    BlockDevInitializer block_dev_init_;
+    pid_t old_pid_;
+    std::vector<std::string> argv_;
+};
+
+// Remove /dev/socket/snapuserd. This ensures that (1) the existing snapuserd
+// will receive no new requests, and (2) the next copy we transition to can
+// own the socket.
+void CleanupSnapuserdSocket();
+
+// Kill an instance of snapuserd given a pid.
+void KillFirstStageSnapuserd(pid_t pid);
+
+// Save an open fd to /system/bin (in the ramdisk) into an environment. This is
+// used to later execveat() snapuserd.
+void SaveRamdiskPathToSnapuserd();
+
+// Returns true if first-stage snapuserd is running.
+bool IsFirstStageSnapuserdRunning();
+
+// Return the pid of the first-stage instances of snapuserd, if it was started.
+std::optional<pid_t> GetSnapuserdFirstStagePid();
+
+// Save an open fd to /system/bin (in the ramdisk) into an environment. This is
+// used to later execveat() snapuserd.
+void SaveRamdiskPathToSnapuserd();
+
+// Returns true if first-stage snapuserd is running.
+bool IsFirstStageSnapuserdRunning();
+
+}  // namespace init
+}  // namespace android
diff --git a/libprocessgroup/profiles/cgroups.json b/libprocessgroup/profiles/cgroups.json
index 792af6f..5b7a28a 100644
--- a/libprocessgroup/profiles/cgroups.json
+++ b/libprocessgroup/profiles/cgroups.json
@@ -42,7 +42,7 @@
     "Controllers": [
       {
         "Controller": "freezer",
-        "Path": ".",
+        "Path": "freezer",
         "Mode": "0755",
         "UID": "system",
         "GID": "system"
diff --git a/libprocessgroup/profiles/task_profiles.json b/libprocessgroup/profiles/task_profiles.json
index 628098b..5b57bdd 100644
--- a/libprocessgroup/profiles/task_profiles.json
+++ b/libprocessgroup/profiles/task_profiles.json
@@ -46,7 +46,7 @@
       "File": "cpu.uclamp.latency_sensitive"
     },
     {
-      "Name": "Freezer",
+      "Name": "FreezerState",
       "Controller": "freezer",
       "File": "cgroup.freeze"
     }
@@ -70,11 +70,11 @@
       "Name": "Frozen",
       "Actions": [
         {
-          "Name": "SetAttribute",
+          "Name": "JoinCgroup",
           "Params":
           {
-            "Name": "Freezer",
-            "Value": "1"
+            "Controller": "freezer",
+            "Path": ""
           }
         }
       ]
@@ -83,11 +83,11 @@
       "Name": "Unfrozen",
       "Actions": [
         {
-          "Name": "SetAttribute",
+          "Name": "JoinCgroup",
           "Params":
           {
-            "Name": "Freezer",
-            "Value": "0"
+            "Controller": "freezer",
+            "Path": "../"
           }
         }
       ]
diff --git a/libsparse/Android.bp b/libsparse/Android.bp
index 860b9ae..5b6ce41 100644
--- a/libsparse/Android.bp
+++ b/libsparse/Android.bp
@@ -28,6 +28,10 @@
             enabled: true,
         },
     },
+    apex_available: [
+        "//apex_available:platform",
+        "com.android.virt",
+    ],
 }
 
 cc_binary {
diff --git a/rootdir/Android.mk b/rootdir/Android.mk
index 17d71f7..73d1101 100644
--- a/rootdir/Android.mk
+++ b/rootdir/Android.mk
@@ -69,7 +69,7 @@
 
 EXPORT_GLOBAL_CLANG_COVERAGE_OPTIONS :=
 ifeq ($(CLANG_COVERAGE),true)
-  EXPORT_GLOBAL_CLANG_COVERAGE_OPTIONS := export LLVM_PROFILE_FILE /data/misc/trace/clang-%p-%m.profraw
+  EXPORT_GLOBAL_CLANG_COVERAGE_OPTIONS := export LLVM_PROFILE_FILE /data/misc/trace/clang-%20m.profraw
 endif
 
 # Put it here instead of in init.rc module definition,
diff --git a/rootdir/avb/Android.bp b/rootdir/avb/Android.bp
index 85d2786..8c7caf3 100644
--- a/rootdir/avb/Android.bp
+++ b/rootdir/avb/Android.bp
@@ -18,3 +18,10 @@
         "s-gsi.avbpubkey",
     ],
 }
+
+filegroup {
+    name: "qcar-gsi_avbpubkey",
+    srcs: [
+        "qcar-gsi.avbpubkey",
+    ],
+}
diff --git a/rootdir/avb/Android.mk b/rootdir/avb/Android.mk
index c8fc1d6..f71f205 100644
--- a/rootdir/avb/Android.mk
+++ b/rootdir/avb/Android.mk
@@ -80,4 +80,15 @@
 
 include $(BUILD_PREBUILT)
 
+#######################################
+# qcar-gsi.avbpubkey
+include $(CLEAR_VARS)
+
+LOCAL_MODULE := qcar-gsi.avbpubkey
+LOCAL_MODULE_CLASS := ETC
+LOCAL_SRC_FILES := $(LOCAL_MODULE)
+LOCAL_MODULE_PATH := $(my_gsi_avb_keys_path)
+
+include $(BUILD_PREBUILT)
+
 my_gsi_avb_keys_path :=
diff --git a/rootdir/avb/qcar-gsi.avbpubkey b/rootdir/avb/qcar-gsi.avbpubkey
new file mode 100644
index 0000000..ce56646
--- /dev/null
+++ b/rootdir/avb/qcar-gsi.avbpubkey
Binary files differ
diff --git a/rootdir/init.rc b/rootdir/init.rc
index d033afc..4478bea 100644
--- a/rootdir/init.rc
+++ b/rootdir/init.rc
@@ -592,6 +592,7 @@
     mkdir /metadata/ota 0700 root system
     mkdir /metadata/ota/snapshots 0700 root system
     mkdir /metadata/userspacereboot 0770 root system
+    mkdir /metadata/watchdog 0770 root system
 
     mkdir /metadata/apex 0700 root system
     mkdir /metadata/apex/sessions 0700 root system
@@ -635,20 +636,21 @@
     mkdir /data/bootchart 0755 shell shell encryption=Require
     bootchart start
 
-    # Make sure that apexd is started in the default namespace
-    enter_default_mount_ns
-
     mkdir /data/vendor 0771 root root encryption=Require
     mkdir /data/vendor_ce 0771 root root encryption=None
     mkdir /data/vendor_de 0771 root root encryption=None
     mkdir /data/vendor/hardware 0771 root root
 
     # Start tombstoned early to be able to store tombstones.
+    mkdir /data/anr 0775 system system encryption=Require
     mkdir /data/tombstones 0771 system system encryption=Require
     mkdir /data/vendor/tombstones 0771 root root
     mkdir /data/vendor/tombstones/wifi 0771 wifi wifi
     start tombstoned
 
+    # Make sure that apexd is started in the default namespace
+    enter_default_mount_ns
+
     # /data/apex is now available. Start apexd to scan and activate APEXes.
     mkdir /data/apex 0755 root system encryption=None
     mkdir /data/apex/active 0755 root system
@@ -781,8 +783,6 @@
     # the following directory.
     mkdir /data/mediadrm 0770 mediadrm mediadrm encryption=Require
 
-    mkdir /data/anr 0775 system system encryption=Require
-
     # NFC: create data/nfc for nv storage
     mkdir /data/nfc 0770 nfc nfc encryption=Require
     mkdir /data/nfc/param 0770 nfc nfc
@@ -1072,6 +1072,7 @@
     class_start main
     class_start late_start
     setprop service.bootanim.exit 0
+    setprop service.bootanim.progress 0
     start bootanim
 
 on property:vold.decrypt=trigger_shutdown_framework
@@ -1173,6 +1174,7 @@
   setprop sys.user.0.ce_available ""
   setprop sys.shutdown.requested ""
   setprop service.bootanim.exit ""
+  setprop service.bootanim.progress ""
 
 on userspace-reboot-fs-remount
   # Make sure that vold is running.
diff --git a/trusty/keymaster/3.0/TrustyKeymaster3Device.cpp b/trusty/keymaster/3.0/TrustyKeymaster3Device.cpp
index 98cbcc3..7184e4d 100644
--- a/trusty/keymaster/3.0/TrustyKeymaster3Device.cpp
+++ b/trusty/keymaster/3.0/TrustyKeymaster3Device.cpp
@@ -221,10 +221,10 @@
 
 Return<ErrorCode> TrustyKeymaster3Device::addRngEntropy(const hidl_vec<uint8_t>& data) {
     if (data.size() == 0) return ErrorCode::OK;
-    AddEntropyRequest request;
+    AddEntropyRequest request(impl_->message_version());
     request.random_data.Reinitialize(data.data(), data.size());
 
-    AddEntropyResponse response;
+    AddEntropyResponse response(impl_->message_version());
     impl_->AddRngEntropy(request, &response);
 
     return legacy_enum_conversion(response.error);
@@ -232,10 +232,10 @@
 
 Return<void> TrustyKeymaster3Device::generateKey(const hidl_vec<KeyParameter>& keyParams,
                                                  generateKey_cb _hidl_cb) {
-    GenerateKeyRequest request;
+    GenerateKeyRequest request(impl_->message_version());
     request.key_description.Reinitialize(KmParamSet(keyParams));
 
-    GenerateKeyResponse response;
+    GenerateKeyResponse response(impl_->message_version());
     impl_->GenerateKey(request, &response);
 
     KeyCharacteristics resultCharacteristics;
@@ -253,11 +253,11 @@
                                                            const hidl_vec<uint8_t>& clientId,
                                                            const hidl_vec<uint8_t>& appData,
                                                            getKeyCharacteristics_cb _hidl_cb) {
-    GetKeyCharacteristicsRequest request;
+    GetKeyCharacteristicsRequest request(impl_->message_version());
     request.SetKeyMaterial(keyBlob.data(), keyBlob.size());
     addClientAndAppData(clientId, appData, &request.additional_params);
 
-    GetKeyCharacteristicsResponse response;
+    GetKeyCharacteristicsResponse response(impl_->message_version());
     impl_->GetKeyCharacteristics(request, &response);
 
     KeyCharacteristics resultCharacteristics;
@@ -273,12 +273,12 @@
                                                KeyFormat keyFormat,
                                                const hidl_vec<uint8_t>& keyData,
                                                importKey_cb _hidl_cb) {
-    ImportKeyRequest request;
+    ImportKeyRequest request(impl_->message_version());
     request.key_description.Reinitialize(KmParamSet(params));
     request.key_format = legacy_enum_conversion(keyFormat);
     request.SetKeyMaterial(keyData.data(), keyData.size());
 
-    ImportKeyResponse response;
+    ImportKeyResponse response(impl_->message_version());
     impl_->ImportKey(request, &response);
 
     KeyCharacteristics resultCharacteristics;
@@ -297,12 +297,12 @@
                                                const hidl_vec<uint8_t>& clientId,
                                                const hidl_vec<uint8_t>& appData,
                                                exportKey_cb _hidl_cb) {
-    ExportKeyRequest request;
+    ExportKeyRequest request(impl_->message_version());
     request.key_format = legacy_enum_conversion(exportFormat);
     request.SetKeyMaterial(keyBlob.data(), keyBlob.size());
     addClientAndAppData(clientId, appData, &request.additional_params);
 
-    ExportKeyResponse response;
+    ExportKeyResponse response(impl_->message_version());
     impl_->ExportKey(request, &response);
 
     hidl_vec<uint8_t> resultKeyBlob;
@@ -316,11 +316,11 @@
 Return<void> TrustyKeymaster3Device::attestKey(const hidl_vec<uint8_t>& keyToAttest,
                                                const hidl_vec<KeyParameter>& attestParams,
                                                attestKey_cb _hidl_cb) {
-    AttestKeyRequest request;
+    AttestKeyRequest request(impl_->message_version());
     request.SetKeyMaterial(keyToAttest.data(), keyToAttest.size());
     request.attest_params.Reinitialize(KmParamSet(attestParams));
 
-    AttestKeyResponse response;
+    AttestKeyResponse response(impl_->message_version());
     impl_->AttestKey(request, &response);
 
     hidl_vec<hidl_vec<uint8_t>> resultCertChain;
@@ -334,11 +334,11 @@
 Return<void> TrustyKeymaster3Device::upgradeKey(const hidl_vec<uint8_t>& keyBlobToUpgrade,
                                                 const hidl_vec<KeyParameter>& upgradeParams,
                                                 upgradeKey_cb _hidl_cb) {
-    UpgradeKeyRequest request;
+    UpgradeKeyRequest request(impl_->message_version());
     request.SetKeyMaterial(keyBlobToUpgrade.data(), keyBlobToUpgrade.size());
     request.upgrade_params.Reinitialize(KmParamSet(upgradeParams));
 
-    UpgradeKeyResponse response;
+    UpgradeKeyResponse response(impl_->message_version());
     impl_->UpgradeKey(request, &response);
 
     if (response.error == KM_ERROR_OK) {
@@ -350,18 +350,18 @@
 }
 
 Return<ErrorCode> TrustyKeymaster3Device::deleteKey(const hidl_vec<uint8_t>& keyBlob) {
-    DeleteKeyRequest request;
+    DeleteKeyRequest request(impl_->message_version());
     request.SetKeyMaterial(keyBlob.data(), keyBlob.size());
 
-    DeleteKeyResponse response;
+    DeleteKeyResponse response(impl_->message_version());
     impl_->DeleteKey(request, &response);
 
     return legacy_enum_conversion(response.error);
 }
 
 Return<ErrorCode> TrustyKeymaster3Device::deleteAllKeys() {
-    DeleteAllKeysRequest request;
-    DeleteAllKeysResponse response;
+    DeleteAllKeysRequest request(impl_->message_version());
+    DeleteAllKeysResponse response(impl_->message_version());
     impl_->DeleteAllKeys(request, &response);
 
     return legacy_enum_conversion(response.error);
@@ -374,15 +374,15 @@
 Return<void> TrustyKeymaster3Device::begin(KeyPurpose purpose, const hidl_vec<uint8_t>& key,
                                            const hidl_vec<KeyParameter>& inParams,
                                            begin_cb _hidl_cb) {
-    BeginOperationRequest request;
+    BeginOperationRequest request(impl_->message_version());
     request.purpose = legacy_enum_conversion(purpose);
     request.SetKeyMaterial(key.data(), key.size());
     request.additional_params.Reinitialize(KmParamSet(inParams));
 
-    BeginOperationResponse response;
+    BeginOperationResponse response(impl_->message_version());
     impl_->BeginOperation(request, &response);
 
-    hidl_vec<KeyParameter> resultParams;
+    hidl_vec<KeyParameter> resultParams(impl_->message_version());
     if (response.error == KM_ERROR_OK) {
         resultParams = kmParamSet2Hidl(response.output_params);
     }
@@ -394,8 +394,8 @@
 Return<void> TrustyKeymaster3Device::update(uint64_t operationHandle,
                                             const hidl_vec<KeyParameter>& inParams,
                                             const hidl_vec<uint8_t>& input, update_cb _hidl_cb) {
-    UpdateOperationRequest request;
-    UpdateOperationResponse response;
+    UpdateOperationRequest request(impl_->message_version());
+    UpdateOperationResponse response(impl_->message_version());
     hidl_vec<KeyParameter> resultParams;
     hidl_vec<uint8_t> resultBlob;
     uint32_t resultConsumed = 0;
@@ -431,13 +431,13 @@
                                             const hidl_vec<uint8_t>& input,
                                             const hidl_vec<uint8_t>& signature,
                                             finish_cb _hidl_cb) {
-    FinishOperationRequest request;
+    FinishOperationRequest request(impl_->message_version());
     request.op_handle = operationHandle;
     request.input.Reinitialize(input.data(), input.size());
     request.signature.Reinitialize(signature.data(), signature.size());
     request.additional_params.Reinitialize(KmParamSet(inParams));
 
-    FinishOperationResponse response;
+    FinishOperationResponse response(impl_->message_version());
     impl_->FinishOperation(request, &response);
 
     hidl_vec<KeyParameter> resultParams;
@@ -451,10 +451,10 @@
 }
 
 Return<ErrorCode> TrustyKeymaster3Device::abort(uint64_t operationHandle) {
-    AbortOperationRequest request;
+    AbortOperationRequest request(impl_->message_version());
     request.op_handle = operationHandle;
 
-    AbortOperationResponse response;
+    AbortOperationResponse response(impl_->message_version());
     impl_->AbortOperation(request, &response);
 
     return legacy_enum_conversion(response.error);
diff --git a/trusty/keymaster/4.0/TrustyKeymaster4Device.cpp b/trusty/keymaster/4.0/TrustyKeymaster4Device.cpp
index ec2ba12..73ad6ae 100644
--- a/trusty/keymaster/4.0/TrustyKeymaster4Device.cpp
+++ b/trusty/keymaster/4.0/TrustyKeymaster4Device.cpp
@@ -284,7 +284,7 @@
 
 Return<void> TrustyKeymaster4Device::computeSharedHmac(
         const hidl_vec<HmacSharingParameters>& params, computeSharedHmac_cb _hidl_cb) {
-    ComputeSharedHmacRequest request;
+    ComputeSharedHmacRequest request(impl_->message_version());
     request.params_array.params_array = new keymaster::HmacSharingParameters[params.size()];
     request.params_array.num_params = params.size();
     for (size_t i = 0; i < params.size(); ++i) {
@@ -309,7 +309,7 @@
 Return<void> TrustyKeymaster4Device::verifyAuthorization(
         uint64_t challenge, const hidl_vec<KeyParameter>& parametersToVerify,
         const HardwareAuthToken& authToken, verifyAuthorization_cb _hidl_cb) {
-    VerifyAuthorizationRequest request;
+    VerifyAuthorizationRequest request(impl_->message_version());
     request.challenge = challenge;
     request.parameters_to_verify.Reinitialize(KmParamSet(parametersToVerify));
     request.auth_token.challenge = authToken.challenge;
@@ -336,10 +336,10 @@
 
 Return<ErrorCode> TrustyKeymaster4Device::addRngEntropy(const hidl_vec<uint8_t>& data) {
     if (data.size() == 0) return ErrorCode::OK;
-    AddEntropyRequest request;
+    AddEntropyRequest request(impl_->message_version());
     request.random_data.Reinitialize(data.data(), data.size());
 
-    AddEntropyResponse response;
+    AddEntropyResponse response(impl_->message_version());
     impl_->AddRngEntropy(request, &response);
 
     return legacy_enum_conversion(response.error);
@@ -347,10 +347,10 @@
 
 Return<void> TrustyKeymaster4Device::generateKey(const hidl_vec<KeyParameter>& keyParams,
                                                  generateKey_cb _hidl_cb) {
-    GenerateKeyRequest request;
+    GenerateKeyRequest request(impl_->message_version());
     request.key_description.Reinitialize(KmParamSet(keyParams));
 
-    GenerateKeyResponse response;
+    GenerateKeyResponse response(impl_->message_version());
     impl_->GenerateKey(request, &response);
 
     KeyCharacteristics resultCharacteristics;
@@ -368,11 +368,11 @@
                                                            const hidl_vec<uint8_t>& clientId,
                                                            const hidl_vec<uint8_t>& appData,
                                                            getKeyCharacteristics_cb _hidl_cb) {
-    GetKeyCharacteristicsRequest request;
+    GetKeyCharacteristicsRequest request(impl_->message_version());
     request.SetKeyMaterial(keyBlob.data(), keyBlob.size());
     addClientAndAppData(clientId, appData, &request.additional_params);
 
-    GetKeyCharacteristicsResponse response;
+    GetKeyCharacteristicsResponse response(impl_->message_version());
     impl_->GetKeyCharacteristics(request, &response);
 
     KeyCharacteristics resultCharacteristics;
@@ -388,12 +388,12 @@
                                                KeyFormat keyFormat,
                                                const hidl_vec<uint8_t>& keyData,
                                                importKey_cb _hidl_cb) {
-    ImportKeyRequest request;
+    ImportKeyRequest request(impl_->message_version());
     request.key_description.Reinitialize(KmParamSet(params));
     request.key_format = legacy_enum_conversion(keyFormat);
     request.SetKeyMaterial(keyData.data(), keyData.size());
 
-    ImportKeyResponse response;
+    ImportKeyResponse response(impl_->message_version());
     impl_->ImportKey(request, &response);
 
     KeyCharacteristics resultCharacteristics;
@@ -411,7 +411,7 @@
         const hidl_vec<uint8_t>& wrappedKeyData, const hidl_vec<uint8_t>& wrappingKeyBlob,
         const hidl_vec<uint8_t>& maskingKey, const hidl_vec<KeyParameter>& unwrappingParams,
         uint64_t passwordSid, uint64_t biometricSid, importWrappedKey_cb _hidl_cb) {
-    ImportWrappedKeyRequest request;
+    ImportWrappedKeyRequest request(impl_->message_version());
     request.SetWrappedMaterial(wrappedKeyData.data(), wrappedKeyData.size());
     request.SetWrappingMaterial(wrappingKeyBlob.data(), wrappingKeyBlob.size());
     request.SetMaskingKeyMaterial(maskingKey.data(), maskingKey.size());
@@ -419,7 +419,7 @@
     request.password_sid = passwordSid;
     request.biometric_sid = biometricSid;
 
-    ImportWrappedKeyResponse response;
+    ImportWrappedKeyResponse response(impl_->message_version());
     impl_->ImportWrappedKey(request, &response);
 
     KeyCharacteristics resultCharacteristics;
@@ -438,12 +438,12 @@
                                                const hidl_vec<uint8_t>& clientId,
                                                const hidl_vec<uint8_t>& appData,
                                                exportKey_cb _hidl_cb) {
-    ExportKeyRequest request;
+    ExportKeyRequest request(impl_->message_version());
     request.key_format = legacy_enum_conversion(exportFormat);
     request.SetKeyMaterial(keyBlob.data(), keyBlob.size());
     addClientAndAppData(clientId, appData, &request.additional_params);
 
-    ExportKeyResponse response;
+    ExportKeyResponse response(impl_->message_version());
     impl_->ExportKey(request, &response);
 
     hidl_vec<uint8_t> resultKeyBlob;
@@ -457,11 +457,11 @@
 Return<void> TrustyKeymaster4Device::attestKey(const hidl_vec<uint8_t>& keyToAttest,
                                                const hidl_vec<KeyParameter>& attestParams,
                                                attestKey_cb _hidl_cb) {
-    AttestKeyRequest request;
+    AttestKeyRequest request(impl_->message_version());
     request.SetKeyMaterial(keyToAttest.data(), keyToAttest.size());
     request.attest_params.Reinitialize(KmParamSet(attestParams));
 
-    AttestKeyResponse response;
+    AttestKeyResponse response(impl_->message_version());
     impl_->AttestKey(request, &response);
 
     hidl_vec<hidl_vec<uint8_t>> resultCertChain;
@@ -475,11 +475,11 @@
 Return<void> TrustyKeymaster4Device::upgradeKey(const hidl_vec<uint8_t>& keyBlobToUpgrade,
                                                 const hidl_vec<KeyParameter>& upgradeParams,
                                                 upgradeKey_cb _hidl_cb) {
-    UpgradeKeyRequest request;
+    UpgradeKeyRequest request(impl_->message_version());
     request.SetKeyMaterial(keyBlobToUpgrade.data(), keyBlobToUpgrade.size());
     request.upgrade_params.Reinitialize(KmParamSet(upgradeParams));
 
-    UpgradeKeyResponse response;
+    UpgradeKeyResponse response(impl_->message_version());
     impl_->UpgradeKey(request, &response);
 
     if (response.error == KM_ERROR_OK) {
@@ -491,18 +491,18 @@
 }
 
 Return<ErrorCode> TrustyKeymaster4Device::deleteKey(const hidl_vec<uint8_t>& keyBlob) {
-    DeleteKeyRequest request;
+    DeleteKeyRequest request(impl_->message_version());
     request.SetKeyMaterial(keyBlob.data(), keyBlob.size());
 
-    DeleteKeyResponse response;
+    DeleteKeyResponse response(impl_->message_version());
     impl_->DeleteKey(request, &response);
 
     return legacy_enum_conversion(response.error);
 }
 
 Return<ErrorCode> TrustyKeymaster4Device::deleteAllKeys() {
-    DeleteAllKeysRequest request;
-    DeleteAllKeysResponse response;
+    DeleteAllKeysRequest request(impl_->message_version());
+    DeleteAllKeysResponse response(impl_->message_version());
     impl_->DeleteAllKeys(request, &response);
 
     return legacy_enum_conversion(response.error);
@@ -516,12 +516,12 @@
                                            const hidl_vec<KeyParameter>& inParams,
                                            const HardwareAuthToken& authToken, begin_cb _hidl_cb) {
     hidl_vec<KeyParameter> extendedParams = injectAuthToken(inParams, authToken);
-    BeginOperationRequest request;
+    BeginOperationRequest request(impl_->message_version());
     request.purpose = legacy_enum_conversion(purpose);
     request.SetKeyMaterial(key.data(), key.size());
     request.additional_params.Reinitialize(KmParamSet(extendedParams));
 
-    BeginOperationResponse response;
+    BeginOperationResponse response(impl_->message_version());
     impl_->BeginOperation(request, &response);
 
     hidl_vec<KeyParameter> resultParams;
@@ -540,8 +540,8 @@
                                             const VerificationToken& verificationToken,
                                             update_cb _hidl_cb) {
     (void)verificationToken;
-    UpdateOperationRequest request;
-    UpdateOperationResponse response;
+    UpdateOperationRequest request(impl_->message_version());
+    UpdateOperationResponse response(impl_->message_version());
     hidl_vec<KeyParameter> resultParams;
     hidl_vec<uint8_t> resultBlob;
     hidl_vec<KeyParameter> extendedParams = injectAuthToken(inParams, authToken);
@@ -581,14 +581,14 @@
                                             const VerificationToken& verificationToken,
                                             finish_cb _hidl_cb) {
     (void)verificationToken;
-    FinishOperationRequest request;
+    FinishOperationRequest request(impl_->message_version());
     hidl_vec<KeyParameter> extendedParams = injectAuthToken(inParams, authToken);
     request.op_handle = operationHandle;
     request.input.Reinitialize(input.data(), input.size());
     request.signature.Reinitialize(signature.data(), signature.size());
     request.additional_params.Reinitialize(KmParamSet(extendedParams));
 
-    FinishOperationResponse response;
+    FinishOperationResponse response(impl_->message_version());
     impl_->FinishOperation(request, &response);
 
     hidl_vec<KeyParameter> resultParams;
@@ -602,10 +602,10 @@
 }
 
 Return<ErrorCode> TrustyKeymaster4Device::abort(uint64_t operationHandle) {
-    AbortOperationRequest request;
+    AbortOperationRequest request(impl_->message_version());
     request.op_handle = operationHandle;
 
-    AbortOperationResponse response;
+    AbortOperationResponse response(impl_->message_version());
     impl_->AbortOperation(request, &response);
 
     return legacy_enum_conversion(response.error);
diff --git a/trusty/keymaster/TrustyKeymaster.cpp b/trusty/keymaster/TrustyKeymaster.cpp
index 750a9d7..5690031 100644
--- a/trusty/keymaster/TrustyKeymaster.cpp
+++ b/trusty/keymaster/TrustyKeymaster.cpp
@@ -31,11 +31,35 @@
         return err;
     }
 
-    ConfigureRequest req;
+    // Try GetVersion2 first.
+    GetVersion2Request versionReq;
+    GetVersion2Response versionRsp = GetVersion2(versionReq);
+    if (versionRsp.error != KM_ERROR_OK) {
+        ALOGW("TA appears not to support GetVersion2, falling back (err = %d)", versionRsp.error);
+
+        GetVersionRequest versionReq;
+        GetVersionResponse versionRsp;
+        GetVersion(versionReq, &versionRsp);
+        if (versionRsp.error != KM_ERROR_OK) {
+            ALOGE("Failed to get TA version %d", versionRsp.error);
+            return -1;
+        } else {
+            keymaster_error_t error;
+            message_version_ = NegotiateMessageVersion(versionRsp, &error);
+            if (error != KM_ERROR_OK) {
+                ALOGE("Failed to negotiate message version %d", error);
+                return -1;
+            }
+        }
+    } else {
+        message_version_ = NegotiateMessageVersion(versionReq, versionRsp);
+    }
+
+    ConfigureRequest req(message_version());
     req.os_version = GetOsVersion();
     req.os_patchlevel = GetOsPatchlevel();
 
-    ConfigureResponse rsp;
+    ConfigureResponse rsp(message_version());
     Configure(req, &rsp);
 
     if (rsp.error != KM_ERROR_OK) {
@@ -52,7 +76,7 @@
     trusty_keymaster_disconnect();
 }
 
-static void ForwardCommand(enum keymaster_command command, const Serializable& req,
+static void ForwardCommand(enum keymaster_command command, const KeymasterMessage& req,
                            KeymasterResponse* rsp) {
     keymaster_error_t err;
     err = trusty_keymaster_send(command, req, rsp);
@@ -173,25 +197,30 @@
 }
 
 GetHmacSharingParametersResponse TrustyKeymaster::GetHmacSharingParameters() {
-    // Empty buffer to allow ForwardCommand to have something to serialize
-    Buffer request;
-    GetHmacSharingParametersResponse response;
+    GetHmacSharingParametersRequest request(message_version());
+    GetHmacSharingParametersResponse response(message_version());
     ForwardCommand(KM_GET_HMAC_SHARING_PARAMETERS, request, &response);
     return response;
 }
 
 ComputeSharedHmacResponse TrustyKeymaster::ComputeSharedHmac(
         const ComputeSharedHmacRequest& request) {
-    ComputeSharedHmacResponse response;
+    ComputeSharedHmacResponse response(message_version());
     ForwardCommand(KM_COMPUTE_SHARED_HMAC, request, &response);
     return response;
 }
 
 VerifyAuthorizationResponse TrustyKeymaster::VerifyAuthorization(
         const VerifyAuthorizationRequest& request) {
-    VerifyAuthorizationResponse response;
+    VerifyAuthorizationResponse response(message_version());
     ForwardCommand(KM_VERIFY_AUTHORIZATION, request, &response);
     return response;
 }
 
+GetVersion2Response TrustyKeymaster::GetVersion2(const GetVersion2Request& request) {
+    GetVersion2Response response(message_version());
+    ForwardCommand(KM_GET_VERSION_2, request, &response);
+    return response;
+}
+
 }  // namespace keymaster
diff --git a/trusty/keymaster/include/trusty_keymaster/TrustyKeymaster.h b/trusty/keymaster/include/trusty_keymaster/TrustyKeymaster.h
index 030b645..bec2a2a 100644
--- a/trusty/keymaster/include/trusty_keymaster/TrustyKeymaster.h
+++ b/trusty/keymaster/include/trusty_keymaster/TrustyKeymaster.h
@@ -59,6 +59,12 @@
     GetHmacSharingParametersResponse GetHmacSharingParameters();
     ComputeSharedHmacResponse ComputeSharedHmac(const ComputeSharedHmacRequest& request);
     VerifyAuthorizationResponse VerifyAuthorization(const VerifyAuthorizationRequest& request);
+    GetVersion2Response GetVersion2(const GetVersion2Request& request);
+
+    uint32_t message_version() const { return message_version_; }
+
+  private:
+    uint32_t message_version_;
 };
 
 }  // namespace keymaster
diff --git a/trusty/keymaster/include/trusty_keymaster/ipc/keymaster_ipc.h b/trusty/keymaster/include/trusty_keymaster/ipc/keymaster_ipc.h
index ce2cc2e..419c96f 100644
--- a/trusty/keymaster/include/trusty_keymaster/ipc/keymaster_ipc.h
+++ b/trusty/keymaster/include/trusty_keymaster/ipc/keymaster_ipc.h
@@ -53,6 +53,7 @@
     KM_DELETE_ALL_KEYS              = (23 << KEYMASTER_REQ_SHIFT),
     KM_DESTROY_ATTESTATION_IDS      = (24 << KEYMASTER_REQ_SHIFT),
     KM_IMPORT_WRAPPED_KEY           = (25 << KEYMASTER_REQ_SHIFT),
+    KM_GET_VERSION_2                = (28 << KEYMASTER_REQ_SHIFT),
 
     // Bootloader/provisioning calls.
     KM_SET_BOOT_PARAMS = (0x1000 << KEYMASTER_REQ_SHIFT),
diff --git a/trusty/libtrusty/tipc-test/Android.bp b/trusty/libtrusty/tipc-test/Android.bp
index 9676b79..5e60d28 100644
--- a/trusty/libtrusty/tipc-test/Android.bp
+++ b/trusty/libtrusty/tipc-test/Android.bp
@@ -19,6 +19,7 @@
     srcs: ["tipc_test.c"],
     shared_libs: [
         "libc",
+        "libdmabufheap",
         "liblog",
         "libtrusty",
     ],
diff --git a/trusty/libtrusty/tipc-test/tipc_test.c b/trusty/libtrusty/tipc-test/tipc_test.c
index ca581dc..94aedd7 100644
--- a/trusty/libtrusty/tipc-test/tipc_test.c
+++ b/trusty/libtrusty/tipc-test/tipc_test.c
@@ -25,6 +25,8 @@
 #include <sys/mman.h>
 #include <sys/uio.h>
 
+#include <BufferAllocator/BufferAllocatorWrapper.h>
+
 #include <trusty/tipc.h>
 
 #define TIPC_DEFAULT_DEVNAME "/dev/trusty-ipc-dev0"
@@ -86,7 +88,7 @@
         "   ta-access    - test ta-access flags\n"
         "   writev       - writev test\n"
         "   readv        - readv test\n"
-        "   send-fd      - transmit memfd to trusty, use as shm\n"
+        "   send-fd      - transmit dma_buf to trusty, use as shm\n"
         "\n";
 
 static uint opt_repeat  = 1;
@@ -890,9 +892,12 @@
 
 static int send_fd_test(void) {
     int ret;
-    int memfd = -1;
+    int dma_buf = -1;
     int fd = -1;
     volatile char* buf = MAP_FAILED;
+    BufferAllocator* allocator = NULL;
+
+    const size_t num_pages = 10;
 
     fd = tipc_connect(dev_name, receiver_name);
     if (fd < 0) {
@@ -901,22 +906,24 @@
         goto cleanup;
     }
 
-    memfd = memfd_create("tipc-send-fd", 0);
-    if (memfd < 0) {
-        fprintf(stderr, "Failed to create memfd: %s\n", strerror(errno));
+    allocator = CreateDmabufHeapBufferAllocator();
+    if (!allocator) {
+        fprintf(stderr, "Failed to create dma-buf allocator.\n");
         ret = -1;
         goto cleanup;
     }
 
-    if (ftruncate(memfd, PAGE_SIZE) < 0) {
-        fprintf(stderr, "Failed to resize memfd: %s\n", strerror(errno));
-        ret = -1;
+    size_t buf_size = PAGE_SIZE * num_pages;
+    dma_buf = DmabufHeapAlloc(allocator, "system", buf_size, 0);
+    if (dma_buf < 0) {
+        ret = dma_buf;
+        fprintf(stderr, "Failed to create dma-buf fd of size %zu err (%d)\n", buf_size, ret);
         goto cleanup;
     }
 
-    buf = mmap(0, PAGE_SIZE, PROT_READ | PROT_WRITE, MAP_SHARED, memfd, 0);
+    buf = mmap(0, buf_size, PROT_READ | PROT_WRITE, MAP_SHARED, dma_buf, 0);
     if (buf == MAP_FAILED) {
-        fprintf(stderr, "Failed to map memfd: %s\n", strerror(errno));
+        fprintf(stderr, "Failed to map dma-buf: %s\n", strerror(errno));
         ret = -1;
         goto cleanup;
     }
@@ -924,13 +931,13 @@
     strcpy((char*)buf, "From NS");
 
     struct trusty_shm shm = {
-            .fd = memfd,
+            .fd = dma_buf,
             .transfer = TRUSTY_SHARE,
     };
 
     ssize_t rc = tipc_send(fd, NULL, 0, &shm, 1);
     if (rc < 0) {
-        fprintf(stderr, "tipc_send failed\n");
+        fprintf(stderr, "tipc_send failed: %zd\n", rc);
         ret = rc;
         goto cleanup;
     }
@@ -938,13 +945,19 @@
     read(fd, &c, 1);
     tipc_close(fd);
 
-    ret = strcmp("Hello from Trusty!", (const char*)buf) ? (-1) : 0;
+    ret = 0;
+    for (size_t skip = 0; skip < num_pages; skip++) {
+        ret |= strcmp("Hello from Trusty!", (const char*)&buf[skip * PAGE_SIZE]) ? (-1) : 0;
+    }
 
 cleanup:
     if (buf != MAP_FAILED) {
         munmap((char*)buf, PAGE_SIZE);
     }
-    close(memfd);
+    close(dma_buf);
+    if (allocator) {
+        FreeDmabufHeapBufferAllocator(allocator);
+    }
     tipc_close(fd);
     return ret;
 }