diff --git a/fastboot/device/flashing.cpp b/fastboot/device/flashing.cpp
index ee0aa58..9b5d2cd 100644
--- a/fastboot/device/flashing.cpp
+++ b/fastboot/device/flashing.cpp
@@ -187,11 +187,17 @@
                                  ", build may be missing broken or missing boot_devices");
     }
 
+    std::string slot_suffix = device->GetCurrentSlot();
+    uint32_t slot_number = SlotNumberForSlotSuffix(slot_suffix);
+
+    std::string other_slot_suffix;
+    if (!slot_suffix.empty()) {
+        other_slot_suffix = (slot_suffix == "_a") ? "_b" : "_a";
+    }
+
     // If we are unable to read the existing metadata, then the super partition
     // is corrupt. In this case we reflash the whole thing using the provided
     // image.
-    std::string slot_suffix = device->GetCurrentSlot();
-    uint32_t slot_number = SlotNumberForSlotSuffix(slot_suffix);
     std::unique_ptr<LpMetadata> old_metadata = ReadMetadata(super_name, slot_number);
     if (wipe || !old_metadata) {
         if (!FlashPartitionTable(super_name, *new_metadata.get())) {
@@ -203,11 +209,15 @@
     }
 
     std::set<std::string> partitions_to_keep;
+    bool virtual_ab = android::base::GetBoolProperty("ro.virtual_ab.enabled", false);
     for (const auto& partition : old_metadata->partitions) {
         // Preserve partitions in the other slot, but not the current slot.
         std::string partition_name = GetPartitionName(partition);
-        if (!slot_suffix.empty() && GetPartitionSlotSuffix(partition_name) == slot_suffix) {
-            continue;
+        if (!slot_suffix.empty()) {
+            auto part_suffix = GetPartitionSlotSuffix(partition_name);
+            if (part_suffix == slot_suffix || (part_suffix == other_slot_suffix && virtual_ab)) {
+                continue;
+            }
         }
         std::string group_name = GetPartitionGroupName(old_metadata->groups[partition.group_index]);
         // Skip partitions in the COW group
diff --git a/fs_mgr/Android.bp b/fs_mgr/Android.bp
index 5356b00..3c83aab 100644
--- a/fs_mgr/Android.bp
+++ b/fs_mgr/Android.bp
@@ -63,6 +63,7 @@
         "-D_FILE_OFFSET_BITS=64",
     ],
     srcs: [
+        "blockdev.cpp",
         "file_wait.cpp",
         "fs_mgr.cpp",
         "fs_mgr_format.cpp",
diff --git a/fs_mgr/blockdev.cpp b/fs_mgr/blockdev.cpp
new file mode 100644
index 0000000..14b217c
--- /dev/null
+++ b/fs_mgr/blockdev.cpp
@@ -0,0 +1,159 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <android-base/file.h>
+#include <android-base/logging.h>
+#include <android-base/stringprintf.h>
+#include <android-base/strings.h>
+
+#include <dirent.h>
+#include <libdm/dm.h>
+#include <sys/stat.h>
+#include <sys/sysmacros.h>
+#include <sys/types.h>
+#include "blockdev.h"
+
+using android::base::Basename;
+using android::base::ErrnoError;
+using android::base::Error;
+using android::base::Result;
+using android::base::ResultError;
+using android::base::StartsWith;
+using android::base::StringPrintf;
+using android::base::unique_fd;
+using android::dm::DeviceMapper;
+
+// Return the parent device of a partition. Converts e.g. "sda26" into "sda".
+static std::string PartitionParent(const std::string& blockdev) {
+    if (blockdev.find('/') != std::string::npos) {
+        LOG(ERROR) << __func__ << ": invalid argument " << blockdev;
+        return blockdev;
+    }
+    auto dir = std::unique_ptr<DIR, decltype(&closedir)>{opendir("/sys/class/block"), closedir};
+    if (!dir) {
+        return blockdev;
+    }
+    for (struct dirent* ent = readdir(dir.get()); ent; ent = readdir(dir.get())) {
+        if (ent->d_name[0] == '.') {
+            continue;
+        }
+        std::string path = StringPrintf("/sys/class/block/%s/%s", ent->d_name, blockdev.c_str());
+        struct stat statbuf;
+        if (stat(path.c_str(), &statbuf) >= 0) {
+            return ent->d_name;
+        }
+    }
+    return blockdev;
+}
+
+// Convert a major:minor pair into a block device name.
+static std::string BlockdevName(dev_t dev) {
+    auto dir = std::unique_ptr<DIR, decltype(&closedir)>{opendir("/dev/block"), closedir};
+    if (!dir) {
+        return {};
+    }
+    for (struct dirent* ent = readdir(dir.get()); ent; ent = readdir(dir.get())) {
+        if (ent->d_name[0] == '.') {
+            continue;
+        }
+        const std::string path = std::string("/dev/block/") + ent->d_name;
+        struct stat statbuf;
+        if (stat(path.c_str(), &statbuf) >= 0 && dev == statbuf.st_rdev) {
+            return ent->d_name;
+        }
+    }
+    return {};
+}
+
+// Trim whitespace from the end of a string.
+static void rtrim(std::string& s) {
+    s.erase(s.find_last_not_of('\n') + 1, s.length());
+}
+
+// For file `file_path`, retrieve the block device backing the filesystem on
+// which the file exists and return the queue depth of the block device.
+static Result<uint32_t> BlockDeviceQueueDepth(const std::string& file_path) {
+    struct stat statbuf;
+    int res = stat(file_path.c_str(), &statbuf);
+    if (res < 0) {
+        return ErrnoError() << "stat(" << file_path << ")";
+    }
+    std::string blockdev = "/dev/block/" + BlockdevName(statbuf.st_dev);
+    LOG(DEBUG) << __func__ << ": " << file_path << " -> " << blockdev;
+    if (blockdev.empty()) {
+        const std::string err_msg =
+                StringPrintf("Failed to convert %u:%u (path %s)", major(statbuf.st_dev),
+                             minor(statbuf.st_dev), file_path.c_str());
+        return ResultError(err_msg, 0);
+    }
+    auto& dm = DeviceMapper::Instance();
+    for (;;) {
+        std::optional<std::string> child = dm.GetParentBlockDeviceByPath(blockdev);
+        if (!child) {
+            break;
+        }
+        LOG(DEBUG) << __func__ << ": " << blockdev << " -> " << *child;
+        blockdev = *child;
+    }
+    std::optional<std::string> maybe_blockdev = android::dm::ExtractBlockDeviceName(blockdev);
+    if (!maybe_blockdev) {
+        return ResultError("Failed to remove /dev/block/ prefix from " + blockdev, 0);
+    }
+    blockdev = PartitionParent(*maybe_blockdev);
+    LOG(DEBUG) << __func__ << ": "
+               << "Partition parent: " << blockdev;
+    const std::string nr_tags_path =
+            StringPrintf("/sys/class/block/%s/mq/0/nr_tags", blockdev.c_str());
+    std::string nr_tags;
+    if (!android::base::ReadFileToString(nr_tags_path, &nr_tags)) {
+        return ResultError("Failed to read " + nr_tags_path, 0);
+    }
+    rtrim(nr_tags);
+    LOG(DEBUG) << __func__ << ": " << file_path << " is backed by /dev/" << blockdev
+               << " and that block device supports queue depth " << nr_tags;
+    return strtol(nr_tags.c_str(), NULL, 0);
+}
+
+// Set 'nr_requests' of `loop_device_path` to the queue depth of the block
+// device backing `file_path`.
+Result<void> ConfigureQueueDepth(const std::string& loop_device_path,
+                                 const std::string& file_path) {
+    if (!StartsWith(loop_device_path, "/dev/")) {
+        return Error() << "Invalid argument " << loop_device_path;
+    }
+
+    const std::string loop_device_name = Basename(loop_device_path);
+
+    const Result<uint32_t> qd = BlockDeviceQueueDepth(file_path);
+    if (!qd.ok()) {
+        LOG(DEBUG) << __func__ << ": "
+                   << "BlockDeviceQueueDepth() returned " << qd.error();
+        return ResultError(qd.error());
+    }
+    const std::string nr_requests = StringPrintf("%u", *qd);
+    const std::string sysfs_path =
+            StringPrintf("/sys/class/block/%s/queue/nr_requests", loop_device_name.c_str());
+    unique_fd sysfs_fd(open(sysfs_path.c_str(), O_RDWR | O_CLOEXEC));
+    if (sysfs_fd == -1) {
+        return ErrnoError() << "Failed to open " << sysfs_path;
+    }
+
+    const int res = write(sysfs_fd.get(), nr_requests.data(), nr_requests.length());
+    if (res < 0) {
+        return ErrnoError() << "Failed to write to " << sysfs_path;
+    }
+    return {};
+}
diff --git a/fs_mgr/blockdev.h b/fs_mgr/blockdev.h
new file mode 100644
index 0000000..2c0d68a
--- /dev/null
+++ b/fs_mgr/blockdev.h
@@ -0,0 +1,21 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include <android-base/result.h>
+#include <string>
+
+android::base::Result<void> ConfigureQueueDepth(const std::string& loop_device_path,
+                                                const std::string& file_path);
diff --git a/fs_mgr/fs_mgr.cpp b/fs_mgr/fs_mgr.cpp
index af71fe6..21df8af 100644
--- a/fs_mgr/fs_mgr.cpp
+++ b/fs_mgr/fs_mgr.cpp
@@ -70,6 +70,7 @@
 #include <log/log_properties.h>
 #include <logwrap/logwrap.h>
 
+#include "blockdev.h"
 #include "fs_mgr_priv.h"
 
 #define KEY_LOC_PROP   "ro.crypto.keyfile.userdata"
@@ -2065,22 +2066,24 @@
 
     // Allocate loop device and attach it to file_path.
     LoopControl loop_control;
-    std::string device;
-    if (!loop_control.Attach(target_fd.get(), 5s, &device)) {
+    std::string loop_device;
+    if (!loop_control.Attach(target_fd.get(), 5s, &loop_device)) {
         return false;
     }
 
+    ConfigureQueueDepth(loop_device, "/");
+
     // set block size & direct IO
-    unique_fd device_fd(TEMP_FAILURE_RETRY(open(device.c_str(), O_RDWR | O_CLOEXEC)));
-    if (device_fd.get() == -1) {
-        PERROR << "Cannot open " << device;
+    unique_fd loop_fd(TEMP_FAILURE_RETRY(open(loop_device.c_str(), O_RDWR | O_CLOEXEC)));
+    if (loop_fd.get() == -1) {
+        PERROR << "Cannot open " << loop_device;
         return false;
     }
-    if (!LoopControl::EnableDirectIo(device_fd.get())) {
+    if (!LoopControl::EnableDirectIo(loop_fd.get())) {
         return false;
     }
 
-    return InstallZramDevice(device);
+    return InstallZramDevice(loop_device);
 }
 
 bool fs_mgr_swapon_all(const Fstab& fstab) {
diff --git a/fs_mgr/fs_mgr_fstab.cpp b/fs_mgr/fs_mgr_fstab.cpp
index d0c89b9..f5ab557 100644
--- a/fs_mgr/fs_mgr_fstab.cpp
+++ b/fs_mgr/fs_mgr_fstab.cpp
@@ -413,17 +413,24 @@
     return fstab_result;
 }
 
-// Identify path to fstab file. Lookup is based on pattern
-// fstab.<fstab_suffix>, fstab.<hardware>, fstab.<hardware.platform> in
-// folders /odm/etc, vendor/etc, or /.
+// Return the path to the fstab file.  There may be multiple fstab files; the
+// one that is returned will be the first that exists of fstab.<fstab_suffix>,
+// fstab.<hardware>, and fstab.<hardware.platform>.  The fstab is searched for
+// in /odm/etc/ and /vendor/etc/, as well as in the locations where it may be in
+// the first stage ramdisk during early boot.  Previously, the first stage
+// ramdisk's copy of the fstab had to be located in the root directory, but now
+// the system/etc directory is supported too and is the preferred location.
 std::string GetFstabPath() {
     for (const char* prop : {"fstab_suffix", "hardware", "hardware.platform"}) {
         std::string suffix;
 
         if (!fs_mgr_get_boot_config(prop, &suffix)) continue;
 
-        for (const char* prefix :
-             {"/odm/etc/fstab.", "/vendor/etc/fstab.", "/fstab.", "/first_stage_ramdisk/fstab."}) {
+        for (const char* prefix : {// late-boot/post-boot locations
+                                   "/odm/etc/fstab.", "/vendor/etc/fstab.",
+                                   // early boot locations
+                                   "/system/etc/fstab.", "/first_stage_ramdisk/system/etc/fstab.",
+                                   "/fstab.", "/first_stage_ramdisk/fstab."}) {
             std::string fstab_path = prefix + suffix;
             if (access(fstab_path.c_str(), F_OK) == 0) {
                 return fstab_path;
diff --git a/fs_mgr/libdm/dm.cpp b/fs_mgr/libdm/dm.cpp
index c4874b8..b1d5b39 100644
--- a/fs_mgr/libdm/dm.cpp
+++ b/fs_mgr/libdm/dm.cpp
@@ -170,19 +170,18 @@
     return access("/system/bin/recovery", F_OK) == 0;
 }
 
-bool DeviceMapper::CreateDevice(const std::string& name, const DmTable& table, std::string* path,
-                                const std::chrono::milliseconds& timeout_ms) {
+bool DeviceMapper::CreateEmptyDevice(const std::string& name) {
     std::string uuid = GenerateUuid();
-    if (!CreateDevice(name, uuid)) {
-        return false;
-    }
+    return CreateDevice(name, uuid);
+}
 
+bool DeviceMapper::WaitForDevice(const std::string& name,
+                                 const std::chrono::milliseconds& timeout_ms, std::string* path) {
     // We use the unique path for testing whether the device is ready. After
     // that, it's safe to use the dm-N path which is compatible with callers
     // that expect it to be formatted as such.
     std::string unique_path;
-    if (!LoadTableAndActivate(name, table) || !GetDeviceUniquePath(name, &unique_path) ||
-        !GetDmDevicePathByName(name, path)) {
+    if (!GetDeviceUniquePath(name, &unique_path) || !GetDmDevicePathByName(name, path)) {
         DeleteDevice(name);
         return false;
     }
@@ -208,6 +207,25 @@
     return true;
 }
 
+bool DeviceMapper::CreateDevice(const std::string& name, const DmTable& table, std::string* path,
+                                const std::chrono::milliseconds& timeout_ms) {
+    if (!CreateEmptyDevice(name)) {
+        return false;
+    }
+
+    if (!LoadTableAndActivate(name, table)) {
+        DeleteDevice(name);
+        return false;
+    }
+
+    if (!WaitForDevice(name, timeout_ms, path)) {
+        DeleteDevice(name);
+        return false;
+    }
+
+    return true;
+}
+
 bool DeviceMapper::GetDeviceUniquePath(const std::string& name, std::string* path) {
     struct dm_ioctl io;
     InitIo(&io, name);
@@ -560,34 +578,30 @@
     return std::string{spec.target_type, sizeof(spec.target_type)};
 }
 
-static bool ExtractBlockDeviceName(const std::string& path, std::string* name) {
+std::optional<std::string> ExtractBlockDeviceName(const std::string& path) {
     static constexpr std::string_view kDevBlockPrefix("/dev/block/");
     if (android::base::StartsWith(path, kDevBlockPrefix)) {
-        *name = path.substr(kDevBlockPrefix.length());
-        return true;
+        return path.substr(kDevBlockPrefix.length());
     }
-    return false;
+    return {};
 }
 
 bool DeviceMapper::IsDmBlockDevice(const std::string& path) {
-    std::string name;
-    if (!ExtractBlockDeviceName(path, &name)) {
-        return false;
-    }
-    return android::base::StartsWith(name, "dm-");
+    std::optional<std::string> name = ExtractBlockDeviceName(path);
+    return name && android::base::StartsWith(*name, "dm-");
 }
 
 std::optional<std::string> DeviceMapper::GetDmDeviceNameByPath(const std::string& path) {
-    std::string name;
-    if (!ExtractBlockDeviceName(path, &name)) {
+    std::optional<std::string> name = ExtractBlockDeviceName(path);
+    if (!name) {
         LOG(WARNING) << path << " is not a block device";
         return std::nullopt;
     }
-    if (!android::base::StartsWith(name, "dm-")) {
+    if (!android::base::StartsWith(*name, "dm-")) {
         LOG(WARNING) << path << " is not a dm device";
         return std::nullopt;
     }
-    std::string dm_name_file = "/sys/block/" + name + "/dm/name";
+    std::string dm_name_file = "/sys/block/" + *name + "/dm/name";
     std::string dm_name;
     if (!android::base::ReadFileToString(dm_name_file, &dm_name)) {
         PLOG(ERROR) << "Failed to read file " << dm_name_file;
@@ -598,16 +612,16 @@
 }
 
 std::optional<std::string> DeviceMapper::GetParentBlockDeviceByPath(const std::string& path) {
-    std::string name;
-    if (!ExtractBlockDeviceName(path, &name)) {
+    std::optional<std::string> name = ExtractBlockDeviceName(path);
+    if (!name) {
         LOG(WARNING) << path << " is not a block device";
         return std::nullopt;
     }
-    if (!android::base::StartsWith(name, "dm-")) {
+    if (!android::base::StartsWith(*name, "dm-")) {
         // Reached bottom of the device mapper stack.
         return std::nullopt;
     }
-    auto slaves_dir = "/sys/block/" + name + "/slaves";
+    auto slaves_dir = "/sys/block/" + *name + "/slaves";
     auto dir = std::unique_ptr<DIR, decltype(&closedir)>(opendir(slaves_dir.c_str()), closedir);
     if (dir == nullptr) {
         PLOG(ERROR) << "Failed to open: " << slaves_dir;
diff --git a/fs_mgr/libdm/dm_test.cpp b/fs_mgr/libdm/dm_test.cpp
index 8006db2..541f254 100644
--- a/fs_mgr/libdm/dm_test.cpp
+++ b/fs_mgr/libdm/dm_test.cpp
@@ -29,6 +29,7 @@
 #include <thread>
 
 #include <android-base/file.h>
+#include <android-base/scopeguard.h>
 #include <android-base/strings.h>
 #include <android-base/unique_fd.h>
 #include <gtest/gtest.h>
@@ -679,3 +680,13 @@
     ASSERT_NE(0, access(path.c_str(), F_OK));
     ASSERT_EQ(ENOENT, errno);
 }
+
+TEST(libdm, CreateEmptyDevice) {
+    DeviceMapper& dm = DeviceMapper::Instance();
+    ASSERT_TRUE(dm.CreateEmptyDevice("empty-device"));
+    auto guard =
+            android::base::make_scope_guard([&]() { dm.DeleteDeviceIfExists("empty-device", 5s); });
+
+    // Empty device should be in suspended state.
+    ASSERT_EQ(DmDeviceState::SUSPENDED, dm.GetState("empty-device"));
+}
diff --git a/fs_mgr/libdm/include/libdm/dm.h b/fs_mgr/libdm/include/libdm/dm.h
index 70b14fa..e6698ea 100644
--- a/fs_mgr/libdm/include/libdm/dm.h
+++ b/fs_mgr/libdm/include/libdm/dm.h
@@ -49,6 +49,10 @@
 
 static constexpr uint64_t kSectorSize = 512;
 
+// Returns `path` without /dev/block prefix if and only if `path` starts with
+// that prefix.
+std::optional<std::string> ExtractBlockDeviceName(const std::string& path);
+
 class DeviceMapper final {
   public:
     class DmBlockDevice final {
@@ -115,6 +119,19 @@
     // - ACTIVE: resumes the device.
     bool ChangeState(const std::string& name, DmDeviceState state);
 
+    // Creates empty device.
+    // This supports a use case when a caller doesn't need a device straight away, but instead
+    // asks kernel to create it beforehand, thus avoiding blocking itself from waiting for ueventd
+    // to create user space paths.
+    // Callers are expected to then activate their device by calling LoadTableAndActivate function.
+    // To avoid race conditions, callers must still synchronize with ueventd by calling
+    // WaitForDevice function.
+    bool CreateEmptyDevice(const std::string& name);
+
+    // Waits for device paths to be created in the user space.
+    bool WaitForDevice(const std::string& name, const std::chrono::milliseconds& timeout_ms,
+                       std::string* path);
+
     // Creates a device, loads the given table, and activates it. If the device
     // is not able to be activated, it is destroyed, and false is returned.
     // After creation, |path| contains the result of calling
diff --git a/fs_mgr/libsnapshot/cow_snapuserd_test.cpp b/fs_mgr/libsnapshot/cow_snapuserd_test.cpp
index d09c6e9..bd432bb 100644
--- a/fs_mgr/libsnapshot/cow_snapuserd_test.cpp
+++ b/fs_mgr/libsnapshot/cow_snapuserd_test.cpp
@@ -108,6 +108,7 @@
     void MergeInterruptFixed(int duration);
     void MergeInterruptRandomly(int max_duration);
     void ReadDmUserBlockWithoutDaemon();
+    void ReadLastBlock();
 
     std::string snapshot_dev() const { return snapshot_dev_->path(); }
 
@@ -256,6 +257,73 @@
     }
 }
 
+void CowSnapuserdTest::ReadLastBlock() {
+    unique_fd rnd_fd;
+    total_base_size_ = BLOCK_SZ * 2;
+
+    base_fd_ = CreateTempFile("base_device", total_base_size_);
+    ASSERT_GE(base_fd_, 0);
+
+    rnd_fd.reset(open("/dev/random", O_RDONLY));
+    ASSERT_TRUE(rnd_fd > 0);
+
+    std::unique_ptr<uint8_t[]> random_buffer = std::make_unique<uint8_t[]>(BLOCK_SZ);
+
+    for (size_t j = 0; j < ((total_base_size_) / BLOCK_SZ); j++) {
+        ASSERT_EQ(ReadFullyAtOffset(rnd_fd, (char*)random_buffer.get(), BLOCK_SZ, 0), true);
+        ASSERT_EQ(android::base::WriteFully(base_fd_, random_buffer.get(), BLOCK_SZ), true);
+    }
+
+    ASSERT_EQ(lseek(base_fd_, 0, SEEK_SET), 0);
+
+    base_loop_ = std::make_unique<LoopDevice>(base_fd_, 10s);
+    ASSERT_TRUE(base_loop_->valid());
+
+    std::string path = android::base::GetExecutableDirectory();
+    cow_system_ = std::make_unique<TemporaryFile>(path);
+
+    std::unique_ptr<uint8_t[]> random_buffer_1_ = std::make_unique<uint8_t[]>(total_base_size_);
+    loff_t offset = 0;
+
+    // Fill random data
+    for (size_t j = 0; j < (total_base_size_ / BLOCK_SZ); j++) {
+        ASSERT_EQ(ReadFullyAtOffset(rnd_fd, (char*)random_buffer_1_.get() + offset, BLOCK_SZ, 0),
+                  true);
+
+        offset += BLOCK_SZ;
+    }
+
+    CowOptions options;
+    options.compression = "gz";
+    CowWriter writer(options);
+
+    ASSERT_TRUE(writer.Initialize(cow_system_->fd));
+
+    ASSERT_TRUE(writer.AddRawBlocks(0, random_buffer_1_.get(), BLOCK_SZ));
+    ASSERT_TRUE(writer.AddRawBlocks(1, (char*)random_buffer_1_.get() + BLOCK_SZ, BLOCK_SZ));
+
+    ASSERT_TRUE(writer.Finalize());
+
+    SetDeviceControlName();
+
+    StartSnapuserdDaemon();
+    InitCowDevice();
+
+    CreateDmUserDevice();
+    InitDaemon();
+
+    CreateSnapshotDevice();
+
+    unique_fd snapshot_fd(open(snapshot_dev_->path().c_str(), O_RDONLY));
+    ASSERT_TRUE(snapshot_fd > 0);
+
+    std::unique_ptr<uint8_t[]> snapuserd_buffer = std::make_unique<uint8_t[]>(BLOCK_SZ);
+
+    offset = 7680;
+    ASSERT_EQ(ReadFullyAtOffset(snapshot_fd, snapuserd_buffer.get(), 512, offset), true);
+    ASSERT_EQ(memcmp(snapuserd_buffer.get(), (char*)random_buffer_1_.get() + offset, 512), 0);
+}
+
 void CowSnapuserdTest::CreateBaseDevice() {
     unique_fd rnd_fd;
 
@@ -1068,6 +1136,12 @@
     harness.Shutdown();
 }
 
+TEST(Snapuserd_Test, Snapshot_END_IO_TEST) {
+    CowSnapuserdTest harness;
+    harness.ReadLastBlock();
+    harness.Shutdown();
+}
+
 TEST(Snapuserd_Test, Snapshot_COPY_Overlap_TEST_1) {
     CowSnapuserdTest harness;
     ASSERT_TRUE(harness.SetupCopyOverlap_1());
diff --git a/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h b/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
index 15882b3..69d8967 100644
--- a/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
+++ b/fs_mgr/libsnapshot/include/libsnapshot/snapshot.h
@@ -386,6 +386,17 @@
     // first-stage to decide whether to launch snapuserd.
     bool IsSnapuserdRequired();
 
+    // Add new public entries above this line.
+
+    // Helpers for failure injection.
+    using MergeConsistencyChecker =
+            std::function<MergeFailureCode(const std::string& name, const SnapshotStatus& status)>;
+
+    void set_merge_consistency_checker(MergeConsistencyChecker checker) {
+        merge_consistency_checker_ = checker;
+    }
+    MergeConsistencyChecker merge_consistency_checker() const { return merge_consistency_checker_; }
+
   private:
     FRIEND_TEST(SnapshotTest, CleanFirstStageMount);
     FRIEND_TEST(SnapshotTest, CreateSnapshot);
@@ -399,6 +410,8 @@
     FRIEND_TEST(SnapshotTest, MergeFailureCode);
     FRIEND_TEST(SnapshotTest, NoMergeBeforeReboot);
     FRIEND_TEST(SnapshotTest, UpdateBootControlHal);
+    FRIEND_TEST(SnapshotUpdateTest, AddPartition);
+    FRIEND_TEST(SnapshotUpdateTest, ConsistencyCheckResume);
     FRIEND_TEST(SnapshotUpdateTest, DaemonTransition);
     FRIEND_TEST(SnapshotUpdateTest, DataWipeAfterRollback);
     FRIEND_TEST(SnapshotUpdateTest, DataWipeRollbackInRecovery);
@@ -781,6 +794,7 @@
     std::function<bool(const std::string&)> uevent_regen_callback_;
     std::unique_ptr<SnapuserdClient> snapuserd_client_;
     std::unique_ptr<LpMetadata> old_partition_metadata_;
+    MergeConsistencyChecker merge_consistency_checker_;
 };
 
 }  // namespace snapshot
diff --git a/fs_mgr/libsnapshot/snapshot.cpp b/fs_mgr/libsnapshot/snapshot.cpp
index 2773be7..9a9f400 100644
--- a/fs_mgr/libsnapshot/snapshot.cpp
+++ b/fs_mgr/libsnapshot/snapshot.cpp
@@ -87,6 +87,8 @@
 static constexpr char kRollbackIndicatorPath[] = "/metadata/ota/rollback-indicator";
 static constexpr auto kUpdateStateCheckInterval = 2s;
 
+MergeFailureCode CheckMergeConsistency(const std::string& name, const SnapshotStatus& status);
+
 // Note: IImageManager is an incomplete type in the header, so the default
 // destructor doesn't work.
 SnapshotManager::~SnapshotManager() {}
@@ -116,6 +118,7 @@
 
 SnapshotManager::SnapshotManager(IDeviceInfo* device) : device_(device) {
     metadata_dir_ = device_->GetMetadataDir();
+    merge_consistency_checker_ = android::snapshot::CheckMergeConsistency;
 }
 
 static std::string GetCowName(const std::string& snapshot_name) {
@@ -518,6 +521,13 @@
             break;
     }
 
+    if (mode == SnapshotStorageMode::Persistent && status.state() == SnapshotState::MERGING) {
+        LOG(ERROR) << "Snapshot: " << name
+                   << " has snapshot status Merging but mode set to Persistent."
+                   << " Changing mode to Snapshot-Merge.";
+        mode = SnapshotStorageMode::Merge;
+    }
+
     DmTable table;
     table.Emplace<DmTargetSnapshot>(0, snapshot_sectors, base_device, cow_device, mode,
                                     kSnapshotChunkSize);
@@ -886,6 +896,10 @@
     if (target_type) {
         *target_type = DeviceMapper::GetTargetType(target.spec);
     }
+    if (!status->error.empty()) {
+        LOG(ERROR) << "Snapshot: " << dm_name << " returned error code: " << status->error;
+        return false;
+    }
     return true;
 }
 
@@ -1164,6 +1178,10 @@
                                                         const SnapshotStatus& status) {
     CHECK(lock);
 
+    return merge_consistency_checker_(name, status);
+}
+
+MergeFailureCode CheckMergeConsistency(const std::string& name, const SnapshotStatus& status) {
     if (!status.compression_enabled()) {
         // Do not try to verify old-style COWs yet.
         return MergeFailureCode::Ok;
@@ -1241,9 +1259,11 @@
     }
 
     SnapshotUpdateStatus update_status = ReadSnapshotUpdateStatus(lock);
-    CHECK(update_status.state() == UpdateState::Merging);
+    CHECK(update_status.state() == UpdateState::Merging ||
+          update_status.state() == UpdateState::MergeFailed);
     CHECK(update_status.merge_phase() == MergePhase::FIRST_PHASE);
 
+    update_status.set_state(UpdateState::Merging);
     update_status.set_merge_phase(MergePhase::SECOND_PHASE);
     if (!WriteSnapshotUpdateStatus(lock, update_status)) {
         return MergeFailureCode::WriteStatus;
@@ -1456,7 +1476,7 @@
                                             std::vector<std::string>* snapuserd_argv) {
     LOG(INFO) << "Performing transition for snapuserd.";
 
-    // Don't use EnsuerSnapuserdConnected() because this is called from init,
+    // Don't use EnsureSnapuserdConnected() because this is called from init,
     // and attempting to do so will deadlock.
     if (!snapuserd_client_ && transition != InitTransition::SELINUX_DETACH) {
         snapuserd_client_ = SnapuserdClient::Connect(kSnapuserdSocket, 10s);
@@ -1513,8 +1533,15 @@
             continue;
         }
 
+        std::string source_device_name;
+        if (snapshot_status.old_partition_size() > 0) {
+            source_device_name = GetSourceDeviceName(snapshot);
+        } else {
+            source_device_name = GetBaseDeviceName(snapshot);
+        }
+
         std::string source_device;
-        if (!dm.GetDmDevicePathByName(GetSourceDeviceName(snapshot), &source_device)) {
+        if (!dm.GetDmDevicePathByName(source_device_name, &source_device)) {
             LOG(ERROR) << "Could not get device path for " << GetSourceDeviceName(snapshot);
             continue;
         }
@@ -2091,14 +2118,18 @@
     if (live_snapshot_status->compression_enabled()) {
         // Get the source device (eg the view of the partition from before it was resized).
         std::string source_device_path;
-        if (!MapSourceDevice(lock, params.GetPartitionName(), remaining_time,
-                             &source_device_path)) {
-            LOG(ERROR) << "Could not map source device for: " << cow_name;
-            return false;
-        }
+        if (live_snapshot_status->old_partition_size() > 0) {
+            if (!MapSourceDevice(lock, params.GetPartitionName(), remaining_time,
+                                 &source_device_path)) {
+                LOG(ERROR) << "Could not map source device for: " << cow_name;
+                return false;
+            }
 
-        auto source_device = GetSourceDeviceName(params.GetPartitionName());
-        created_devices.EmplaceBack<AutoUnmapDevice>(&dm, source_device);
+            auto source_device = GetSourceDeviceName(params.GetPartitionName());
+            created_devices.EmplaceBack<AutoUnmapDevice>(&dm, source_device);
+        } else {
+            source_device_path = base_path;
+        }
 
         if (!WaitForDevice(source_device_path, remaining_time)) {
             return false;
@@ -2534,6 +2565,7 @@
         SnapshotUpdateStatus old_status = ReadSnapshotUpdateStatus(lock);
         status.set_compression_enabled(old_status.compression_enabled());
         status.set_source_build_fingerprint(old_status.source_build_fingerprint());
+        status.set_merge_phase(old_status.merge_phase());
     }
     return WriteSnapshotUpdateStatus(lock, status);
 }
diff --git a/fs_mgr/libsnapshot/snapshot_test.cpp b/fs_mgr/libsnapshot/snapshot_test.cpp
index 57a2854..272bc2e 100644
--- a/fs_mgr/libsnapshot/snapshot_test.cpp
+++ b/fs_mgr/libsnapshot/snapshot_test.cpp
@@ -289,7 +289,7 @@
     }
 
     AssertionResult DeleteDevice(const std::string& device) {
-        if (!dm_.DeleteDeviceIfExists(device)) {
+        if (!sm->DeleteDeviceIfExists(device, 1s)) {
             return AssertionFailure() << "Can't delete " << device;
         }
         return AssertionSuccess();
@@ -967,7 +967,7 @@
     }
 
     AssertionResult UnmapAll() {
-        for (const auto& name : {"sys", "vnd", "prd"}) {
+        for (const auto& name : {"sys", "vnd", "prd", "dlkm"}) {
             if (!dm_.DeleteDeviceIfExists(name + "_a"s)) {
                 return AssertionFailure() << "Cannot unmap " << name << "_a";
             }
@@ -1301,6 +1301,92 @@
     }
 }
 
+// Test that a transient merge consistency check failure can resume properly.
+TEST_F(SnapshotUpdateTest, ConsistencyCheckResume) {
+    if (!IsCompressionEnabled()) {
+        // b/179111359
+        GTEST_SKIP() << "Skipping Virtual A/B Compression test";
+    }
+
+    auto old_sys_size = GetSize(sys_);
+    auto old_prd_size = GetSize(prd_);
+
+    // Grow |sys| but shrink |prd|.
+    SetSize(sys_, old_sys_size * 2);
+    sys_->set_estimate_cow_size(8_MiB);
+    SetSize(prd_, old_prd_size / 2);
+    prd_->set_estimate_cow_size(1_MiB);
+
+    AddOperationForPartitions();
+
+    ASSERT_TRUE(sm->BeginUpdate());
+    ASSERT_TRUE(sm->CreateUpdateSnapshots(manifest_));
+    ASSERT_TRUE(WriteSnapshotAndHash("sys_b"));
+    ASSERT_TRUE(WriteSnapshotAndHash("vnd_b"));
+    ASSERT_TRUE(ShiftAllSnapshotBlocks("prd_b", old_prd_size));
+
+    sync();
+
+    // Assert that source partitions aren't affected.
+    for (const auto& name : {"sys_a", "vnd_a", "prd_a"}) {
+        ASSERT_TRUE(IsPartitionUnchanged(name));
+    }
+
+    ASSERT_TRUE(sm->FinishedSnapshotWrites(false));
+
+    // Simulate shutting down the device.
+    ASSERT_TRUE(UnmapAll());
+
+    // After reboot, init does first stage mount.
+    auto init = NewManagerForFirstStageMount("_b");
+    ASSERT_NE(init, nullptr);
+    ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
+    ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
+
+    // Check that the target partitions have the same content.
+    for (const auto& name : {"sys_b", "vnd_b", "prd_b"}) {
+        ASSERT_TRUE(IsPartitionUnchanged(name));
+    }
+
+    auto old_checker = init->merge_consistency_checker();
+
+    init->set_merge_consistency_checker(
+            [](const std::string&, const SnapshotStatus&) -> MergeFailureCode {
+                return MergeFailureCode::WrongMergeCountConsistencyCheck;
+            });
+
+    // Initiate the merge and wait for it to be completed.
+    ASSERT_TRUE(init->InitiateMerge());
+    {
+        // Check that the merge phase is FIRST_PHASE until at least one call
+        // to ProcessUpdateState() occurs.
+        ASSERT_TRUE(AcquireLock());
+        auto local_lock = std::move(lock_);
+        auto status = init->ReadSnapshotUpdateStatus(local_lock.get());
+        ASSERT_EQ(status.merge_phase(), MergePhase::FIRST_PHASE);
+    }
+
+    // Merge should have failed.
+    ASSERT_EQ(UpdateState::MergeFailed, init->ProcessUpdateState());
+
+    // Simulate shutting down the device and creating partitions again.
+    ASSERT_TRUE(UnmapAll());
+
+    // Restore the checker.
+    init->set_merge_consistency_checker(std::move(old_checker));
+
+    ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
+
+    // Complete the merge.
+    ASSERT_EQ(UpdateState::MergeCompleted, init->ProcessUpdateState());
+
+    // Check that the target partitions have the same content after the merge.
+    for (const auto& name : {"sys_b", "vnd_b", "prd_b"}) {
+        ASSERT_TRUE(IsPartitionUnchanged(name))
+                << "Content of " << name << " changes after the merge";
+    }
+}
+
 // Test that if new system partitions uses empty space in super, that region is not snapshotted.
 TEST_F(SnapshotUpdateTest, DirectWriteEmptySpace) {
     GTEST_SKIP() << "b/141889746";
@@ -2030,6 +2116,80 @@
     ASSERT_LT(res.required_size(), 40_MiB);
 }
 
+TEST_F(SnapshotUpdateTest, AddPartition) {
+    // OTA client blindly unmaps all partitions that are possibly mapped.
+    for (const auto& name : {"sys_b", "vnd_b", "prd_b"}) {
+        ASSERT_TRUE(sm->UnmapUpdateSnapshot(name));
+    }
+
+    group_->add_partition_names("dlkm");
+
+    auto dlkm = manifest_.add_partitions();
+    dlkm->set_partition_name("dlkm");
+    dlkm->set_estimate_cow_size(2_MiB);
+    SetSize(dlkm, 3_MiB);
+
+    // Grow all partitions. Set |prd| large enough that |sys| and |vnd|'s COWs
+    // fit in super, but not |prd|.
+    constexpr uint64_t partition_size = 3788_KiB;
+    SetSize(sys_, partition_size);
+    SetSize(vnd_, partition_size);
+    SetSize(prd_, partition_size);
+    SetSize(dlkm, partition_size);
+
+    AddOperationForPartitions({sys_, vnd_, prd_, dlkm});
+
+    // Execute the update.
+    ASSERT_TRUE(sm->BeginUpdate());
+    ASSERT_TRUE(sm->CreateUpdateSnapshots(manifest_));
+
+    // Write some data to target partitions.
+    for (const auto& name : {"sys_b", "vnd_b", "prd_b", "dlkm_b"}) {
+        ASSERT_TRUE(WriteSnapshotAndHash(name));
+    }
+
+    // Assert that source partitions aren't affected.
+    for (const auto& name : {"sys_a", "vnd_a", "prd_a"}) {
+        ASSERT_TRUE(IsPartitionUnchanged(name));
+    }
+
+    ASSERT_TRUE(sm->FinishedSnapshotWrites(false));
+
+    // Simulate shutting down the device.
+    ASSERT_TRUE(UnmapAll());
+
+    // After reboot, init does first stage mount.
+    auto init = NewManagerForFirstStageMount("_b");
+    ASSERT_NE(init, nullptr);
+
+    ASSERT_TRUE(init->EnsureSnapuserdConnected());
+    init->set_use_first_stage_snapuserd(true);
+
+    ASSERT_TRUE(init->NeedSnapshotsInFirstStageMount());
+    ASSERT_TRUE(init->CreateLogicalAndSnapshotPartitions("super", snapshot_timeout_));
+
+    // Check that the target partitions have the same content.
+    std::vector<std::string> partitions = {"sys_b", "vnd_b", "prd_b", "dlkm_b"};
+    for (const auto& name : partitions) {
+        ASSERT_TRUE(IsPartitionUnchanged(name));
+    }
+
+    ASSERT_TRUE(init->PerformInitTransition(SnapshotManager::InitTransition::SECOND_STAGE));
+    for (const auto& name : partitions) {
+        ASSERT_TRUE(init->snapuserd_client()->WaitForDeviceDelete(name + "-user-cow-init"));
+    }
+
+    // Initiate the merge and wait for it to be completed.
+    ASSERT_TRUE(init->InitiateMerge());
+    ASSERT_EQ(UpdateState::MergeCompleted, init->ProcessUpdateState());
+
+    // Check that the target partitions have the same content after the merge.
+    for (const auto& name : {"sys_b", "vnd_b", "prd_b", "dlkm_b"}) {
+        ASSERT_TRUE(IsPartitionUnchanged(name))
+                << "Content of " << name << " changes after the merge";
+    }
+}
+
 class AutoKill final {
   public:
     explicit AutoKill(pid_t pid) : pid_(pid) {}
@@ -2257,7 +2417,6 @@
     }
     void TearDown() override {
         RETURN_IF_NON_VIRTUAL_AB();
-        return;  // BUG(149738928)
 
         EXPECT_TRUE(!image_manager_->BackingImageExists(kImageName) ||
                     image_manager_->DeleteBackingImage(kImageName));
@@ -2266,19 +2425,6 @@
     std::unique_ptr<LowSpaceUserdata> userdata_;
 };
 
-TEST_P(ImageManagerTest, CreateImageEnoughAvailSpace) {
-    if (userdata_->available_space() == 0) {
-        GTEST_SKIP() << "/data is full (" << userdata_->available_space()
-                     << " bytes available), skipping";
-    }
-    ASSERT_TRUE(image_manager_->CreateBackingImage(kImageName, userdata_->available_space(),
-                                                   IImageManager::CREATE_IMAGE_DEFAULT))
-            << "Should be able to create image with size = " << userdata_->available_space()
-            << " bytes";
-    ASSERT_TRUE(image_manager_->DeleteBackingImage(kImageName))
-            << "Should be able to delete created image";
-}
-
 TEST_P(ImageManagerTest, CreateImageNoSpace) {
     uint64_t to_allocate = userdata_->free_space() + userdata_->bsize();
     auto res = image_manager_->CreateBackingImage(kImageName, to_allocate,
diff --git a/fs_mgr/libsnapshot/snapuserd_readahead.cpp b/fs_mgr/libsnapshot/snapuserd_readahead.cpp
index 16d5919..e55fea3 100644
--- a/fs_mgr/libsnapshot/snapuserd_readahead.cpp
+++ b/fs_mgr/libsnapshot/snapuserd_readahead.cpp
@@ -226,9 +226,15 @@
     int num_ops = 0;
     int total_blocks_merged = 0;
 
+    // This memcpy is important as metadata_buffer_ will be an unaligned address and will fault
+    // on 32-bit systems
+    std::unique_ptr<uint8_t[]> metadata_buffer =
+            std::make_unique<uint8_t[]>(snapuserd_->GetBufferMetadataSize());
+    memcpy(metadata_buffer.get(), metadata_buffer_, snapuserd_->GetBufferMetadataSize());
+
     while (true) {
         struct ScratchMetadata* bm = reinterpret_cast<struct ScratchMetadata*>(
-                (char*)metadata_buffer_ + metadata_offset);
+                (char*)metadata_buffer.get() + metadata_offset);
 
         // Done reading metadata
         if (bm->new_block == 0 && bm->file_offset == 0) {
diff --git a/fs_mgr/libsnapshot/snapuserd_worker.cpp b/fs_mgr/libsnapshot/snapuserd_worker.cpp
index 682f9da..defb5bb 100644
--- a/fs_mgr/libsnapshot/snapuserd_worker.cpp
+++ b/fs_mgr/libsnapshot/snapuserd_worker.cpp
@@ -287,16 +287,36 @@
     it = std::lower_bound(chunk_vec.begin(), chunk_vec.end(), std::make_pair(sector, nullptr),
                           Snapuserd::compare);
 
-    if (!(it != chunk_vec.end())) {
-        SNAP_LOG(ERROR) << "ReadData: Sector " << sector << " not found in chunk_vec";
-        return -1;
+    bool read_end_of_device = false;
+    if (it == chunk_vec.end()) {
+        // |-------|-------|-------|
+        // 0       1       2       3
+        //
+        // Block 0 - op 1
+        // Block 1 - op 2
+        // Block 2 - op 3
+        //
+        // chunk_vec will have block 0, 1, 2 which maps to relavant COW ops.
+        //
+        // Each block is 4k bytes. Thus, the last block will span 8 sectors
+        // ranging till block 3 (However, block 3 won't be in chunk_vec as
+        // it doesn't have any mapping to COW ops. Now, if we get an I/O request for a sector
+        // spanning between block 2 and block 3, we need to step back
+        // and get hold of the last element.
+        //
+        // Additionally, dm-snapshot makes sure that I/O request beyond block 3
+        // will not be routed to the daemon. Hence, it is safe to assume that
+        // if a sector is not available in the chunk_vec, the I/O falls in the
+        // end of region.
+        it = std::prev(chunk_vec.end());
+        read_end_of_device = true;
     }
 
     // We didn't find the required sector; hence find the previous sector
     // as lower_bound will gives us the value greater than
     // the requested sector
     if (it->first != sector) {
-        if (it != chunk_vec.begin()) {
+        if (it != chunk_vec.begin() && !read_end_of_device) {
             --it;
         }
 
diff --git a/init/Android.bp b/init/Android.bp
index 7eeafa2..a57f3a4 100644
--- a/init/Android.bp
+++ b/init/Android.bp
@@ -89,7 +89,19 @@
     "host_init_verifier.cpp",
 ]
 
-cc_defaults {
+soong_config_module_type {
+    name: "libinit_cc_defaults",
+    module_type: "cc_defaults",
+    config_namespace: "ANDROID",
+    bool_variables: [
+        "PRODUCT_INSTALL_DEBUG_POLICY_TO_SYSTEM_EXT",
+    ],
+    properties: [
+        "cflags",
+    ],
+}
+
+libinit_cc_defaults {
     name: "init_defaults",
     sanitize: {
         misc_undefined: ["signed-integer-overflow"],
@@ -109,6 +121,7 @@
         "-DDUMP_ON_UMOUNT_FAILURE=0",
         "-DSHUTDOWN_ZERO_TIMEOUT=0",
         "-DINIT_FULL_SOURCES",
+        "-DINSTALL_DEBUG_POLICY_TO_SYSTEM_EXT=0",
     ],
     product_variables: {
         debuggable: {
@@ -137,6 +150,14 @@
             cppflags: ["-DUSER_MODE_LINUX"],
         },
     },
+    soong_config_variables: {
+        PRODUCT_INSTALL_DEBUG_POLICY_TO_SYSTEM_EXT: {
+            cflags: [
+                "-UINSTALL_DEBUG_POLICY_TO_SYSTEM_EXT",
+                "-DINSTALL_DEBUG_POLICY_TO_SYSTEM_EXT=1",
+            ],
+        },
+    },
     static_libs: [
         "libavb",
         "libc++fs",
diff --git a/init/first_stage_init.cpp b/init/first_stage_init.cpp
index 78e5b60..c7b7b0c 100644
--- a/init/first_stage_init.cpp
+++ b/init/first_stage_init.cpp
@@ -330,14 +330,21 @@
     // If "/force_debuggable" is present, the second-stage init will use a userdebug
     // sepolicy and load adb_debug.prop to allow adb root, if the device is unlocked.
     if (access("/force_debuggable", F_OK) == 0) {
+        constexpr const char adb_debug_prop_src[] = "/adb_debug.prop";
+        constexpr const char userdebug_plat_sepolicy_cil_src[] = "/userdebug_plat_sepolicy.cil";
         std::error_code ec;  // to invoke the overloaded copy_file() that won't throw.
-        if (!fs::copy_file("/adb_debug.prop", kDebugRamdiskProp, ec) ||
-            !fs::copy_file("/userdebug_plat_sepolicy.cil", kDebugRamdiskSEPolicy, ec)) {
-            LOG(ERROR) << "Failed to setup debug ramdisk";
-        } else {
-            // setenv for second-stage init to read above kDebugRamdisk* files.
-            setenv("INIT_FORCE_DEBUGGABLE", "true", 1);
+        if (access(adb_debug_prop_src, F_OK) == 0 &&
+            !fs::copy_file(adb_debug_prop_src, kDebugRamdiskProp, ec)) {
+            LOG(WARNING) << "Can't copy " << adb_debug_prop_src << " to " << kDebugRamdiskProp
+                         << ": " << ec.message();
         }
+        if (access(userdebug_plat_sepolicy_cil_src, F_OK) == 0 &&
+            !fs::copy_file(userdebug_plat_sepolicy_cil_src, kDebugRamdiskSEPolicy, ec)) {
+            LOG(WARNING) << "Can't copy " << userdebug_plat_sepolicy_cil_src << " to "
+                         << kDebugRamdiskSEPolicy << ": " << ec.message();
+        }
+        // setenv for second-stage init to read above kDebugRamdisk* files.
+        setenv("INIT_FORCE_DEBUGGABLE", "true", 1);
     }
 
     if (ForceNormalBoot(cmdline, bootconfig)) {
diff --git a/init/init.cpp b/init/init.cpp
index a7325ca..942feb9 100644
--- a/init/init.cpp
+++ b/init/init.cpp
@@ -27,6 +27,7 @@
 #include <sys/mount.h>
 #include <sys/signalfd.h>
 #include <sys/types.h>
+#include <sys/utsname.h>
 #include <unistd.h>
 
 #define _REALLY_INCLUDE_SYS__SYSTEM_PROPERTIES_H_
@@ -554,6 +555,19 @@
     }
 }
 
+/// Set ro.kernel.version property to contain the major.minor pair as returned
+/// by uname(2).
+static void SetKernelVersion() {
+    struct utsname uts;
+    unsigned int major, minor;
+
+    if ((uname(&uts) != 0) || (sscanf(uts.release, "%u.%u", &major, &minor) != 2)) {
+        LOG(ERROR) << "Could not parse the kernel version from uname";
+        return;
+    }
+    SetProperty("ro.kernel.version", android::base::StringPrintf("%u.%u", major, minor));
+}
+
 static void HandleSigtermSignal(const signalfd_siginfo& siginfo) {
     if (siginfo.ssi_pid != 0) {
         // Drop any userspace SIGTERM requests.
@@ -824,6 +838,7 @@
     export_oem_lock_status();
     MountHandler mount_handler(&epoll);
     SetUsbController();
+    SetKernelVersion();
 
     const BuiltinFunctionMap& function_map = GetBuiltinFunctionMap();
     Action::set_function_map(&function_map);
diff --git a/init/selinux.cpp b/init/selinux.cpp
index 42d3023..29c0ff3 100644
--- a/init/selinux.cpp
+++ b/init/selinux.cpp
@@ -295,6 +295,25 @@
     return access(plat_policy_cil_file, R_OK) != -1;
 }
 
+std::optional<const char*> GetUserdebugPlatformPolicyFile() {
+    // See if we need to load userdebug_plat_sepolicy.cil instead of plat_sepolicy.cil.
+    const char* force_debuggable_env = getenv("INIT_FORCE_DEBUGGABLE");
+    if (force_debuggable_env && "true"s == force_debuggable_env && AvbHandle::IsDeviceUnlocked()) {
+        const std::vector<const char*> debug_policy_candidates = {
+#if INSTALL_DEBUG_POLICY_TO_SYSTEM_EXT == 1
+            "/system_ext/etc/selinux/userdebug_plat_sepolicy.cil",
+#endif
+            kDebugRamdiskSEPolicy,
+        };
+        for (const char* debug_policy : debug_policy_candidates) {
+            if (access(debug_policy, F_OK) == 0) {
+                return debug_policy;
+            }
+        }
+    }
+    return std::nullopt;
+}
+
 struct PolicyFile {
     unique_fd fd;
     std::string path;
@@ -310,13 +329,10 @@
     // secilc is invoked to compile the above three policy files into a single monolithic policy
     // file. This file is then loaded into the kernel.
 
-    // See if we need to load userdebug_plat_sepolicy.cil instead of plat_sepolicy.cil.
-    const char* force_debuggable_env = getenv("INIT_FORCE_DEBUGGABLE");
-    bool use_userdebug_policy =
-            ((force_debuggable_env && "true"s == force_debuggable_env) &&
-             AvbHandle::IsDeviceUnlocked() && access(kDebugRamdiskSEPolicy, F_OK) == 0);
+    const auto userdebug_plat_sepolicy = GetUserdebugPlatformPolicyFile();
+    const bool use_userdebug_policy = userdebug_plat_sepolicy.has_value();
     if (use_userdebug_policy) {
-        LOG(WARNING) << "Using userdebug system sepolicy";
+        LOG(INFO) << "Using userdebug system sepolicy " << *userdebug_plat_sepolicy;
     }
 
     // Load precompiled policy from vendor image, if a matching policy is found there. The policy
@@ -413,7 +429,7 @@
     // clang-format off
     std::vector<const char*> compile_args {
         "/system/bin/secilc",
-        use_userdebug_policy ? kDebugRamdiskSEPolicy: plat_policy_cil_file,
+        use_userdebug_policy ? *userdebug_plat_sepolicy : plat_policy_cil_file,
         "-m", "-M", "true", "-G", "-N",
         "-c", version_as_string.c_str(),
         plat_mapping_file.c_str(),
diff --git a/init/snapuserd_transition.cpp b/init/snapuserd_transition.cpp
index 40467b7..7fd3f65 100644
--- a/init/snapuserd_transition.cpp
+++ b/init/snapuserd_transition.cpp
@@ -32,6 +32,7 @@
 #include <android-base/strings.h>
 #include <android-base/unique_fd.h>
 #include <cutils/sockets.h>
+#include <fs_avb/fs_avb.h>
 #include <libsnapshot/snapshot.h>
 #include <libsnapshot/snapuserd_client.h>
 #include <private/android_filesystem_config.h>
@@ -227,6 +228,56 @@
     }
 }
 
+/*
+ * Before starting init second stage, we will wait
+ * for snapuserd daemon to be up and running; bionic libc
+ * may read /system/etc/selinux/plat_property_contexts file
+ * before invoking main() function. This will happen if
+ * init initializes property during second stage. Any access
+ * to /system without snapuserd daemon will lead to a deadlock.
+ *
+ * Thus, we do a simple probe by reading system partition. This
+ * read will eventually be serviced by daemon confirming that
+ * daemon is up and running. Furthermore, we are still in the kernel
+ * domain and sepolicy has not been enforced yet. Thus, access
+ * to these device mapper block devices are ok even though
+ * we may see audit logs.
+ */
+bool SnapuserdSelinuxHelper::TestSnapuserdIsReady() {
+    std::string dev = "/dev/block/mapper/system"s + fs_mgr_get_slot_suffix();
+    android::base::unique_fd fd(open(dev.c_str(), O_RDONLY | O_DIRECT));
+    if (fd < 0) {
+        PLOG(ERROR) << "open " << dev << " failed";
+        return false;
+    }
+
+    void* addr;
+    ssize_t page_size = getpagesize();
+    if (posix_memalign(&addr, page_size, page_size) < 0) {
+        PLOG(ERROR) << "posix_memalign with page size " << page_size;
+        return false;
+    }
+
+    std::unique_ptr<void, decltype(&::free)> buffer(addr, ::free);
+
+    int iter = 0;
+    while (iter < 10) {
+        ssize_t n = TEMP_FAILURE_RETRY(pread(fd.get(), buffer.get(), page_size, 0));
+        if (n < 0) {
+            // Wait for sometime before retry
+            std::this_thread::sleep_for(100ms);
+        } else if (n == page_size) {
+            return true;
+        } else {
+            LOG(ERROR) << "pread returned: " << n << " from: " << dev << " expected: " << page_size;
+        }
+
+        iter += 1;
+    }
+
+    return false;
+}
+
 void SnapuserdSelinuxHelper::RelaunchFirstStageSnapuserd() {
     auto fd = GetRamdiskSnapuserdFd();
     if (!fd) {
@@ -248,6 +299,13 @@
         setenv(kSnapuserdFirstStagePidVar, std::to_string(pid).c_str(), 1);
 
         LOG(INFO) << "Relaunched snapuserd with pid: " << pid;
+
+        if (!TestSnapuserdIsReady()) {
+            PLOG(FATAL) << "snapuserd daemon failed to launch";
+        } else {
+            LOG(INFO) << "snapuserd daemon is up and running";
+        }
+
         return;
     }
 
diff --git a/init/snapuserd_transition.h b/init/snapuserd_transition.h
index a5ab652..757af13 100644
--- a/init/snapuserd_transition.h
+++ b/init/snapuserd_transition.h
@@ -51,6 +51,7 @@
   private:
     void RelaunchFirstStageSnapuserd();
     void ExecSnapuserd();
+    bool TestSnapuserdIsReady();
 
     std::unique_ptr<SnapshotManager> sm_;
     BlockDevInitializer block_dev_init_;
diff --git a/libcutils/include/private/android_filesystem_config.h b/libcutils/include/private/android_filesystem_config.h
index 8f22d89..ffd4d12 100644
--- a/libcutils/include/private/android_filesystem_config.h
+++ b/libcutils/include/private/android_filesystem_config.h
@@ -157,6 +157,7 @@
 #define AID_READPROC 3009     /* Allow /proc read access */
 #define AID_WAKELOCK 3010     /* Allow system wakelock read/write access */
 #define AID_UHID 3011         /* Allow read/write to /dev/uhid node */
+#define AID_READTRACEFS 3012  /* Allow tracefs read */
 
 /* The range 5000-5999 is also reserved for vendor partition. */
 #define AID_OEM_RESERVED_2_START 5000
diff --git a/libprocessgroup/profiles/task_profiles.json b/libprocessgroup/profiles/task_profiles.json
index 449a505..45d3c7c 100644
--- a/libprocessgroup/profiles/task_profiles.json
+++ b/libprocessgroup/profiles/task_profiles.json
@@ -183,7 +183,19 @@
         }
       ]
     },
-
+    {
+      "Name": "Dex2oatPerformance",
+      "Actions": [
+        {
+          "Name": "JoinCgroup",
+          "Params":
+          {
+            "Controller": "cpu",
+            "Path": "dex2oat"
+          }
+        }
+      ]
+    },
     {
       "Name": "CpuPolicySpread",
       "Actions": [
@@ -638,7 +650,7 @@
     },
     {
       "Name": "Dex2OatBootComplete",
-      "Profiles": [ "SCHED_SP_BACKGROUND" ]
+      "Profiles": [ "Dex2oatPerformance", "LowIoPriority", "TimerSlackHigh" ]
     }
   ]
 }
diff --git a/libprocessgroup/profiles/task_profiles_28.json b/libprocessgroup/profiles/task_profiles_28.json
index 9f83785..e7be548 100644
--- a/libprocessgroup/profiles/task_profiles_28.json
+++ b/libprocessgroup/profiles/task_profiles_28.json
@@ -40,6 +40,19 @@
       ]
     },
     {
+      "Name": "ServicePerformance",
+      "Actions": [
+        {
+          "Name": "JoinCgroup",
+          "Params":
+          {
+            "Controller": "schedtune",
+            "Path": "background"
+          }
+        }
+      ]
+    },
+    {
       "Name": "HighPerformance",
       "Actions": [
         {
@@ -104,7 +117,19 @@
         }
       ]
     },
-
+    {
+      "Name": "Dex2oatPerformance",
+      "Actions": [
+        {
+          "Name": "JoinCgroup",
+          "Params":
+          {
+            "Controller": "schedtune",
+            "Path": "background"
+          }
+        }
+      ]
+    },
     {
       "Name": "CpuPolicySpread",
       "Actions": [
diff --git a/libprocessgroup/profiles/task_profiles_29.json b/libprocessgroup/profiles/task_profiles_29.json
index 9f83785..6174c8d 100644
--- a/libprocessgroup/profiles/task_profiles_29.json
+++ b/libprocessgroup/profiles/task_profiles_29.json
@@ -53,6 +53,19 @@
       ]
     },
     {
+      "Name": "ServicePerformance",
+      "Actions": [
+        {
+          "Name": "JoinCgroup",
+          "Params":
+          {
+            "Controller": "schedtune",
+            "Path": "background"
+          }
+        }
+      ]
+    },
+    {
       "Name": "MaxPerformance",
       "Actions": [
         {
@@ -104,7 +117,19 @@
         }
       ]
     },
-
+    {
+      "Name": "Dex2oatPerformance",
+      "Actions": [
+        {
+          "Name": "JoinCgroup",
+          "Params":
+          {
+            "Controller": "schedtune",
+            "Path": "background"
+          }
+        }
+      ]
+    },
     {
       "Name": "CpuPolicySpread",
       "Actions": [
diff --git a/libprocessgroup/profiles/task_profiles_30.json b/libprocessgroup/profiles/task_profiles_30.json
index 9f83785..e7be548 100644
--- a/libprocessgroup/profiles/task_profiles_30.json
+++ b/libprocessgroup/profiles/task_profiles_30.json
@@ -40,6 +40,19 @@
       ]
     },
     {
+      "Name": "ServicePerformance",
+      "Actions": [
+        {
+          "Name": "JoinCgroup",
+          "Params":
+          {
+            "Controller": "schedtune",
+            "Path": "background"
+          }
+        }
+      ]
+    },
+    {
       "Name": "HighPerformance",
       "Actions": [
         {
@@ -104,7 +117,19 @@
         }
       ]
     },
-
+    {
+      "Name": "Dex2oatPerformance",
+      "Actions": [
+        {
+          "Name": "JoinCgroup",
+          "Params":
+          {
+            "Controller": "schedtune",
+            "Path": "background"
+          }
+        }
+      ]
+    },
     {
       "Name": "CpuPolicySpread",
       "Actions": [
diff --git a/libprocessgroup/sched_policy.cpp b/libprocessgroup/sched_policy.cpp
index 1a4196a..169b1d3 100644
--- a/libprocessgroup/sched_policy.cpp
+++ b/libprocessgroup/sched_policy.cpp
@@ -165,27 +165,7 @@
     return 0;
 }
 
-int get_sched_policy(int tid, SchedPolicy* policy) {
-    if (tid == 0) {
-        tid = GetThreadId();
-    }
-
-    std::string group;
-    if (schedboost_enabled()) {
-        if ((getCGroupSubsys(tid, "schedtune", group) < 0) &&
-            (getCGroupSubsys(tid, "cpu", group) < 0)) {
-                LOG(ERROR) << "Failed to find cpu cgroup for tid " << tid;
-                return -1;
-        }
-    }
-    if (group.empty() && cpusets_enabled()) {
-        if (getCGroupSubsys(tid, "cpuset", group) < 0) {
-            LOG(ERROR) << "Failed to find cpuset cgroup for tid " << tid;
-            return -1;
-        }
-    }
-
-    // TODO: replace hardcoded directories
+static int get_sched_policy_from_group(const std::string& group, SchedPolicy* policy) {
     if (group.empty()) {
         *policy = SP_FOREGROUND;
     } else if (group == "foreground") {
@@ -205,6 +185,35 @@
     return 0;
 }
 
+int get_sched_policy(int tid, SchedPolicy* policy) {
+    if (tid == 0) {
+        tid = GetThreadId();
+    }
+
+    std::string group;
+    if (schedboost_enabled()) {
+        if ((getCGroupSubsys(tid, "schedtune", group) < 0) &&
+            (getCGroupSubsys(tid, "cpu", group) < 0)) {
+            LOG(ERROR) << "Failed to find cpu cgroup for tid " << tid;
+            return -1;
+        }
+        // Wipe invalid group to fallback to cpuset
+        if (!group.empty()) {
+            if (get_sched_policy_from_group(group, policy) < 0) {
+                group.clear();
+            } else {
+                return 0;
+            }
+        }
+    }
+
+    if (cpusets_enabled() && getCGroupSubsys(tid, "cpuset", group) < 0) {
+        LOG(ERROR) << "Failed to find cpuset cgroup for tid " << tid;
+        return -1;
+    }
+    return get_sched_policy_from_group(group, policy);
+}
+
 #else
 
 /* Stubs for non-Android targets. */
diff --git a/libprocessgroup/task_profiles.cpp b/libprocessgroup/task_profiles.cpp
index cf74e65..3834f91 100644
--- a/libprocessgroup/task_profiles.cpp
+++ b/libprocessgroup/task_profiles.cpp
@@ -144,30 +144,13 @@
     return true;
 }
 
-bool SetCgroupAction::IsAppDependentPath(const std::string& path) {
-    return path.find("<uid>", 0) != std::string::npos || path.find("<pid>", 0) != std::string::npos;
-}
-
-SetCgroupAction::SetCgroupAction(const CgroupController& c, const std::string& p)
-    : controller_(c), path_(p) {
-    // file descriptors for app-dependent paths can't be cached
-    if (IsAppDependentPath(path_)) {
-        // file descriptor is not cached
-        fd_.reset(FDS_APP_DEPENDENT);
-        return;
-    }
-
-    // file descriptor can be cached later on request
-    fd_.reset(FDS_NOT_CACHED);
-}
-
-void SetCgroupAction::EnableResourceCaching() {
+void CachedFdProfileAction::EnableResourceCaching() {
     std::lock_guard<std::mutex> lock(fd_mutex_);
     if (fd_ != FDS_NOT_CACHED) {
         return;
     }
 
-    std::string tasks_path = controller_.GetTasksFilePath(path_);
+    std::string tasks_path = GetPath();
 
     if (access(tasks_path.c_str(), W_OK) != 0) {
         // file is not accessible
@@ -185,7 +168,7 @@
     fd_ = std::move(fd);
 }
 
-void SetCgroupAction::DropResourceCaching() {
+void CachedFdProfileAction::DropResourceCaching() {
     std::lock_guard<std::mutex> lock(fd_mutex_);
     if (fd_ == FDS_NOT_CACHED) {
         return;
@@ -194,22 +177,59 @@
     fd_.reset(FDS_NOT_CACHED);
 }
 
-bool SetCgroupAction::AddTidToCgroup(int tid, int fd) {
+bool CachedFdProfileAction::IsAppDependentPath(const std::string& path) {
+    return path.find("<uid>", 0) != std::string::npos || path.find("<pid>", 0) != std::string::npos;
+}
+
+void CachedFdProfileAction::InitFd(const std::string& path) {
+    // file descriptors for app-dependent paths can't be cached
+    if (IsAppDependentPath(path)) {
+        // file descriptor is not cached
+        fd_.reset(FDS_APP_DEPENDENT);
+        return;
+    }
+    // file descriptor can be cached later on request
+    fd_.reset(FDS_NOT_CACHED);
+}
+
+SetCgroupAction::SetCgroupAction(const CgroupController& c, const std::string& p)
+    : controller_(c), path_(p) {
+    InitFd(controller_.GetTasksFilePath(path_));
+}
+
+bool SetCgroupAction::AddTidToCgroup(int tid, int fd, const char* controller_name) {
     if (tid <= 0) {
         return true;
     }
 
     std::string value = std::to_string(tid);
 
-    if (TEMP_FAILURE_RETRY(write(fd, value.c_str(), value.length())) < 0) {
-        // If the thread is in the process of exiting, don't flag an error
-        if (errno != ESRCH) {
-            PLOG(ERROR) << "AddTidToCgroup failed to write '" << value << "'; fd=" << fd;
-            return false;
-        }
+    if (TEMP_FAILURE_RETRY(write(fd, value.c_str(), value.length())) == value.length()) {
+        return true;
     }
 
-    return true;
+    // If the thread is in the process of exiting, don't flag an error
+    if (errno == ESRCH) {
+        return true;
+    }
+
+    // ENOSPC is returned when cpuset cgroup that we are joining has no online cpus
+    if (errno == ENOSPC && !strcmp(controller_name, "cpuset")) {
+        // This is an abnormal case happening only in testing, so report it only once
+        static bool empty_cpuset_reported = false;
+
+        if (empty_cpuset_reported) {
+            return true;
+        }
+
+        LOG(ERROR) << "Failed to add task '" << value
+                   << "' into cpuset because all cpus in that cpuset are offline";
+        empty_cpuset_reported = true;
+    } else {
+        PLOG(ERROR) << "AddTidToCgroup failed to write '" << value << "'; fd=" << fd;
+    }
+
+    return false;
 }
 
 bool SetCgroupAction::ExecuteForProcess(uid_t uid, pid_t pid) const {
@@ -219,7 +239,7 @@
         PLOG(WARNING) << "Failed to open " << procs_path;
         return false;
     }
-    if (!AddTidToCgroup(pid, tmp_fd)) {
+    if (!AddTidToCgroup(pid, tmp_fd, controller()->name())) {
         LOG(ERROR) << "Failed to add task into cgroup";
         return false;
     }
@@ -231,7 +251,7 @@
     std::lock_guard<std::mutex> lock(fd_mutex_);
     if (IsFdValid()) {
         // fd is cached, reuse it
-        if (!AddTidToCgroup(tid, fd_)) {
+        if (!AddTidToCgroup(tid, fd_, controller()->name())) {
             LOG(ERROR) << "Failed to add task into cgroup";
             return false;
         }
@@ -253,10 +273,10 @@
     std::string tasks_path = controller()->GetTasksFilePath(path_);
     unique_fd tmp_fd(TEMP_FAILURE_RETRY(open(tasks_path.c_str(), O_WRONLY | O_CLOEXEC)));
     if (tmp_fd < 0) {
-        PLOG(WARNING) << "Failed to open " << tasks_path << ": " << strerror(errno);
+        PLOG(WARNING) << "Failed to open " << tasks_path;
         return false;
     }
-    if (!AddTidToCgroup(tid, tmp_fd)) {
+    if (!AddTidToCgroup(tid, tmp_fd, controller()->name())) {
         LOG(ERROR) << "Failed to add task into cgroup";
         return false;
     }
@@ -264,37 +284,73 @@
     return true;
 }
 
-bool WriteFileAction::ExecuteForProcess(uid_t uid, pid_t pid) const {
-    std::string filepath(filepath_), value(value_);
+WriteFileAction::WriteFileAction(const std::string& path, const std::string& value,
+                                 bool logfailures)
+    : path_(path), value_(value), logfailures_(logfailures) {
+    InitFd(path_);
+}
 
-    filepath = StringReplace(filepath, "<uid>", std::to_string(uid), true);
-    filepath = StringReplace(filepath, "<pid>", std::to_string(pid), true);
-    value = StringReplace(value, "<uid>", std::to_string(uid), true);
-    value = StringReplace(value, "<pid>", std::to_string(pid), true);
+bool WriteFileAction::WriteValueToFile(const std::string& value, const std::string& path,
+                                       bool logfailures) {
+    // Use WriteStringToFd instead of WriteStringToFile because the latter will open file with
+    // O_TRUNC which causes kernfs_mutex contention
+    unique_fd tmp_fd(TEMP_FAILURE_RETRY(open(path.c_str(), O_WRONLY | O_CLOEXEC)));
 
-    if (!WriteStringToFile(value, filepath)) {
-        if (logfailures_) PLOG(ERROR) << "Failed to write '" << value << "' to " << filepath;
+    if (tmp_fd < 0) {
+        if (logfailures) PLOG(WARNING) << "Failed to open " << path;
+        return false;
+    }
+
+    if (!WriteStringToFd(value, tmp_fd)) {
+        if (logfailures) PLOG(ERROR) << "Failed to write '" << value << "' to " << path;
         return false;
     }
 
     return true;
 }
 
+bool WriteFileAction::ExecuteForProcess(uid_t uid, pid_t pid) const {
+    std::lock_guard<std::mutex> lock(fd_mutex_);
+    std::string value(value_);
+    std::string path(path_);
+
+    value = StringReplace(value, "<uid>", std::to_string(uid), true);
+    value = StringReplace(value, "<pid>", std::to_string(pid), true);
+    path = StringReplace(path, "<uid>", std::to_string(uid), true);
+    path = StringReplace(path, "<pid>", std::to_string(pid), true);
+
+    return WriteValueToFile(value, path, logfailures_);
+}
+
 bool WriteFileAction::ExecuteForTask(int tid) const {
-    std::string filepath(filepath_), value(value_);
+    std::lock_guard<std::mutex> lock(fd_mutex_);
+    std::string value(value_);
     int uid = getuid();
 
-    filepath = StringReplace(filepath, "<uid>", std::to_string(uid), true);
-    filepath = StringReplace(filepath, "<pid>", std::to_string(tid), true);
     value = StringReplace(value, "<uid>", std::to_string(uid), true);
     value = StringReplace(value, "<pid>", std::to_string(tid), true);
 
-    if (!WriteStringToFile(value, filepath)) {
-        if (logfailures_) PLOG(ERROR) << "Failed to write '" << value << "' to " << filepath;
+    if (IsFdValid()) {
+        // fd is cached, reuse it
+        if (!WriteStringToFd(value, fd_)) {
+            if (logfailures_) PLOG(ERROR) << "Failed to write '" << value << "' to " << path_;
+            return false;
+        }
+        return true;
+    }
+
+    if (fd_ == FDS_INACCESSIBLE) {
+        // no permissions to access the file, ignore
+        return true;
+    }
+
+    if (fd_ == FDS_APP_DEPENDENT) {
+        // application-dependent path can't be used with tid
+        PLOG(ERROR) << "Application profile can't be applied to a thread";
         return false;
     }
 
-    return true;
+    return WriteValueToFile(value, path_, logfailures_);
 }
 
 bool ApplyProfileAction::ExecuteForProcess(uid_t uid, pid_t pid) const {
diff --git a/libprocessgroup/task_profiles.h b/libprocessgroup/task_profiles.h
index 25a84b0..278892d 100644
--- a/libprocessgroup/task_profiles.h
+++ b/libprocessgroup/task_profiles.h
@@ -108,50 +108,67 @@
     std::string value_;
 };
 
-// Set cgroup profile element
-class SetCgroupAction : public ProfileAction {
+// Abstract profile element for cached fd
+class CachedFdProfileAction : public ProfileAction {
   public:
-    SetCgroupAction(const CgroupController& c, const std::string& p);
-
-    virtual bool ExecuteForProcess(uid_t uid, pid_t pid) const;
-    virtual bool ExecuteForTask(int tid) const;
     virtual void EnableResourceCaching();
     virtual void DropResourceCaching();
 
-    const CgroupController* controller() const { return &controller_; }
-    std::string path() const { return path_; }
-
-  private:
+  protected:
     enum FdState {
         FDS_INACCESSIBLE = -1,
         FDS_APP_DEPENDENT = -2,
         FDS_NOT_CACHED = -3,
     };
 
-    CgroupController controller_;
-    std::string path_;
     android::base::unique_fd fd_;
     mutable std::mutex fd_mutex_;
 
     static bool IsAppDependentPath(const std::string& path);
-    static bool AddTidToCgroup(int tid, int fd);
 
+    void InitFd(const std::string& path);
     bool IsFdValid() const { return fd_ > FDS_INACCESSIBLE; }
+
+    virtual const std::string GetPath() const = 0;
 };
 
-// Write to file action
-class WriteFileAction : public ProfileAction {
+// Set cgroup profile element
+class SetCgroupAction : public CachedFdProfileAction {
   public:
-    WriteFileAction(const std::string& filepath, const std::string& value,
-                    bool logfailures) noexcept
-        : filepath_(filepath), value_(value), logfailures_(logfailures) {}
+    SetCgroupAction(const CgroupController& c, const std::string& p);
 
     virtual bool ExecuteForProcess(uid_t uid, pid_t pid) const;
     virtual bool ExecuteForTask(int tid) const;
 
+    const CgroupController* controller() const { return &controller_; }
+
+  protected:
+    const std::string GetPath() const override { return controller_.GetTasksFilePath(path_); }
+
   private:
-    std::string filepath_, value_;
+    CgroupController controller_;
+    std::string path_;
+
+    static bool AddTidToCgroup(int tid, int fd, const char* controller_name);
+};
+
+// Write to file action
+class WriteFileAction : public CachedFdProfileAction {
+  public:
+    WriteFileAction(const std::string& path, const std::string& value, bool logfailures);
+
+    virtual bool ExecuteForProcess(uid_t uid, pid_t pid) const;
+    virtual bool ExecuteForTask(int tid) const;
+
+  protected:
+    const std::string GetPath() const override { return path_; }
+
+  private:
+    std::string path_, value_;
     bool logfailures_;
+
+    static bool WriteValueToFile(const std::string& value, const std::string& path,
+                                 bool logfailures);
 };
 
 class TaskProfile {
diff --git a/libutils/Threads.cpp b/libutils/Threads.cpp
index 540dcf4..3bf5779 100644
--- a/libutils/Threads.cpp
+++ b/libutils/Threads.cpp
@@ -86,8 +86,10 @@
 
         // A new thread will be in its parent's sched group by default,
         // so we just need to handle the background case.
+        // currently set to system_background group which is different
+        // from background group for app.
         if (prio >= ANDROID_PRIORITY_BACKGROUND) {
-            SetTaskProfiles(0, {"SCHED_SP_BACKGROUND"}, true);
+            SetTaskProfiles(0, {"SCHED_SP_SYSTEM"}, true);
         }
 
         if (name) {
@@ -313,12 +315,9 @@
     }
 
     if (pri >= ANDROID_PRIORITY_BACKGROUND) {
-        rc = SetTaskProfiles(tid, {"SCHED_SP_BACKGROUND"}, true) ? 0 : -1;
+        rc = SetTaskProfiles(tid, {"SCHED_SP_SYSTEM"}, true) ? 0 : -1;
     } else if (curr_pri >= ANDROID_PRIORITY_BACKGROUND) {
-        SchedPolicy policy = SP_FOREGROUND;
-        // Change to the sched policy group of the process.
-        get_sched_policy(getpid(), &policy);
-        rc = SetTaskProfiles(tid, {get_sched_policy_profile_name(policy)}, true) ? 0 : -1;
+        rc = SetTaskProfiles(tid, {"SCHED_SP_FOREGROUND"}, true) ? 0 : -1;
     }
 
     if (rc) {
diff --git a/llkd/libllkd.cpp b/llkd/libllkd.cpp
index c4c58ee..42602e9 100644
--- a/llkd/libllkd.cpp
+++ b/llkd/libllkd.cpp
@@ -1283,8 +1283,7 @@
     llkEnableSysrqT &= !llkLowRam;
     if (debuggable) {
         llkEnableSysrqT |= llkCheckEng(LLK_ENABLE_SYSRQ_T_PROPERTY);
-        if (!LLK_ENABLE_DEFAULT) {  // NB: default is currently true ...
-            llkEnable |= llkCheckEng(LLK_ENABLE_PROPERTY);
+        if (!LLK_ENABLE_DEFAULT) {
             khtEnable |= llkCheckEng(KHT_ENABLE_PROPERTY);
         }
     }
diff --git a/llkd/llkd-debuggable.rc b/llkd/llkd-debuggable.rc
index 4b11b1c..c075609 100644
--- a/llkd/llkd-debuggable.rc
+++ b/llkd/llkd-debuggable.rc
@@ -1,5 +1,5 @@
 on property:ro.debuggable=1
-    setprop llk.enable ${ro.llk.enable:-1}
+    setprop llk.enable ${ro.llk.enable:-0}
     setprop khungtask.enable ${ro.khungtask.enable:-1}
 
 on property:ro.llk.enable=eng
diff --git a/llkd/tests/llkd_test.cpp b/llkd/tests/llkd_test.cpp
index 475512c..8eb9b00 100644
--- a/llkd/tests/llkd_test.cpp
+++ b/llkd/tests/llkd_test.cpp
@@ -69,13 +69,9 @@
 seconds llkdSleepPeriod(char state) {
     auto default_eng = android::base::GetProperty(LLK_ENABLE_PROPERTY, "eng") == "eng";
     auto default_enable = LLK_ENABLE_DEFAULT;
-    if (!LLK_ENABLE_DEFAULT && default_eng &&
-        android::base::GetBoolProperty("ro.debuggable", false)) {
-        default_enable = true;
-    }
     default_enable = android::base::GetBoolProperty(LLK_ENABLE_PROPERTY, default_enable);
     if (default_eng) {
-        GTEST_LOG_INFO << LLK_ENABLE_PROPERTY " defaults to \"eng\" thus "
+        GTEST_LOG_INFO << LLK_ENABLE_PROPERTY " defaults to "
                        << (default_enable ? "true" : "false") << "\n";
     }
     // Hail Mary hope is unconfigured.
@@ -108,10 +104,6 @@
         rest();
     }
     default_enable = LLK_ENABLE_DEFAULT;
-    if (!LLK_ENABLE_DEFAULT && (android::base::GetProperty(LLK_ENABLE_PROPERTY, "eng") == "eng") &&
-        android::base::GetBoolProperty("ro.debuggable", false)) {
-        default_enable = true;
-    }
     default_enable = android::base::GetBoolProperty(LLK_ENABLE_PROPERTY, default_enable);
     if (default_enable) {
         execute("start llkd-1");
diff --git a/rootdir/init.rc b/rootdir/init.rc
index 376a678..5116c0f 100644
--- a/rootdir/init.rc
+++ b/rootdir/init.rc
@@ -78,8 +78,8 @@
     mkdir /dev/boringssl 0755 root root
     mkdir /dev/boringssl/selftest 0755 root root
 
-    # Mount tracefs
-    mount tracefs tracefs /sys/kernel/tracing
+    # Mount tracefs (with GID=AID_READTRACEFS)
+    mount tracefs tracefs /sys/kernel/tracing gid=3012
 
     # create sys dirctory
     mkdir /dev/sys 0755 system system
@@ -155,6 +155,7 @@
     mkdir /dev/cpuctl/rt
     mkdir /dev/cpuctl/system
     mkdir /dev/cpuctl/system-background
+    mkdir /dev/cpuctl/dex2oat
     chown system system /dev/cpuctl
     chown system system /dev/cpuctl/foreground
     chown system system /dev/cpuctl/background
@@ -162,6 +163,7 @@
     chown system system /dev/cpuctl/rt
     chown system system /dev/cpuctl/system
     chown system system /dev/cpuctl/system-background
+    chown system system /dev/cpuctl/dex2oat
     chown system system /dev/cpuctl/tasks
     chown system system /dev/cpuctl/foreground/tasks
     chown system system /dev/cpuctl/background/tasks
@@ -169,6 +171,7 @@
     chown system system /dev/cpuctl/rt/tasks
     chown system system /dev/cpuctl/system/tasks
     chown system system /dev/cpuctl/system-background/tasks
+    chown system system /dev/cpuctl/dex2oat/tasks
     chmod 0664 /dev/cpuctl/tasks
     chmod 0664 /dev/cpuctl/foreground/tasks
     chmod 0664 /dev/cpuctl/background/tasks
@@ -176,6 +179,7 @@
     chmod 0664 /dev/cpuctl/rt/tasks
     chmod 0664 /dev/cpuctl/system/tasks
     chmod 0664 /dev/cpuctl/system-background/tasks
+    chmod 0664 /dev/cpuctl/dex2oat/tasks
 
     # Create a cpu group for NNAPI HAL processes
     mkdir /dev/cpuctl/nnapi-hal
@@ -590,9 +594,23 @@
     # Load trusted keys from dm-verity protected partitions
     exec -- /system/bin/fsverity_init --load-verified-keys
 
+# Only enable the bootreceiver tracing instance for kernels 5.10 and above.
+on late-fs && property:ro.kernel.version=4.9
+    setprop bootreceiver.enable 0
+on late-fs && property:ro.kernel.version=4.14
+    setprop bootreceiver.enable 0
+on late-fs && property:ro.kernel.version=4.19
+    setprop bootreceiver.enable 0
+on late-fs && property:ro.kernel.version=5.4
+    setprop bootreceiver.enable 0
+on late-fs
+    # Bootreceiver tracing instance is enabled by default.
+    setprop bootreceiver.enable ${bootreceiver.enable:-1}
+
+on property:ro.product.cpu.abilist64=* && property:bootreceiver.enable=1
     # Set up a tracing instance for system_server to monitor error_report_end events.
     # These are sent by kernel tools like KASAN and KFENCE when a memory corruption
-    # is detected.
+    # is detected. This is only needed for 64-bit systems.
     mkdir /sys/kernel/tracing/instances/bootreceiver 0700 system system
     restorecon_recursive /sys/kernel/tracing/instances/bootreceiver
     write /sys/kernel/tracing/instances/bootreceiver/buffer_size_kb 1
diff --git a/trusty/keymaster/Android.bp b/trusty/keymaster/Android.bp
index ff6460d..cf056f0 100644
--- a/trusty/keymaster/Android.bp
+++ b/trusty/keymaster/Android.bp
@@ -120,6 +120,7 @@
     ],
     required: [
         "android.hardware.hardware_keystore.xml",
+        "RemoteProvisioner",
     ],
 }
 
diff --git a/trusty/keymaster/ipc/trusty_keymaster_ipc.cpp b/trusty/keymaster/ipc/trusty_keymaster_ipc.cpp
index 0956fe6..2d44009 100644
--- a/trusty/keymaster/ipc/trusty_keymaster_ipc.cpp
+++ b/trusty/keymaster/ipc/trusty_keymaster_ipc.cpp
@@ -25,6 +25,8 @@
 #include <unistd.h>
 
 #include <algorithm>
+#include <variant>
+#include <vector>
 
 #include <log/log.h>
 #include <trusty/tipc.h>
@@ -46,8 +48,27 @@
     return 0;
 }
 
-int trusty_keymaster_call(uint32_t cmd, void* in, uint32_t in_size, uint8_t* out,
-                          uint32_t* out_size) {
+class VectorEraser {
+  public:
+    VectorEraser(std::vector<uint8_t>* v) : _v(v) {}
+    ~VectorEraser() {
+        if (_v) {
+            std::fill(const_cast<volatile uint8_t*>(_v->data()),
+                      const_cast<volatile uint8_t*>(_v->data() + _v->size()), 0);
+        }
+    }
+    void disarm() { _v = nullptr; }
+    VectorEraser(const VectorEraser&) = delete;
+    VectorEraser& operator=(const VectorEraser&) = delete;
+    VectorEraser(VectorEraser&& other) = delete;
+    VectorEraser& operator=(VectorEraser&&) = delete;
+
+  private:
+    std::vector<uint8_t>* _v;
+};
+
+std::variant<int, std::vector<uint8_t>> trusty_keymaster_call_2(uint32_t cmd, void* in,
+                                                                uint32_t in_size) {
     if (handle_ < 0) {
         ALOGE("not connected\n");
         return -EINVAL;
@@ -70,15 +91,38 @@
         ALOGE("failed to send cmd (%d) to %s: %s\n", cmd, KEYMASTER_PORT, strerror(errno));
         return -errno;
     }
-    size_t out_max_size = *out_size;
-    *out_size = 0;
+
+    std::vector<uint8_t> out(TRUSTY_KEYMASTER_RECV_BUF_SIZE);
+    VectorEraser out_eraser(&out);
+    uint8_t* write_pos = out.data();
+    uint8_t* out_end = out.data() + out.size();
+
     struct iovec iov[2];
     struct keymaster_message header;
     iov[0] = {.iov_base = &header, .iov_len = sizeof(struct keymaster_message)};
     while (true) {
-        iov[1] = {.iov_base = out + *out_size,
-                  .iov_len = std::min<uint32_t>(KEYMASTER_MAX_BUFFER_LENGTH,
-                                                out_max_size - *out_size)};
+        if (out_end - write_pos < KEYMASTER_MAX_BUFFER_LENGTH) {
+            // In stead of using std::vector.resize(), allocate a new one to have chance
+            // at zeroing the old buffer.
+            std::vector<uint8_t> new_out(out.size() + KEYMASTER_MAX_BUFFER_LENGTH);
+            // After the swap below this erases the old out buffer.
+            VectorEraser new_out_eraser(&new_out);
+            std::copy(out.data(), write_pos, new_out.begin());
+
+            auto write_offset = write_pos - out.data();
+
+            std::swap(new_out, out);
+
+            write_pos = out.data() + write_offset;
+            out_end = out.data() + out.size();
+        }
+        size_t buffer_size = 0;
+        if (__builtin_sub_overflow(reinterpret_cast<uintptr_t>(out_end),
+                                   reinterpret_cast<uintptr_t>(write_pos), &buffer_size)) {
+            return -EOVERFLOW;
+        }
+        iov[1] = {.iov_base = write_pos, .iov_len = buffer_size};
+
         rc = readv(handle_, iov, 2);
         if (rc < 0) {
             ALOGE("failed to retrieve response for cmd (%d) to %s: %s\n", cmd, KEYMASTER_PORT,
@@ -95,13 +139,36 @@
             ALOGE("invalid command (%d)", header.cmd);
             return -EINVAL;
         }
-        *out_size += ((size_t)rc - sizeof(struct keymaster_message));
+        write_pos += ((size_t)rc - sizeof(struct keymaster_message));
         if (header.cmd & KEYMASTER_STOP_BIT) {
             break;
         }
     }
 
-    return rc;
+    out.resize(write_pos - out.data());
+    out_eraser.disarm();
+    return out;
+}
+
+int trusty_keymaster_call(uint32_t cmd, void* in, uint32_t in_size, uint8_t* out,
+                          uint32_t* out_size) {
+    auto result = trusty_keymaster_call_2(cmd, in, in_size);
+    if (auto out_buffer = std::get_if<std::vector<uint8_t>>(&result)) {
+        if (out_buffer->size() <= *out_size) {
+            std::copy(out_buffer->begin(), out_buffer->end(), out);
+            std::fill(const_cast<volatile uint8_t*>(&*out_buffer->begin()),
+                      const_cast<volatile uint8_t*>(&*out_buffer->end()), 0);
+
+            *out_size = out_buffer->size();
+            return 0;
+        } else {
+            ALOGE("Message was to large (%zu) for the provided buffer (%u)", out_buffer->size(),
+                  *out_size);
+            return -EMSGSIZE;
+        }
+    } else {
+        return std::get<int>(result);
+    }
 }
 
 void trusty_keymaster_disconnect() {
@@ -155,28 +222,27 @@
     req.Serialize(send_buf, send_buf + req_size);
 
     // Send it
-    uint8_t recv_buf[TRUSTY_KEYMASTER_RECV_BUF_SIZE];
-    keymaster::Eraser recv_buf_eraser(recv_buf, TRUSTY_KEYMASTER_RECV_BUF_SIZE);
-    uint32_t rsp_size = TRUSTY_KEYMASTER_RECV_BUF_SIZE;
-    int rc = trusty_keymaster_call(command, send_buf, req_size, recv_buf, &rsp_size);
-    if (rc < 0) {
+    auto response = trusty_keymaster_call_2(command, send_buf, req_size);
+    if (auto response_buffer = std::get_if<std::vector<uint8_t>>(&response)) {
+        keymaster::Eraser response_buffer_erasor(response_buffer->data(), response_buffer->size());
+        ALOGV("Received %zu byte response\n", response_buffer->size());
+
+        const uint8_t* p = response_buffer->data();
+        if (!rsp->Deserialize(&p, p + response_buffer->size())) {
+            ALOGE("Error deserializing response of size %zu\n", response_buffer->size());
+            return KM_ERROR_UNKNOWN_ERROR;
+        } else if (rsp->error != KM_ERROR_OK) {
+            ALOGE("Response of size %zu contained error code %d\n", response_buffer->size(),
+                  (int)rsp->error);
+        }
+        return rsp->error;
+    } else {
+        auto rc = std::get<int>(response);
         // Reset the connection on tipc error
         trusty_keymaster_disconnect();
         trusty_keymaster_connect();
         ALOGE("tipc error: %d\n", rc);
         // TODO(swillden): Distinguish permanent from transient errors and set error_ appropriately.
         return translate_error(rc);
-    } else {
-        ALOGV("Received %d byte response\n", rsp_size);
     }
-
-    const uint8_t* p = recv_buf;
-    if (!rsp->Deserialize(&p, p + rsp_size)) {
-        ALOGE("Error deserializing response of size %d\n", (int)rsp_size);
-        return KM_ERROR_UNKNOWN_ERROR;
-    } else if (rsp->error != KM_ERROR_OK) {
-        ALOGE("Response of size %d contained error code %d\n", (int)rsp_size, (int)rsp->error);
-        return rsp->error;
-    }
-    return rsp->error;
 }
diff --git a/trusty/storage/interface/include/trusty/interface/storage.h b/trusty/storage/interface/include/trusty/interface/storage.h
index b196d88..3f1dcb8 100644
--- a/trusty/storage/interface/include/trusty/interface/storage.h
+++ b/trusty/storage/interface/include/trusty/interface/storage.h
@@ -112,26 +112,30 @@
 
 /**
  * enum storage_msg_flag - protocol-level flags in struct storage_msg
- * @STORAGE_MSG_FLAG_BATCH:             if set, command belongs to a batch transaction.
- *                                      No response will be sent by the server until
- *                                      it receives a command with this flag unset, at
- *                                      which point a cummulative result for all messages
- *                                      sent with STORAGE_MSG_FLAG_BATCH will be sent.
- *                                      This is only supported by the non-secure disk proxy
- *                                      server.
- * @STORAGE_MSG_FLAG_PRE_COMMIT:        if set, indicates that server need to commit
- *                                      pending changes before processing this message.
- * @STORAGE_MSG_FLAG_POST_COMMIT:       if set, indicates that server need to commit
- *                                      pending changes after processing this message.
- * @STORAGE_MSG_FLAG_TRANSACT_COMPLETE: if set, indicates that server need to commit
- *                                      current transaction after processing this message.
- *                                      It is an alias for STORAGE_MSG_FLAG_POST_COMMIT.
+ * @STORAGE_MSG_FLAG_BATCH:                 if set, command belongs to a batch transaction.
+ *                                          No response will be sent by the server until
+ *                                          it receives a command with this flag unset, at
+ *                                          which point a cumulative result for all messages
+ *                                          sent with STORAGE_MSG_FLAG_BATCH will be sent.
+ *                                          This is only supported by the non-secure disk proxy
+ *                                          server.
+ * @STORAGE_MSG_FLAG_PRE_COMMIT:            if set, indicates that server need to commit
+ *                                          pending changes before processing this message.
+ * @STORAGE_MSG_FLAG_POST_COMMIT:           if set, indicates that server need to commit
+ *                                          pending changes after processing this message.
+ * @STORAGE_MSG_FLAG_TRANSACT_COMPLETE:     if set, indicates that server need to commit
+ *                                          current transaction after processing this message.
+ *                                          It is an alias for STORAGE_MSG_FLAG_POST_COMMIT.
+ * @STORAGE_MSG_FLAG_PRE_COMMIT_CHECKPOINT: if set, indicates that server needs to ensure
+ *                                          that there is not a pending checkpoint for
+ *                                          userdata before processing this message.
  */
 enum storage_msg_flag {
-	STORAGE_MSG_FLAG_BATCH = 0x1,
-	STORAGE_MSG_FLAG_PRE_COMMIT = 0x2,
-	STORAGE_MSG_FLAG_POST_COMMIT = 0x4,
-	STORAGE_MSG_FLAG_TRANSACT_COMPLETE = STORAGE_MSG_FLAG_POST_COMMIT,
+    STORAGE_MSG_FLAG_BATCH = 0x1,
+    STORAGE_MSG_FLAG_PRE_COMMIT = 0x2,
+    STORAGE_MSG_FLAG_POST_COMMIT = 0x4,
+    STORAGE_MSG_FLAG_TRANSACT_COMPLETE = STORAGE_MSG_FLAG_POST_COMMIT,
+    STORAGE_MSG_FLAG_PRE_COMMIT_CHECKPOINT = 0x8,
 };
 
 /*
diff --git a/trusty/storage/proxy/Android.bp b/trusty/storage/proxy/Android.bp
index d67089f..38d8685 100644
--- a/trusty/storage/proxy/Android.bp
+++ b/trusty/storage/proxy/Android.bp
@@ -23,6 +23,7 @@
     vendor: true,
 
     srcs: [
+        "checkpoint_handling.cpp",
         "ipc.c",
         "rpmb.c",
         "storage.c",
@@ -30,12 +31,14 @@
     ],
 
     shared_libs: [
+        "libbase",
         "liblog",
         "libhardware_legacy",
     ],
     header_libs: ["libcutils_headers"],
 
     static_libs: [
+        "libfstab",
         "libtrustystorageinterface",
         "libtrusty",
     ],
diff --git a/trusty/storage/proxy/checkpoint_handling.cpp b/trusty/storage/proxy/checkpoint_handling.cpp
new file mode 100644
index 0000000..6c2fd36
--- /dev/null
+++ b/trusty/storage/proxy/checkpoint_handling.cpp
@@ -0,0 +1,77 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#include "checkpoint_handling.h"
+#include "log.h"
+
+#include <fstab/fstab.h>
+#include <cstring>
+#include <string>
+
+namespace {
+
+bool checkpointingDoneForever = false;
+
+}  // namespace
+
+int is_data_checkpoint_active(bool* active) {
+    if (!active) {
+        ALOGE("active out parameter is null");
+        return 0;
+    }
+
+    *active = false;
+
+    if (checkpointingDoneForever) {
+        return 0;
+    }
+
+    android::fs_mgr::Fstab procMounts;
+    bool success = android::fs_mgr::ReadFstabFromFile("/proc/mounts", &procMounts);
+    if (!success) {
+        ALOGE("Could not parse /proc/mounts\n");
+        /* Really bad. Tell the caller to abort the write. */
+        return -1;
+    }
+
+    android::fs_mgr::FstabEntry* dataEntry =
+            android::fs_mgr::GetEntryForMountPoint(&procMounts, "/data");
+    if (dataEntry == NULL) {
+        ALOGE("/data is not mounted yet\n");
+        return 0;
+    }
+
+    /* We can't handle e.g., ext4. Nothing we can do about it for now. */
+    if (dataEntry->fs_type != "f2fs") {
+        ALOGW("Checkpoint status not supported for filesystem %s\n", dataEntry->fs_type.c_str());
+        checkpointingDoneForever = true;
+        return 0;
+    }
+
+    /*
+     * The data entry looks like "... blah,checkpoint=disable:0,blah ...".
+     * checkpoint=disable means checkpointing is on (yes, arguably reversed).
+     */
+    size_t checkpointPos = dataEntry->fs_options.find("checkpoint=disable");
+    if (checkpointPos == std::string::npos) {
+        /* Assumption is that once checkpointing turns off, it stays off */
+        checkpointingDoneForever = true;
+    } else {
+        *active = true;
+    }
+
+    return 0;
+}
diff --git a/trusty/storage/proxy/checkpoint_handling.h b/trusty/storage/proxy/checkpoint_handling.h
new file mode 100644
index 0000000..f1bf27c
--- /dev/null
+++ b/trusty/storage/proxy/checkpoint_handling.h
@@ -0,0 +1,37 @@
+/*
+ * Copyright (C) 2021 The Android Open Source Project
+ *
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ *      http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ */
+
+#pragma once
+
+#include <stdbool.h>
+
+#ifdef __cplusplus
+extern "C" {
+#endif
+
+/**
+ * is_data_checkpoint_active() - Check for an active, uncommitted checkpoint of
+ * /data. If a checkpoint is active, storage should not commit any
+ * rollback-protected writes to /data.
+ * @active: Out parameter that will be set to the result of the check.
+ *
+ * Return: 0 if active was set and is valid, non-zero otherwise.
+ */
+int is_data_checkpoint_active(bool* active);
+
+#ifdef __cplusplus
+}
+#endif
diff --git a/trusty/storage/proxy/proxy.c b/trusty/storage/proxy/proxy.c
index e230941..c690a28 100644
--- a/trusty/storage/proxy/proxy.c
+++ b/trusty/storage/proxy/proxy.c
@@ -26,6 +26,7 @@
 
 #include <cutils/android_filesystem_config.h>
 
+#include "checkpoint_handling.h"
 #include "ipc.h"
 #include "log.h"
 #include "rpmb.h"
@@ -130,6 +131,21 @@
         }
     }
 
+    if (msg->flags & STORAGE_MSG_FLAG_PRE_COMMIT_CHECKPOINT) {
+        bool is_checkpoint_active = false;
+
+        rc = is_data_checkpoint_active(&is_checkpoint_active);
+        if (rc != 0) {
+            ALOGE("is_data_checkpoint_active failed in an unexpected way. Aborting.\n");
+            msg->result = STORAGE_ERR_GENERIC;
+            return ipc_respond(msg, NULL, 0);
+        } else if (is_checkpoint_active) {
+            ALOGE("Checkpoint in progress, dropping write ...\n");
+            msg->result = STORAGE_ERR_GENERIC;
+            return ipc_respond(msg, NULL, 0);
+        }
+    }
+
     switch (msg->cmd) {
         case STORAGE_FILE_DELETE:
             rc = storage_file_delete(msg, req, req_len);
diff --git a/trusty/storage/proxy/rpmb.c b/trusty/storage/proxy/rpmb.c
index b59fb67..f059935 100644
--- a/trusty/storage/proxy/rpmb.c
+++ b/trusty/storage/proxy/rpmb.c
@@ -16,7 +16,10 @@
 
 #include <errno.h>
 #include <fcntl.h>
+#include <scsi/scsi.h>
+#include <scsi/scsi_proto.h>
 #include <scsi/sg.h>
+#include <stdbool.h>
 #include <stdint.h>
 #include <stdio.h>
 #include <stdlib.h>
@@ -55,6 +58,17 @@
 #define MMC_BLOCK_SIZE 512
 
 /*
+ * Number of retry attempts when an RPMB authenticated write triggers a UNIT
+ * ATTENTION
+ */
+#define UFS_RPMB_WRITE_RETRY_COUNT 1
+/*
+ * Number of retry attempts when an RPMB read operation triggers a UNIT
+ * ATTENTION
+ */
+#define UFS_RPMB_READ_RETRY_COUNT 3
+
+/*
  * There should be no timeout for security protocol ioctl call, so we choose a
  * large number for timeout.
  * 20000 millisecs == 20 seconds
@@ -104,21 +118,62 @@
 
 static const char* UFS_WAKE_LOCK_NAME = "ufs_seq_wakelock";
 
-#ifdef RPMB_DEBUG
-
-static void print_buf(const char* prefix, const uint8_t* buf, size_t size) {
+/**
+ * log_buf - Log a byte buffer to the android log.
+ * @priority: One of ANDROID_LOG_* priority levels from android_LogPriority in
+ *            android/log.h
+ * @prefix:   A null-terminated string that identifies this buffer. Must be less
+ *            than 128 bytes.
+ * @buf:      Buffer to dump.
+ * @size:     Length of @buf in bytes.
+ */
+#define LOG_BUF_SIZE 256
+static int log_buf(int priority, const char* prefix, const uint8_t* buf, size_t size) {
+    int rc;
     size_t i;
+    char line[LOG_BUF_SIZE] = {0};
+    char* cur = line;
 
-    printf("%s @%p [%zu]", prefix, buf, size);
-    for (i = 0; i < size; i++) {
-        if (i && i % 32 == 0) printf("\n%*s", (int)strlen(prefix), "");
-        printf(" %02x", buf[i]);
+    rc = snprintf(line, LOG_BUF_SIZE, "%s @%p [%zu]", prefix, buf, size);
+    if (rc < 0 || rc >= LOG_BUF_SIZE) {
+        goto err;
     }
-    printf("\n");
-    fflush(stdout);
-}
+    cur += rc;
+    for (i = 0; i < size; i++) {
+        if (i % 32 == 0) {
+            /*
+             * Flush the line out to the log after we have printed 32 bytes
+             * (also flushes the header line on the first iteration and sets up
+             * for printing the buffer itself)
+             */
+            LOG_PRI(priority, LOG_TAG, "%s", line);
+            memset(line, 0, LOG_BUF_SIZE);
+            cur = line;
+            /* Shift output over by the length of the prefix */
+            rc = snprintf(line, LOG_BUF_SIZE, "%*s", (int)strlen(prefix), "");
+            if (rc < 0 || rc >= LOG_BUF_SIZE) {
+                goto err;
+            }
+            cur += rc;
+        }
+        rc = snprintf(cur, LOG_BUF_SIZE - (cur - line), "%02x ", buf[i]);
+        if (rc < 0 || rc >= LOG_BUF_SIZE - (cur - line)) {
+            goto err;
+        }
+        cur += rc;
+    }
+    LOG_PRI(priority, LOG_TAG, "%s", line);
 
-#endif
+    return 0;
+
+err:
+    if (rc < 0) {
+        return rc;
+    } else {
+        ALOGE("log_buf prefix was too long");
+        return -1;
+    }
+}
 
 static void set_sg_io_hdr(sg_io_hdr_t* io_hdrp, int dxfer_direction, unsigned char cmd_len,
                           unsigned char mx_sb_len, unsigned int dxfer_len, void* dxferp,
@@ -135,6 +190,137 @@
     io_hdrp->timeout = TIMEOUT;
 }
 
+/**
+ * enum scsi_result - Results of checking the SCSI status and sense buffer
+ *
+ * @SCSI_RES_OK:    SCSI status and sense are good
+ * @SCSI_RES_ERR:   SCSI status or sense contain an unhandled error
+ * @SCSI_RES_RETRY: SCSI sense buffer contains a status that indicates that the
+ *                  command should be retried
+ */
+enum scsi_result {
+    SCSI_RES_OK = 0,
+    SCSI_RES_ERR,
+    SCSI_RES_RETRY,
+};
+
+static enum scsi_result check_scsi_sense(const uint8_t* sense_buf, size_t len) {
+    uint8_t response_code = 0;
+    uint8_t sense_key = 0;
+    uint8_t additional_sense_code = 0;
+    uint8_t additional_sense_code_qualifier = 0;
+    uint8_t additional_length = 0;
+
+    if (!sense_buf || len == 0) {
+        ALOGE("Invalid SCSI sense buffer, length: %zu\n", len);
+        return SCSI_RES_ERR;
+    }
+
+    response_code = 0x7f & sense_buf[0];
+
+    if (response_code < 0x70 || response_code > 0x73) {
+        ALOGE("Invalid SCSI sense response code: %hhu\n", response_code);
+        return SCSI_RES_ERR;
+    }
+
+    if (response_code >= 0x72) {
+        /* descriptor format, SPC-6 4.4.2 */
+        if (len > 1) {
+            sense_key = 0xf & sense_buf[1];
+        }
+        if (len > 2) {
+            additional_sense_code = sense_buf[2];
+        }
+        if (len > 3) {
+            additional_sense_code_qualifier = sense_buf[3];
+        }
+        if (len > 7) {
+            additional_length = sense_buf[7];
+        }
+    } else {
+        /* fixed format, SPC-6 4.4.3 */
+        if (len > 2) {
+            sense_key = 0xf & sense_buf[2];
+        }
+        if (len > 7) {
+            additional_length = sense_buf[7];
+        }
+        if (len > 12) {
+            additional_sense_code = sense_buf[12];
+        }
+        if (len > 13) {
+            additional_sense_code_qualifier = sense_buf[13];
+        }
+    }
+
+    switch (sense_key) {
+        case NO_SENSE:
+        case 0x0f: /* COMPLETED, not present in kernel headers */
+            ALOGD("SCSI success with sense data: key=%hhu, asc=%hhu, ascq=%hhu\n", sense_key,
+                  additional_sense_code, additional_sense_code_qualifier);
+            return SCSI_RES_OK;
+        case UNIT_ATTENTION:
+            ALOGD("UNIT ATTENTION with sense data: key=%hhu, asc=%hhu, ascq=%hhu\n", sense_key,
+                  additional_sense_code, additional_sense_code_qualifier);
+            if (additional_sense_code == 0x29) {
+                /* POWER ON or RESET condition */
+                return SCSI_RES_RETRY;
+            }
+
+            /* treat this UNIT ATTENTION as an error if we don't recognize it */
+            break;
+    }
+
+    ALOGE("Unexpected SCSI sense data: key=%hhu, asc=%hhu, ascq=%hhu\n", sense_key,
+          additional_sense_code, additional_sense_code_qualifier);
+    log_buf(ANDROID_LOG_ERROR, "sense buffer: ", sense_buf, len);
+    return SCSI_RES_ERR;
+}
+
+static enum scsi_result check_sg_io_hdr(const sg_io_hdr_t* io_hdrp) {
+    if (io_hdrp->status == 0 && io_hdrp->host_status == 0 && io_hdrp->driver_status == 0) {
+        return SCSI_RES_OK;
+    }
+
+    if (io_hdrp->status & 0x01) {
+        ALOGE("SG_IO received unknown status, LSB is set: %hhu", io_hdrp->status);
+    }
+
+    if (io_hdrp->masked_status != GOOD && io_hdrp->sb_len_wr > 0) {
+        enum scsi_result scsi_res = check_scsi_sense(io_hdrp->sbp, io_hdrp->sb_len_wr);
+        if (scsi_res == SCSI_RES_RETRY) {
+            return SCSI_RES_RETRY;
+        } else if (scsi_res != SCSI_RES_OK) {
+            ALOGE("Unexpected SCSI sense. masked_status: %hhu, host_status: %hu, driver_status: "
+                  "%hu\n",
+                  io_hdrp->masked_status, io_hdrp->host_status, io_hdrp->driver_status);
+            return scsi_res;
+        }
+    }
+
+    switch (io_hdrp->masked_status) {
+        case GOOD:
+            break;
+        case CHECK_CONDITION:
+            /* handled by check_sg_sense above */
+            break;
+        default:
+            ALOGE("SG_IO failed with masked_status: %hhu, host_status: %hu, driver_status: %hu\n",
+                  io_hdrp->masked_status, io_hdrp->host_status, io_hdrp->driver_status);
+            return SCSI_RES_ERR;
+    }
+
+    if (io_hdrp->host_status != 0) {
+        ALOGE("SG_IO failed with host_status: %hu, driver_status: %hu\n", io_hdrp->host_status,
+              io_hdrp->driver_status);
+    }
+
+    if (io_hdrp->resid != 0) {
+        ALOGE("SG_IO resid was non-zero: %d\n", io_hdrp->resid);
+    }
+    return SCSI_RES_ERR;
+}
+
 static int send_mmc_rpmb_req(int mmc_fd, const struct storage_rpmb_send_req* req) {
     struct {
         struct mmc_ioc_multi_cmd multi;
@@ -153,7 +339,7 @@
         mmc_ioc_cmd_set_data((*cmd), write_buf);
 #ifdef RPMB_DEBUG
         ALOGI("opcode: 0x%x, write_flag: 0x%x\n", cmd->opcode, cmd->write_flag);
-        print_buf("request: ", write_buf, req->reliable_write_size);
+        log_buf(ANDROID_LOG_INFO, "request: ", write_buf, req->reliable_write_size);
 #endif
         write_buf += req->reliable_write_size;
         mmc.multi.num_of_cmds++;
@@ -169,7 +355,7 @@
         mmc_ioc_cmd_set_data((*cmd), write_buf);
 #ifdef RPMB_DEBUG
         ALOGI("opcode: 0x%x, write_flag: 0x%x\n", cmd->opcode, cmd->write_flag);
-        print_buf("request: ", write_buf, req->write_size);
+        log_buf(ANDROID_LOG_INFO, "request: ", write_buf, req->write_size);
 #endif
         write_buf += req->write_size;
         mmc.multi.num_of_cmds++;
@@ -207,6 +393,8 @@
     struct sec_proto_cdb out_cdb = {0xB5, 0xEC, 0x00, 0x01, 0x00, 0x00, 0, 0x00, 0x00};
     unsigned char sense_buffer[32];
 
+    bool is_request_write = req->reliable_write_size > 0;
+
     wl_rc = acquire_wake_lock(PARTIAL_WAKE_LOCK, UFS_WAKE_LOCK_NAME);
     if (wl_rc < 0) {
         ALOGE("%s: failed to acquire wakelock: %d, %s\n", __func__, wl_rc, strerror(errno));
@@ -215,30 +403,44 @@
 
     if (req->reliable_write_size) {
         /* Prepare SECURITY PROTOCOL OUT command. */
-        out_cdb.length = __builtin_bswap32(req->reliable_write_size);
         sg_io_hdr_t io_hdr;
-        set_sg_io_hdr(&io_hdr, SG_DXFER_TO_DEV, sizeof(out_cdb), sizeof(sense_buffer),
-                      req->reliable_write_size, (void*)write_buf, (unsigned char*)&out_cdb,
-                      sense_buffer);
-        rc = ioctl(sg_fd, SG_IO, &io_hdr);
-        if (rc < 0) {
-            ALOGE("%s: ufs ioctl failed: %d, %s\n", __func__, rc, strerror(errno));
-            goto err_op;
-        }
+        int retry_count = UFS_RPMB_WRITE_RETRY_COUNT;
+        do {
+            out_cdb.length = __builtin_bswap32(req->reliable_write_size);
+            set_sg_io_hdr(&io_hdr, SG_DXFER_TO_DEV, sizeof(out_cdb), sizeof(sense_buffer),
+                          req->reliable_write_size, (void*)write_buf, (unsigned char*)&out_cdb,
+                          sense_buffer);
+            rc = ioctl(sg_fd, SG_IO, &io_hdr);
+            if (rc < 0) {
+                ALOGE("%s: ufs ioctl failed: %d, %s\n", __func__, rc, strerror(errno));
+                goto err_op;
+            }
+        } while (check_sg_io_hdr(&io_hdr) == SCSI_RES_RETRY && retry_count-- > 0);
         write_buf += req->reliable_write_size;
     }
 
     if (req->write_size) {
         /* Prepare SECURITY PROTOCOL OUT command. */
-        out_cdb.length = __builtin_bswap32(req->write_size);
         sg_io_hdr_t io_hdr;
-        set_sg_io_hdr(&io_hdr, SG_DXFER_TO_DEV, sizeof(out_cdb), sizeof(sense_buffer),
-                      req->write_size, (void*)write_buf, (unsigned char*)&out_cdb, sense_buffer);
-        rc = ioctl(sg_fd, SG_IO, &io_hdr);
-        if (rc < 0) {
-            ALOGE("%s: ufs ioctl failed: %d, %s\n", __func__, rc, strerror(errno));
-            goto err_op;
-        }
+        /*
+         * We don't retry write response request messages (is_request_write ==
+         * true) because a unit attention condition between the write and
+         * requesting a response means that the device was reset and we can't
+         * get a response to our original write. We can only retry this SG_IO
+         * call when it is the first call in our sequence.
+         */
+        int retry_count = is_request_write ? 0 : UFS_RPMB_READ_RETRY_COUNT;
+        do {
+            out_cdb.length = __builtin_bswap32(req->write_size);
+            set_sg_io_hdr(&io_hdr, SG_DXFER_TO_DEV, sizeof(out_cdb), sizeof(sense_buffer),
+                          req->write_size, (void*)write_buf, (unsigned char*)&out_cdb,
+                          sense_buffer);
+            rc = ioctl(sg_fd, SG_IO, &io_hdr);
+            if (rc < 0) {
+                ALOGE("%s: ufs ioctl failed: %d, %s\n", __func__, rc, strerror(errno));
+                goto err_op;
+            }
+        } while (check_sg_io_hdr(&io_hdr) == SCSI_RES_RETRY && retry_count-- > 0);
         write_buf += req->write_size;
     }
 
@@ -252,6 +454,7 @@
         if (rc < 0) {
             ALOGE("%s: ufs ioctl failed: %d, %s\n", __func__, rc, strerror(errno));
         }
+        check_sg_io_hdr(&io_hdr);
     }
 
 err_op:
@@ -353,7 +556,7 @@
         goto err_response;
     }
 #ifdef RPMB_DEBUG
-    if (req->read_size) print_buf("response: ", read_buf, req->read_size);
+    if (req->read_size) log_buf(ANDROID_LOG_INFO, "response: ", read_buf, req->read_size);
 #endif
 
     if (msg->flags & STORAGE_MSG_FLAG_POST_COMMIT) {