diff --git a/debuggerd/crash_dump.cpp b/debuggerd/crash_dump.cpp
index 5bb1d75..967b942 100644
--- a/debuggerd/crash_dump.cpp
+++ b/debuggerd/crash_dump.cpp
@@ -502,15 +502,24 @@
         continue;
       }
 
-      struct iovec iov = {
+      struct iovec tagged_addr_iov = {
           &info.tagged_addr_ctrl,
           sizeof(info.tagged_addr_ctrl),
       };
       if (ptrace(PTRACE_GETREGSET, thread, NT_ARM_TAGGED_ADDR_CTRL,
-                 reinterpret_cast<void*>(&iov)) == -1) {
+                 reinterpret_cast<void*>(&tagged_addr_iov)) == -1) {
         info.tagged_addr_ctrl = -1;
       }
 
+      struct iovec pac_enabled_keys_iov = {
+          &info.pac_enabled_keys,
+          sizeof(info.pac_enabled_keys),
+      };
+      if (ptrace(PTRACE_GETREGSET, thread, NT_ARM_PAC_ENABLED_KEYS,
+                 reinterpret_cast<void*>(&pac_enabled_keys_iov)) == -1) {
+        info.pac_enabled_keys = -1;
+      }
+
       if (thread == g_target_thread) {
         // Read the thread's registers along with the rest of the crash info out of the pipe.
         ReadCrashInfo(input_pipe, &siginfo, &info.registers, &process_info);
diff --git a/debuggerd/debuggerd_test.cpp b/debuggerd/debuggerd_test.cpp
index b107767..2cf5b18 100644
--- a/debuggerd/debuggerd_test.cpp
+++ b/debuggerd/debuggerd_test.cpp
@@ -173,6 +173,14 @@
   *status = response.status;
 }
 
+static bool pac_supported() {
+#if defined(__aarch64__)
+  return getauxval(AT_HWCAP) & HWCAP_PACA;
+#else
+  return false;
+#endif
+}
+
 class CrasherTest : public ::testing::Test {
  public:
   pid_t crasher_pid = -1;
@@ -357,6 +365,12 @@
     ASSERT_MATCH(result, R"(tagged_addr_ctrl: 000000000007fff3)"
                          R"( \(PR_TAGGED_ADDR_ENABLE, PR_MTE_TCF_SYNC, mask 0xfffe\))");
   }
+
+  if (pac_supported()) {
+    // Test that the default PAC_ENABLED_KEYS value is set.
+    ASSERT_MATCH(result, R"(pac_enabled_keys: 000000000000000f)"
+                         R"( \(PR_PAC_APIAKEY, PR_PAC_APIBKEY, PR_PAC_APDAKEY, PR_PAC_APDBKEY\))");
+  }
 }
 
 TEST_F(CrasherTest, tagged_fault_addr) {
diff --git a/debuggerd/libdebuggerd/gwp_asan.cpp b/debuggerd/libdebuggerd/gwp_asan.cpp
index 3ee309f..ed2b974 100644
--- a/debuggerd/libdebuggerd/gwp_asan.cpp
+++ b/debuggerd/libdebuggerd/gwp_asan.cpp
@@ -147,7 +147,7 @@
   for (size_t i = 0; i != num_frames; ++i) {
     unwindstack::FrameData frame_data = unwinder->BuildFrameFromPcOnly(frames[i]);
     BacktraceFrame* f = heap_object->add_allocation_backtrace();
-    fill_in_backtrace_frame(f, frame_data, unwinder->GetMaps());
+    fill_in_backtrace_frame(f, frame_data);
   }
 
   heap_object->set_deallocation_tid(__gwp_asan_get_deallocation_thread_id(responsible_allocation_));
@@ -156,7 +156,7 @@
   for (size_t i = 0; i != num_frames; ++i) {
     unwindstack::FrameData frame_data = unwinder->BuildFrameFromPcOnly(frames[i]);
     BacktraceFrame* f = heap_object->add_deallocation_backtrace();
-    fill_in_backtrace_frame(f, frame_data, unwinder->GetMaps());
+    fill_in_backtrace_frame(f, frame_data);
   }
 
   set_human_readable_cause(cause, crash_address_);
diff --git a/debuggerd/libdebuggerd/include/libdebuggerd/tombstone.h b/debuggerd/libdebuggerd/include/libdebuggerd/tombstone.h
index 2331f1e..7bf1688 100644
--- a/debuggerd/libdebuggerd/include/libdebuggerd/tombstone.h
+++ b/debuggerd/libdebuggerd/include/libdebuggerd/tombstone.h
@@ -37,7 +37,6 @@
 
 namespace unwindstack {
 struct FrameData;
-class Maps;
 class Unwinder;
 }
 
@@ -68,8 +67,7 @@
     const Tombstone& tombstone,
     std::function<void(const std::string& line, bool should_log)> callback);
 
-void fill_in_backtrace_frame(BacktraceFrame* f, const unwindstack::FrameData& frame,
-                             unwindstack::Maps* maps);
+void fill_in_backtrace_frame(BacktraceFrame* f, const unwindstack::FrameData& frame);
 void set_human_readable_cause(Cause* cause, uint64_t fault_addr);
 
 #endif  // _DEBUGGERD_TOMBSTONE_H
diff --git a/debuggerd/libdebuggerd/include/libdebuggerd/types.h b/debuggerd/libdebuggerd/include/libdebuggerd/types.h
index 086dc97..a51e276 100644
--- a/debuggerd/libdebuggerd/include/libdebuggerd/types.h
+++ b/debuggerd/libdebuggerd/include/libdebuggerd/types.h
@@ -25,6 +25,7 @@
 struct ThreadInfo {
   std::unique_ptr<unwindstack::Regs> registers;
   long tagged_addr_ctrl = -1;
+  long pac_enabled_keys = -1;
 
   pid_t uid;
 
diff --git a/debuggerd/libdebuggerd/include/libdebuggerd/utility.h b/debuggerd/libdebuggerd/include/libdebuggerd/utility.h
index 002321f..63e142f 100644
--- a/debuggerd/libdebuggerd/include/libdebuggerd/utility.h
+++ b/debuggerd/libdebuggerd/include/libdebuggerd/utility.h
@@ -1,22 +1,20 @@
-/* system/debuggerd/utility.h
-**
-** Copyright 2008, 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.
-*/
+/*
+ * Copyright 2008, 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.
+ */
 
-#ifndef _DEBUGGERD_UTILITY_H
-#define _DEBUGGERD_UTILITY_H
+#pragma once
 
 #include <inttypes.h>
 #include <signal.h>
@@ -93,6 +91,7 @@
 const char* get_signame(const siginfo_t*);
 const char* get_sigcode(const siginfo_t*);
 std::string describe_tagged_addr_ctrl(long ctrl);
+std::string describe_pac_enabled_keys(long keys);
 
 // Number of bytes per MTE granule.
 constexpr size_t kTagGranuleSize = 16;
@@ -100,5 +99,3 @@
 // Number of rows and columns to display in an MTE tag dump.
 constexpr size_t kNumTagColumns = 16;
 constexpr size_t kNumTagRows = 16;
-
-#endif // _DEBUGGERD_UTILITY_H
diff --git a/debuggerd/libdebuggerd/scudo.cpp b/debuggerd/libdebuggerd/scudo.cpp
index a89f385..a2933f2 100644
--- a/debuggerd/libdebuggerd/scudo.cpp
+++ b/debuggerd/libdebuggerd/scudo.cpp
@@ -108,7 +108,7 @@
   for (size_t i = 0; i < arraysize(report->allocation_trace) && report->allocation_trace[i]; ++i) {
     unwindstack::FrameData frame_data = unwinder->BuildFrameFromPcOnly(report->allocation_trace[i]);
     BacktraceFrame* f = heap_object->add_allocation_backtrace();
-    fill_in_backtrace_frame(f, frame_data, unwinder->GetMaps());
+    fill_in_backtrace_frame(f, frame_data);
   }
 
   heap_object->set_deallocation_tid(report->deallocation_tid);
@@ -117,7 +117,7 @@
     unwindstack::FrameData frame_data =
         unwinder->BuildFrameFromPcOnly(report->deallocation_trace[i]);
     BacktraceFrame* f = heap_object->add_deallocation_backtrace();
-    fill_in_backtrace_frame(f, frame_data, unwinder->GetMaps());
+    fill_in_backtrace_frame(f, frame_data);
   }
 
   set_human_readable_cause(cause, untagged_fault_addr_);
diff --git a/debuggerd/libdebuggerd/test/utility_test.cpp b/debuggerd/libdebuggerd/test/utility_test.cpp
index 97328b7..dad3380 100644
--- a/debuggerd/libdebuggerd/test/utility_test.cpp
+++ b/debuggerd/libdebuggerd/test/utility_test.cpp
@@ -31,3 +31,12 @@
       describe_tagged_addr_ctrl(0xf0000000 | PR_TAGGED_ADDR_ENABLE | PR_MTE_TCF_SYNC |
                                 PR_MTE_TCF_ASYNC | (0xfffe << PR_MTE_TAG_SHIFT)));
 }
+
+TEST(UtilityTest, describe_pac_enabled_keys) {
+  EXPECT_EQ("", describe_pac_enabled_keys(0));
+  EXPECT_EQ(" (PR_PAC_APIAKEY)", describe_pac_enabled_keys(PR_PAC_APIAKEY));
+  EXPECT_EQ(" (PR_PAC_APIAKEY, PR_PAC_APDBKEY)",
+            describe_pac_enabled_keys(PR_PAC_APIAKEY | PR_PAC_APDBKEY));
+  EXPECT_EQ(" (PR_PAC_APIAKEY, PR_PAC_APDBKEY, unknown 0x1000)",
+            describe_pac_enabled_keys(PR_PAC_APIAKEY | PR_PAC_APDBKEY | 0x1000));
+}
diff --git a/debuggerd/libdebuggerd/tombstone_proto.cpp b/debuggerd/libdebuggerd/tombstone_proto.cpp
index b1c4ef3..b7d5bc4 100644
--- a/debuggerd/libdebuggerd/tombstone_proto.cpp
+++ b/debuggerd/libdebuggerd/tombstone_proto.cpp
@@ -312,8 +312,7 @@
   }
 }
 
-void fill_in_backtrace_frame(BacktraceFrame* f, const unwindstack::FrameData& frame,
-                             unwindstack::Maps* maps) {
+void fill_in_backtrace_frame(BacktraceFrame* f, const unwindstack::FrameData& frame) {
   f->set_rel_pc(frame.rel_pc);
   f->set_pc(frame.pc);
   f->set_sp(frame.sp);
@@ -331,21 +330,20 @@
 
   f->set_function_offset(frame.function_offset);
 
-  if (frame.map_start == frame.map_end) {
+  if (frame.map_info == nullptr) {
     // No valid map associated with this frame.
     f->set_file_name("<unknown>");
-  } else if (!frame.map_name.empty()) {
-    f->set_file_name(frame.map_name);
+    return;
+  }
+
+  if (!frame.map_info->name().empty()) {
+    f->set_file_name(frame.map_info->GetFullName());
   } else {
-    f->set_file_name(StringPrintf("<anonymous:%" PRIx64 ">", frame.map_start));
+    f->set_file_name(StringPrintf("<anonymous:%" PRIx64 ">", frame.map_info->start()));
   }
+  f->set_file_map_offset(frame.map_info->elf_start_offset());
 
-  f->set_file_map_offset(frame.map_elf_start_offset);
-
-  auto map_info = maps->Find(frame.map_start);
-  if (map_info.get() != nullptr) {
-    f->set_build_id(map_info->GetPrintableBuildID());
-  }
+  f->set_build_id(frame.map_info->GetPrintableBuildID());
 }
 
 static void dump_thread(Tombstone* tombstone, unwindstack::Unwinder* unwinder,
@@ -355,6 +353,7 @@
   thread.set_id(thread_info.tid);
   thread.set_name(thread_info.thread_name);
   thread.set_tagged_addr_ctrl(thread_info.tagged_addr_ctrl);
+  thread.set_pac_enabled_keys(thread_info.pac_enabled_keys);
 
   unwindstack::Maps* maps = unwinder->GetMaps();
   unwindstack::Memory* memory = unwinder->GetProcessMemory().get();
@@ -434,7 +433,7 @@
     unwinder->SetDisplayBuildID(true);
     for (const auto& frame : unwinder->frames()) {
       BacktraceFrame* f = thread.add_current_backtrace();
-      fill_in_backtrace_frame(f, frame, maps);
+      fill_in_backtrace_frame(f, frame);
     }
   }
 
diff --git a/debuggerd/libdebuggerd/tombstone_proto_to_text.cpp b/debuggerd/libdebuggerd/tombstone_proto_to_text.cpp
index de86b0a..0265641 100644
--- a/debuggerd/libdebuggerd/tombstone_proto_to_text.cpp
+++ b/debuggerd/libdebuggerd/tombstone_proto_to_text.cpp
@@ -85,6 +85,10 @@
     CB(should_log, "tagged_addr_ctrl: %016" PRIx64 "%s", thread.tagged_addr_ctrl(),
        describe_tagged_addr_ctrl(thread.tagged_addr_ctrl()).c_str());
   }
+  if (thread.pac_enabled_keys() != -1) {
+    CB(should_log, "pac_enabled_keys: %016" PRIx64 "%s", thread.pac_enabled_keys(),
+       describe_pac_enabled_keys(thread.pac_enabled_keys()).c_str());
+  }
 }
 
 static void print_register_row(CallbackType callback, int word_size,
diff --git a/debuggerd/libdebuggerd/utility.cpp b/debuggerd/libdebuggerd/utility.cpp
index 71f0c09..543a67c 100644
--- a/debuggerd/libdebuggerd/utility.cpp
+++ b/debuggerd/libdebuggerd/utility.cpp
@@ -446,31 +446,40 @@
   return "?";
 }
 
-std::string describe_tagged_addr_ctrl(long ctrl) {
+#define DESCRIBE_FLAG(flag) \
+  if (value & flag) {       \
+    desc += ", ";           \
+    desc += #flag;          \
+    value &= ~flag;         \
+  }
+
+static std::string describe_end(long value, std::string& desc) {
+  if (value) {
+    desc += StringPrintf(", unknown 0x%lx", value);
+  }
+  return desc.empty() ? "" : " (" + desc.substr(2) + ")";
+}
+
+std::string describe_tagged_addr_ctrl(long value) {
   std::string desc;
-  if (ctrl & PR_TAGGED_ADDR_ENABLE) {
-    desc += ", PR_TAGGED_ADDR_ENABLE";
-    ctrl &= ~PR_TAGGED_ADDR_ENABLE;
+  DESCRIBE_FLAG(PR_TAGGED_ADDR_ENABLE);
+  DESCRIBE_FLAG(PR_MTE_TCF_SYNC);
+  DESCRIBE_FLAG(PR_MTE_TCF_ASYNC);
+  if (value & PR_MTE_TAG_MASK) {
+    desc += StringPrintf(", mask 0x%04lx", (value & PR_MTE_TAG_MASK) >> PR_MTE_TAG_SHIFT);
+    value &= ~PR_MTE_TAG_MASK;
   }
-  if (ctrl & PR_MTE_TCF_SYNC) {
-    desc += ", PR_MTE_TCF_SYNC";
-    ctrl &= ~PR_MTE_TCF_SYNC;
-  }
-  if (ctrl & PR_MTE_TCF_ASYNC) {
-    desc += ", PR_MTE_TCF_ASYNC";
-    ctrl &= ~PR_MTE_TCF_ASYNC;
-  }
-  if (ctrl & PR_MTE_TAG_MASK) {
-    desc += StringPrintf(", mask 0x%04lx", (ctrl & PR_MTE_TAG_MASK) >> PR_MTE_TAG_SHIFT);
-    ctrl &= ~PR_MTE_TAG_MASK;
-  }
-  if (ctrl) {
-    desc += StringPrintf(", unknown 0x%lx", ctrl);
-  }
-  if (desc.empty()) {
-    return "";
-  }
-  return " (" + desc.substr(2) + ")";
+  return describe_end(value, desc);
+}
+
+std::string describe_pac_enabled_keys(long value) {
+  std::string desc;
+  DESCRIBE_FLAG(PR_PAC_APIAKEY);
+  DESCRIBE_FLAG(PR_PAC_APIBKEY);
+  DESCRIBE_FLAG(PR_PAC_APDAKEY);
+  DESCRIBE_FLAG(PR_PAC_APDBKEY);
+  DESCRIBE_FLAG(PR_PAC_APGAKEY);
+  return describe_end(value, desc);
 }
 
 void log_backtrace(log_t* log, unwindstack::Unwinder* unwinder, const char* prefix) {
diff --git a/debuggerd/proto/tombstone.proto b/debuggerd/proto/tombstone.proto
index a701212..40a942e 100644
--- a/debuggerd/proto/tombstone.proto
+++ b/debuggerd/proto/tombstone.proto
@@ -126,8 +126,9 @@
   repeated BacktraceFrame current_backtrace = 4;
   repeated MemoryDump memory_dump = 5;
   int64 tagged_addr_ctrl = 6;
+  int64 pac_enabled_keys = 8;
 
-  reserved 8 to 999;
+  reserved 9 to 999;
 }
 
 message BacktraceFrame {
diff --git a/fastboot/Android.bp b/fastboot/Android.bp
index 708a677..9ae2c37 100644
--- a/fastboot/Android.bp
+++ b/fastboot/Android.bp
@@ -15,7 +15,10 @@
 // This is required because no Android.bp can include a library defined in an
 // Android.mk. Eventually should kill libfastboot (defined in Android.mk)
 package {
-    default_applicable_licenses: ["system_core_fastboot_license"],
+    default_applicable_licenses: [
+        "system_core_fastboot_license",
+        "Android-Apache-2.0",
+    ],
 }
 
 // Added automatically by a large-scale-change that took the approach of
@@ -36,10 +39,9 @@
     name: "system_core_fastboot_license",
     visibility: [":__subpackages__"],
     license_kinds: [
-        "SPDX-license-identifier-Apache-2.0",
         "SPDX-license-identifier-BSD",
     ],
-    // large-scale-change unable to identify any license_text files
+    license_text: ["LICENSE"],
 }
 
 cc_library_host_static {
diff --git a/fastboot/LICENSE b/fastboot/LICENSE
new file mode 100644
index 0000000..f0a0e52
--- /dev/null
+++ b/fastboot/LICENSE
@@ -0,0 +1,23 @@
+ Redistribution and use in source and binary forms, with or without
+ modification, are permitted provided that the following conditions
+ are met:
+  * Redistributions of source code must retain the above copyright
+    notice, this list of conditions and the following disclaimer.
+  * Redistributions in binary form must reproduce the above copyright
+    notice, this list of conditions and the following disclaimer in
+    the documentation and/or other materials provided with the
+    distribution.
+
+ THIS SOFTWARE IS PROVIDED BY THE COPYRIGHT HOLDERS AND CONTRIBUTORS
+ "AS IS" AND ANY EXPRESS OR IMPLIED WARRANTIES, INCLUDING, BUT NOT
+ LIMITED TO, THE IMPLIED WARRANTIES OF MERCHANTABILITY AND FITNESS
+ FOR A PARTICULAR PURPOSE ARE DISCLAIMED. IN NO EVENT SHALL THE
+ COPYRIGHT OWNER OR CONTRIBUTORS BE LIABLE FOR ANY DIRECT, INDIRECT,
+ INCIDENTAL, SPECIAL, EXEMPLARY, OR CONSEQUENTIAL DAMAGES (INCLUDING,
+ BUT NOT LIMITED TO, PROCUREMENT OF SUBSTITUTE GOODS OR SERVICES; LOSS
+ OF USE, DATA, OR PROFITS; OR BUSINESS INTERRUPTION) HOWEVER CAUSED
+ AND ON ANY THEORY OF LIABILITY, WHETHER IN CONTRACT, STRICT LIABILITY,
+ OR TORT (INCLUDING NEGLIGENCE OR OTHERWISE) ARISING IN ANY WAY OUT
+ OF THE USE OF THIS SOFTWARE, EVEN IF ADVISED OF THE POSSIBILITY OF
+ SUCH DAMAGE.
+
diff --git a/fastboot/device/flashing.cpp b/fastboot/device/flashing.cpp
index 7bef72a..44dc81f 100644
--- a/fastboot/device/flashing.cpp
+++ b/fastboot/device/flashing.cpp
@@ -119,9 +119,11 @@
 }
 
 int FlashSparseData(int fd, std::vector<char>& downloaded_data) {
-    struct sparse_file* file = sparse_file_import_buf(downloaded_data.data(), true, false);
+    struct sparse_file* file = sparse_file_import_buf(downloaded_data.data(),
+                                                      downloaded_data.size(), true, false);
     if (!file) {
-        return -ENOENT;
+        // Invalid sparse format
+        return -EINVAL;
     }
     return sparse_file_callback(file, false, false, WriteCallback, reinterpret_cast<void*>(fd));
 }
diff --git a/fastboot/fuzzy_fastboot/main.cpp b/fastboot/fuzzy_fastboot/main.cpp
index 8593adc..074306b 100644
--- a/fastboot/fuzzy_fastboot/main.cpp
+++ b/fastboot/fuzzy_fastboot/main.cpp
@@ -874,6 +874,12 @@
             << "Device did not respond with FAIL for malformed download command '" << cmd << "'";
 }
 
+TEST_F(Fuzz, DownloadInvalid9) {
+    std::string cmd("download:2PPPPPPPPPPPPPPPPPPPPPPPPPPPPPP");
+    EXPECT_EQ(fb->RawCommand(cmd), DEVICE_FAIL)
+            << "Device did not respond with FAIL for malformed download command '" << cmd << "'";
+}
+
 TEST_F(Fuzz, GetVarAllSpam) {
     auto start = std::chrono::high_resolution_clock::now();
     std::chrono::duration<double> elapsed;
@@ -977,6 +983,80 @@
     }
 }
 
+TEST_F(Fuzz, SparseZeroBlkSize) {
+    // handcrafted malform sparse file with zero as block size
+    const std::vector<char> buf = {
+        '\x3a', '\xff', '\x26', '\xed', '\x01', '\x00', '\x00', '\x00', '\x1c', '\x00', '\x0c', '\x00',
+        '\x00', '\x00', '\x00', '\x00', '\x01', '\x00', '\x00', '\x00', '\x01', '\x00', '\x00', '\x00',
+        '\x00', '\x00', '\x00', '\x00', '\xc2', '\xca', '\x00', '\x00', '\x01', '\x00', '\x00', '\x00',
+        '\x10', '\x00', '\x00', '\x00', '\x11', '\x22', '\x33', '\x44'
+    };
+
+    ASSERT_EQ(DownloadCommand(buf.size()), SUCCESS) << "Device rejected download command";
+    ASSERT_EQ(SendBuffer(buf), SUCCESS) << "Downloading payload failed";
+
+    // It can either reject this download or reject it during flash
+    if (HandleResponse() != DEVICE_FAIL) {
+        EXPECT_EQ(fb->Flash("userdata"), DEVICE_FAIL)
+                << "Flashing a zero block size in sparse file should fail";
+    }
+}
+
+TEST_F(Fuzz, SparseVeryLargeBlkSize) {
+    // handcrafted sparse file with block size of ~4GB and divisible 4
+    const std::vector<char> buf = {
+        '\x3a', '\xff', '\x26', '\xed', '\x01', '\x00', '\x00', '\x00',
+        '\x1c', '\x00', '\x0c', '\x00', '\xF0', '\xFF', '\xFF', '\xFF',
+        '\x01', '\x00', '\x00', '\x00', '\x01', '\x00', '\x00', '\x00',
+        '\x00', '\x00', '\x00', '\x00', '\xc3', '\xca', '\x00', '\x00',
+        '\x01', '\x00', '\x00', '\x00', '\x0c', '\x00', '\x00', '\x00',
+        '\x11', '\x22', '\x33', '\x44'
+    };
+
+    ASSERT_EQ(DownloadCommand(buf.size()), SUCCESS) << "Device rejected download command";
+    ASSERT_EQ(SendBuffer(buf), SUCCESS) << "Downloading payload failed";
+    ASSERT_EQ(HandleResponse(), SUCCESS) << "Not receive okay";
+    ASSERT_EQ(fb->Flash("userdata"), SUCCESS) << "Flashing sparse failed";
+}
+
+TEST_F(Fuzz, SparseTrimmed) {
+    // handcrafted malform sparse file which is trimmed
+    const std::vector<char> buf = {
+        '\x3a', '\xff', '\x26', '\xed', '\x01', '\x00', '\x00', '\x00', '\x1c', '\x00', '\x0c', '\x00',
+        '\x00', '\x10', '\x00', '\x00', '\x00', '\x00', '\x08', '\x00', '\x01', '\x00', '\x00', '\x00',
+        '\x00', '\x00', '\x00', '\x00', '\xc1', '\xca', '\x00', '\x00', '\x01', '\x00', '\x00', '\x00',
+        '\x00', '\x00', '\x00', '\x80', '\x11', '\x22', '\x33', '\x44'
+    };
+
+    ASSERT_EQ(DownloadCommand(buf.size()), SUCCESS) << "Device rejected download command";
+    ASSERT_EQ(SendBuffer(buf), SUCCESS) << "Downloading payload failed";
+
+    // It can either reject this download or reject it during flash
+    if (HandleResponse() != DEVICE_FAIL) {
+        EXPECT_EQ(fb->Flash("userdata"), DEVICE_FAIL)
+                << "Flashing a trimmed sparse file should fail";
+    }
+}
+
+TEST_F(Fuzz, SparseInvalidChurk) {
+    // handcrafted malform sparse file with invalid churk
+    const std::vector<char> buf = {
+        '\x3a', '\xff', '\x26', '\xed', '\x01', '\x00', '\x00', '\x00', '\x1c', '\x00', '\x0c', '\x00',
+        '\x00', '\x10', '\x00', '\x00', '\x00', '\x00', '\x08', '\x00', '\x01', '\x00', '\x00', '\x00',
+        '\x00', '\x00', '\x00', '\x00', '\xc1', '\xca', '\x00', '\x00', '\x01', '\x00', '\x00', '\x00',
+        '\x10', '\x00', '\x00', '\x00', '\x11', '\x22', '\x33', '\x44'
+    };
+
+    ASSERT_EQ(DownloadCommand(buf.size()), SUCCESS) << "Device rejected download command";
+    ASSERT_EQ(SendBuffer(buf), SUCCESS) << "Downloading payload failed";
+
+    // It can either reject this download or reject it during flash
+    if (HandleResponse() != DEVICE_FAIL) {
+        EXPECT_EQ(fb->Flash("userdata"), DEVICE_FAIL)
+                << "Flashing a sparse file with invalid churk should fail";
+    }
+}
+
 TEST_F(Fuzz, SparseTooManyChunks) {
     SparseWrapper sparse(4096, 4096);  // 1 block, but we send two chunks that will use 2 blocks
     ASSERT_TRUE(*sparse) << "Sparse image creation failed";
diff --git a/fs_mgr/libsnapshot/snapshot.cpp b/fs_mgr/libsnapshot/snapshot.cpp
index 18a9d22..f3de2b4 100644
--- a/fs_mgr/libsnapshot/snapshot.cpp
+++ b/fs_mgr/libsnapshot/snapshot.cpp
@@ -1467,6 +1467,14 @@
     }
 
     RemoveAllUpdateState(lock);
+
+    if (UpdateUsesUserSnapshots(lock) && !device()->IsTestDevice()) {
+        if (snapuserd_client_) {
+            snapuserd_client_->DetachSnapuserd();
+            snapuserd_client_->CloseConnection();
+            snapuserd_client_ = nullptr;
+        }
+    }
 }
 
 void SnapshotManager::AcknowledgeMergeFailure(MergeFailureCode failure_code) {
@@ -3200,7 +3208,7 @@
 
             // Terminate stale daemon if any
             std::unique_ptr<SnapuserdClient> snapuserd_client =
-                    SnapuserdClient::Connect(kSnapuserdSocket, 10s);
+                    SnapuserdClient::Connect(kSnapuserdSocket, 5s);
             if (snapuserd_client) {
                 snapuserd_client->DetachSnapuserd();
                 snapuserd_client->CloseConnection();
diff --git a/fs_mgr/libsnapshot/snapshot_test.cpp b/fs_mgr/libsnapshot/snapshot_test.cpp
index 11cebe1..d76558b 100644
--- a/fs_mgr/libsnapshot/snapshot_test.cpp
+++ b/fs_mgr/libsnapshot/snapshot_test.cpp
@@ -54,6 +54,8 @@
 #include <libsnapshot/mock_snapshot.h>
 
 DEFINE_string(force_config, "", "Force testing mode (dmsnap, vab, vabc) ignoring device config.");
+DEFINE_string(force_iouring_disable, "",
+              "Force testing mode (iouring_disabled) - disable io_uring");
 
 namespace android {
 namespace snapshot {
@@ -2769,10 +2771,22 @@
         }
     }
 
+    if (FLAGS_force_iouring_disable == "iouring_disabled") {
+        if (!android::base::SetProperty("snapuserd.test.io_uring.force_disable", "1")) {
+            return testing::AssertionFailure()
+                   << "Failed to disable property: snapuserd.test.io_uring.disabled";
+        }
+    }
+
     int ret = RUN_ALL_TESTS();
 
     if (FLAGS_force_config == "dmsnap") {
         android::base::SetProperty("snapuserd.test.dm.snapshots", "0");
     }
+
+    if (FLAGS_force_iouring_disable == "iouring_disabled") {
+        android::base::SetProperty("snapuserd.test.io_uring.force_disable", "0");
+    }
+
     return ret;
 }
diff --git a/fs_mgr/libsnapshot/snapuserd/Android.bp b/fs_mgr/libsnapshot/snapuserd/Android.bp
index 84bcb94..bc2bceb 100644
--- a/fs_mgr/libsnapshot/snapuserd/Android.bp
+++ b/fs_mgr/libsnapshot/snapuserd/Android.bp
@@ -86,7 +86,9 @@
         "libsnapshot_cow",
         "libz",
         "libext4_utils",
+        "liburing",
     ],
+    include_dirs: ["bionic/libc/kernel"],
 }
 
 cc_binary {
@@ -182,7 +184,10 @@
         "libfs_mgr",
         "libdm",
         "libext4_utils",
+        "liburing",
+        "libgflags",
     ],
+    include_dirs: ["bionic/libc/kernel"],
     header_libs: [
         "libstorage_literals_headers",
         "libfiemap_headers",
diff --git a/fs_mgr/libsnapshot/snapuserd/snapuserd_daemon.cpp b/fs_mgr/libsnapshot/snapuserd/snapuserd_daemon.cpp
index ddb1f79..a082742 100644
--- a/fs_mgr/libsnapshot/snapuserd/snapuserd_daemon.cpp
+++ b/fs_mgr/libsnapshot/snapuserd/snapuserd_daemon.cpp
@@ -209,6 +209,8 @@
 int main(int argc, char** argv) {
     android::base::InitLogging(argv, &android::base::KernelLogger);
 
+    LOG(INFO) << "snapuserd daemon about to start";
+
     android::snapshot::Daemon& daemon = android::snapshot::Daemon::Instance();
 
     if (!daemon.StartDaemon(argc, argv)) {
diff --git a/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_core.cpp b/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_core.cpp
index 95d95cd..5109d82 100644
--- a/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_core.cpp
+++ b/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_core.cpp
@@ -16,6 +16,10 @@
 
 #include "snapuserd_core.h"
 
+#include <sys/utsname.h>
+
+#include <android-base/properties.h>
+#include <android-base/scopeguard.h>
 #include <android-base/strings.h>
 
 namespace android {
@@ -288,6 +292,136 @@
     return ReadMetadata();
 }
 
+void SnapshotHandler::FinalizeIouring() {
+    io_uring_queue_exit(ring_.get());
+}
+
+bool SnapshotHandler::InitializeIouring(int io_depth) {
+    ring_ = std::make_unique<struct io_uring>();
+
+    int ret = io_uring_queue_init(io_depth, ring_.get(), 0);
+    if (ret) {
+        LOG(ERROR) << "io_uring_queue_init failed with ret: " << ret;
+        return false;
+    }
+
+    LOG(INFO) << "io_uring_queue_init success with io_depth: " << io_depth;
+    return true;
+}
+
+bool SnapshotHandler::ReadBlocksAsync(const std::string& dm_block_device,
+                                      const std::string& partition_name, size_t size) {
+    // 64k block size with io_depth of 64 is optimal
+    // for a single thread. We just need a single thread
+    // to read all the blocks from all dynamic partitions.
+    size_t io_depth = 64;
+    size_t bs = (64 * 1024);
+
+    if (!InitializeIouring(io_depth)) {
+        return false;
+    }
+
+    LOG(INFO) << "ReadBlockAsync start "
+              << " Block-device: " << dm_block_device << " Partition-name: " << partition_name
+              << " Size: " << size;
+
+    auto scope_guard = android::base::make_scope_guard([this]() -> void { FinalizeIouring(); });
+
+    std::vector<std::unique_ptr<struct iovec>> vecs;
+    using AlignedBuf = std::unique_ptr<void, decltype(free)*>;
+    std::vector<AlignedBuf> alignedBufVector;
+
+    /*
+     * TODO: We need aligned memory for DIRECT-IO. However, if we do
+     * a DIRECT-IO and verify the blocks then we need to inform
+     * update-verifier that block verification has been done and
+     * there is no need to repeat the same. We are not there yet
+     * as we need to see if there are any boot time improvements doing
+     * a DIRECT-IO.
+     *
+     * Also, we could you the same function post merge for block verification;
+     * again, we can do a DIRECT-IO instead of thrashing page-cache and
+     * hurting other applications.
+     *
+     * For now, we will just create aligned buffers but rely on buffered
+     * I/O until we have perf numbers to justify DIRECT-IO.
+     */
+    for (int i = 0; i < io_depth; i++) {
+        auto iovec = std::make_unique<struct iovec>();
+        vecs.push_back(std::move(iovec));
+
+        struct iovec* iovec_ptr = vecs[i].get();
+
+        if (posix_memalign(&iovec_ptr->iov_base, BLOCK_SZ, bs)) {
+            LOG(ERROR) << "posix_memalign failed";
+            return false;
+        }
+
+        iovec_ptr->iov_len = bs;
+        alignedBufVector.push_back(
+                std::unique_ptr<void, decltype(free)*>(iovec_ptr->iov_base, free));
+    }
+
+    android::base::unique_fd fd(TEMP_FAILURE_RETRY(open(dm_block_device.c_str(), O_RDONLY)));
+    if (fd.get() == -1) {
+        SNAP_PLOG(ERROR) << "File open failed - block-device " << dm_block_device
+                         << " partition-name: " << partition_name;
+        return false;
+    }
+
+    loff_t offset = 0;
+    size_t remain = size;
+    size_t read_sz = io_depth * bs;
+
+    while (remain > 0) {
+        size_t to_read = std::min(remain, read_sz);
+        size_t queue_size = to_read / bs;
+
+        for (int i = 0; i < queue_size; i++) {
+            struct io_uring_sqe* sqe = io_uring_get_sqe(ring_.get());
+            if (!sqe) {
+                SNAP_LOG(ERROR) << "io_uring_get_sqe() failed";
+                return false;
+            }
+
+            struct iovec* iovec_ptr = vecs[i].get();
+
+            io_uring_prep_read(sqe, fd.get(), iovec_ptr->iov_base, iovec_ptr->iov_len, offset);
+            sqe->flags |= IOSQE_ASYNC;
+            offset += bs;
+        }
+
+        int ret = io_uring_submit(ring_.get());
+        if (ret != queue_size) {
+            SNAP_LOG(ERROR) << "submit got: " << ret << " wanted: " << queue_size;
+            return false;
+        }
+
+        for (int i = 0; i < queue_size; i++) {
+            struct io_uring_cqe* cqe;
+
+            int ret = io_uring_wait_cqe(ring_.get(), &cqe);
+            if (ret) {
+                SNAP_PLOG(ERROR) << "wait_cqe failed" << ret;
+                return false;
+            }
+
+            if (cqe->res < 0) {
+                SNAP_LOG(ERROR) << "io failed with res: " << cqe->res;
+                return false;
+            }
+            io_uring_cqe_seen(ring_.get(), cqe);
+        }
+
+        remain -= to_read;
+    }
+
+    LOG(INFO) << "ReadBlockAsync complete: "
+              << " Block-device: " << dm_block_device << " Partition-name: " << partition_name
+              << " Size: " << size;
+    return true;
+}
+
 void SnapshotHandler::ReadBlocksToCache(const std::string& dm_block_device,
                                         const std::string& partition_name, off_t offset,
                                         size_t size) {
@@ -344,17 +478,22 @@
         return;
     }
 
-    int num_threads = 2;
-    size_t num_blocks = dev_sz >> BLOCK_SHIFT;
-    size_t num_blocks_per_thread = num_blocks / num_threads;
-    size_t read_sz_per_thread = num_blocks_per_thread << BLOCK_SHIFT;
-    off_t offset = 0;
+    if (IsIouringSupported()) {
+        std::async(std::launch::async, &SnapshotHandler::ReadBlocksAsync, this, dm_block_device,
+                   partition_name, dev_sz);
+    } else {
+        int num_threads = 2;
+        size_t num_blocks = dev_sz >> BLOCK_SHIFT;
+        size_t num_blocks_per_thread = num_blocks / num_threads;
+        size_t read_sz_per_thread = num_blocks_per_thread << BLOCK_SHIFT;
+        off_t offset = 0;
 
-    for (int i = 0; i < num_threads; i++) {
-        std::async(std::launch::async, &SnapshotHandler::ReadBlocksToCache, this, dm_block_device,
-                   partition_name, offset, read_sz_per_thread);
+        for (int i = 0; i < num_threads; i++) {
+            std::async(std::launch::async, &SnapshotHandler::ReadBlocksToCache, this,
+                       dm_block_device, partition_name, offset, read_sz_per_thread);
 
-        offset += read_sz_per_thread;
+            offset += read_sz_per_thread;
+        }
     }
 }
 
@@ -513,5 +652,33 @@
     return ra_state;
 }
 
+bool SnapshotHandler::IsIouringSupported() {
+    struct utsname uts;
+    unsigned int major, minor;
+
+    if (android::base::GetBoolProperty("snapuserd.test.io_uring.force_disable", false)) {
+        SNAP_LOG(INFO) << "io_uring disabled for testing";
+        return false;
+    }
+
+    if ((uname(&uts) != 0) || (sscanf(uts.release, "%u.%u", &major, &minor) != 2)) {
+        SNAP_LOG(ERROR) << "Could not parse the kernel version from uname. "
+                        << " io_uring not supported";
+        return false;
+    }
+
+    // We will only support kernels from 5.6 onwards as IOSQE_ASYNC flag and
+    // IO_URING_OP_READ/WRITE opcodes were introduced only on 5.6 kernel
+    if (major >= 5) {
+        if (major == 5 && minor < 6) {
+            return false;
+        }
+    } else {
+        return false;
+    }
+
+    return android::base::GetBoolProperty("ro.virtual_ab.io_uring.enabled", false);
+}
+
 }  // namespace snapshot
 }  // namespace android
diff --git a/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_core.h b/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_core.h
index 1953316..b0f2d65 100644
--- a/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_core.h
+++ b/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_core.h
@@ -39,6 +39,7 @@
 #include <libdm/dm.h>
 #include <libsnapshot/cow_reader.h>
 #include <libsnapshot/cow_writer.h>
+#include <liburing.h>
 #include <snapuserd/snapuserd_buffer.h>
 #include <snapuserd/snapuserd_kernel.h>
 
@@ -113,6 +114,19 @@
     bool ReconstructDataFromCow();
     void CheckOverlap(const CowOperation* cow_op);
 
+    bool ReadAheadAsyncIO();
+    bool ReapIoCompletions(int pending_ios_to_complete);
+    bool ReadXorData(size_t block_index, size_t xor_op_index,
+                     std::vector<const CowOperation*>& xor_op_vec);
+    void ProcessXorData(size_t& block_xor_index, size_t& xor_index,
+                        std::vector<const CowOperation*>& xor_op_vec, void* buffer,
+                        loff_t& buffer_offset);
+    void UpdateScratchMetadata();
+
+    bool ReadAheadSyncIO();
+    bool InitializeIouring();
+    void FinalizeIouring();
+
     void* read_ahead_buffer_;
     void* metadata_buffer_;
 
@@ -131,7 +145,19 @@
     std::unordered_set<uint64_t> dest_blocks_;
     std::unordered_set<uint64_t> source_blocks_;
     bool overlap_;
+    std::vector<uint64_t> blocks_;
+    int total_blocks_merged_ = 0;
+    std::unique_ptr<uint8_t[]> ra_temp_buffer_;
+    std::unique_ptr<uint8_t[]> ra_temp_meta_buffer_;
     BufferSink bufsink_;
+
+    bool read_ahead_async_ = false;
+    // Queue depth of 32 seems optimal. We don't want
+    // to have a huge depth as it may put more memory pressure
+    // on the kernel worker threads given that we use
+    // IOSQE_ASYNC flag.
+    int queue_depth_ = 32;
+    std::unique_ptr<struct io_uring> ring_;
 };
 
 class Worker {
@@ -185,6 +211,7 @@
     // Merge related ops
     bool Merge();
     bool MergeOrderedOps(const std::unique_ptr<ICowOpIter>& cowop_iter);
+    bool MergeOrderedOpsAsync(const std::unique_ptr<ICowOpIter>& cowop_iter);
     bool MergeReplaceZeroOps(const std::unique_ptr<ICowOpIter>& cowop_iter);
     int PrepareMerge(uint64_t* source_offset, int* pending_ops,
                      const std::unique_ptr<ICowOpIter>& cowop_iter,
@@ -193,6 +220,9 @@
     sector_t ChunkToSector(chunk_t chunk) { return chunk << CHUNK_SHIFT; }
     chunk_t SectorToChunk(sector_t sector) { return sector >> CHUNK_SHIFT; }
 
+    bool InitializeIouring();
+    void FinalizeIouring();
+
     std::unique_ptr<CowReader> reader_;
     BufferSink bufsink_;
     XorSink xorsink_;
@@ -208,6 +238,14 @@
     unique_fd base_path_merge_fd_;
     unique_fd ctrl_fd_;
 
+    bool merge_async_ = false;
+    // Queue depth of 32 seems optimal. We don't want
+    // to have a huge depth as it may put more memory pressure
+    // on the kernel worker threads given that we use
+    // IOSQE_ASYNC flag.
+    int queue_depth_ = 32;
+    std::unique_ptr<struct io_uring> ring_;
+
     std::shared_ptr<SnapshotHandler> snapuserd_;
 };
 
@@ -292,6 +330,8 @@
     bool GetRABuffer(std::unique_lock<std::mutex>* lock, uint64_t block, void* buffer);
     MERGE_GROUP_STATE ProcessMergingBlock(uint64_t new_block, void* buffer);
 
+    bool IsIouringSupported();
+
   private:
     bool ReadMetadata();
     sector_t ChunkToSector(chunk_t chunk) { return chunk << CHUNK_SHIFT; }
@@ -304,6 +344,11 @@
     void ReadBlocksToCache(const std::string& dm_block_device, const std::string& partition_name,
                            off_t offset, size_t size);
 
+    bool InitializeIouring(int io_depth);
+    void FinalizeIouring();
+    bool ReadBlocksAsync(const std::string& dm_block_device, const std::string& partition_name,
+                         size_t size);
+
     // COW device
     std::string cow_device_;
     // Source device
@@ -352,6 +397,8 @@
     bool attached_ = false;
     bool is_socket_present_;
     bool scratch_space_ = false;
+
+    std::unique_ptr<struct io_uring> ring_;
 };
 
 }  // namespace snapshot
diff --git a/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_merge.cpp b/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_merge.cpp
index fa055b7..d4d4efe 100644
--- a/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_merge.cpp
+++ b/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_merge.cpp
@@ -72,16 +72,16 @@
 }
 
 bool Worker::MergeReplaceZeroOps(const std::unique_ptr<ICowOpIter>& cowop_iter) {
-    // Flush every 2048 ops. Since all ops are independent and there is no
+    // Flush every 8192 ops. Since all ops are independent and there is no
     // dependency between COW ops, we will flush the data and the number
-    // of ops merged in COW file for every 2048 ops. If there is a crash,
+    // of ops merged in COW file for every 8192 ops. If there is a crash,
     // we will end up replaying some of the COW ops which were already merged.
     // That is ok.
     //
-    // Why 2048 ops ? We can probably increase this to bigger value but just
-    // need to ensure that merge makes forward progress if there are
-    // crashes repeatedly which is highly unlikely.
-    int total_ops_merged_per_commit = (PAYLOAD_BUFFER_SZ / BLOCK_SZ) * 8;
+    // Why 8192 ops ? Increasing this may improve merge time 3-4 seconds but
+    // we need to make sure that we checkpoint; 8k ops seems optimal. In-case
+    // if there is a crash merge should always make forward progress.
+    int total_ops_merged_per_commit = (PAYLOAD_BUFFER_SZ / BLOCK_SZ) * 32;
     int num_ops_merged = 0;
 
     while (!cowop_iter->Done()) {
@@ -128,7 +128,7 @@
 
         num_ops_merged += linear_blocks;
 
-        if (num_ops_merged == total_ops_merged_per_commit) {
+        if (num_ops_merged >= total_ops_merged_per_commit) {
             // Flush the data
             if (fsync(base_path_merge_fd_.get()) < 0) {
                 SNAP_LOG(ERROR) << "Merge: ReplaceZeroOps: Failed to fsync merged data";
@@ -172,6 +172,173 @@
     return true;
 }
 
+bool Worker::MergeOrderedOpsAsync(const std::unique_ptr<ICowOpIter>& cowop_iter) {
+    void* mapped_addr = snapuserd_->GetMappedAddr();
+    void* read_ahead_buffer =
+            static_cast<void*>((char*)mapped_addr + snapuserd_->GetBufferDataOffset());
+    size_t block_index = 0;
+
+    SNAP_LOG(INFO) << "MergeOrderedOpsAsync started....";
+
+    while (!cowop_iter->Done()) {
+        const CowOperation* cow_op = &cowop_iter->Get();
+        if (!IsOrderedOp(*cow_op)) {
+            break;
+        }
+
+        SNAP_LOG(DEBUG) << "Waiting for merge begin...";
+        // Wait for RA thread to notify that the merge window
+        // is ready for merging.
+        if (!snapuserd_->WaitForMergeBegin()) {
+            snapuserd_->SetMergeFailed(block_index);
+            return false;
+        }
+
+        snapuserd_->SetMergeInProgress(block_index);
+
+        loff_t offset = 0;
+        int num_ops = snapuserd_->GetTotalBlocksToMerge();
+
+        int pending_sqe = queue_depth_;
+        int pending_ios_to_submit = 0;
+        bool flush_required = false;
+
+        SNAP_LOG(DEBUG) << "Merging copy-ops of size: " << num_ops;
+        while (num_ops) {
+            uint64_t source_offset;
+
+            int linear_blocks = PrepareMerge(&source_offset, &num_ops, cowop_iter);
+
+            if (linear_blocks != 0) {
+                size_t io_size = (linear_blocks * BLOCK_SZ);
+
+                // Get an SQE entry from the ring and populate the I/O variables
+                struct io_uring_sqe* sqe = io_uring_get_sqe(ring_.get());
+                if (!sqe) {
+                    SNAP_PLOG(ERROR) << "io_uring_get_sqe failed during merge-ordered ops";
+                    snapuserd_->SetMergeFailed(block_index);
+                    return false;
+                }
+
+                io_uring_prep_write(sqe, base_path_merge_fd_.get(),
+                                    (char*)read_ahead_buffer + offset, io_size, source_offset);
+
+                offset += io_size;
+                num_ops -= linear_blocks;
+
+                pending_sqe -= 1;
+                pending_ios_to_submit += 1;
+                sqe->flags |= IOSQE_ASYNC;
+            }
+
+            // Ring is full or no more COW ops to be merged in this batch
+            if (pending_sqe == 0 || num_ops == 0 || (linear_blocks == 0 && pending_ios_to_submit)) {
+                // If this is a last set of COW ops to be merged in this batch, we need
+                // to sync the merged data. We will try to grab an SQE entry
+                // and set the FSYNC command; additionally, make sure that
+                // the fsync is done after all the I/O operations queued
+                // in the ring is completed by setting IOSQE_IO_DRAIN.
+                //
+                // If there is no space in the ring, we will flush it later
+                // by explicitly calling fsync() system call.
+                if (num_ops == 0 || (linear_blocks == 0 && pending_ios_to_submit)) {
+                    if (pending_sqe != 0) {
+                        struct io_uring_sqe* sqe = io_uring_get_sqe(ring_.get());
+                        if (!sqe) {
+                            // very unlikely but let's continue and not fail the
+                            // merge - we will flush it later
+                            SNAP_PLOG(ERROR) << "io_uring_get_sqe failed during merge-ordered ops";
+                            flush_required = true;
+                        } else {
+                            io_uring_prep_fsync(sqe, base_path_merge_fd_.get(), 0);
+                            // Drain the queue before fsync
+                            io_uring_sqe_set_flags(sqe, IOSQE_IO_DRAIN);
+                            pending_sqe -= 1;
+                            flush_required = false;
+                            pending_ios_to_submit += 1;
+                            sqe->flags |= IOSQE_ASYNC;
+                        }
+                    } else {
+                        flush_required = true;
+                    }
+                }
+
+                // Submit the IO for all the COW ops in a single syscall
+                int ret = io_uring_submit(ring_.get());
+                if (ret != pending_ios_to_submit) {
+                    SNAP_PLOG(ERROR)
+                            << "io_uring_submit failed for read-ahead: "
+                            << " io submit: " << ret << " expected: " << pending_ios_to_submit;
+                    snapuserd_->SetMergeFailed(block_index);
+                    return false;
+                }
+
+                int pending_ios_to_complete = pending_ios_to_submit;
+                pending_ios_to_submit = 0;
+
+                // Reap I/O completions
+                while (pending_ios_to_complete) {
+                    struct io_uring_cqe* cqe;
+
+                    ret = io_uring_wait_cqe(ring_.get(), &cqe);
+                    if (ret) {
+                        SNAP_LOG(ERROR) << "Read-ahead - io_uring_wait_cqe failed: " << ret;
+                        snapuserd_->SetMergeFailed(block_index);
+                        return false;
+                    }
+
+                    if (cqe->res < 0) {
+                        SNAP_LOG(ERROR)
+                                << "Read-ahead - io_uring_Wait_cqe failed with res: " << cqe->res;
+                        snapuserd_->SetMergeFailed(block_index);
+                        return false;
+                    }
+
+                    io_uring_cqe_seen(ring_.get(), cqe);
+                    pending_ios_to_complete -= 1;
+                }
+
+                pending_sqe = queue_depth_;
+            }
+
+            if (linear_blocks == 0) {
+                break;
+            }
+        }
+
+        // Verify all ops are merged
+        CHECK(num_ops == 0);
+
+        // Flush the data
+        if (flush_required && (fsync(base_path_merge_fd_.get()) < 0)) {
+            SNAP_LOG(ERROR) << " Failed to fsync merged data";
+            snapuserd_->SetMergeFailed(block_index);
+            return false;
+        }
+
+        // Merge is done and data is on disk. Update the COW Header about
+        // the merge completion
+        if (!snapuserd_->CommitMerge(snapuserd_->GetTotalBlocksToMerge())) {
+            SNAP_LOG(ERROR) << " Failed to commit the merged block in the header";
+            snapuserd_->SetMergeFailed(block_index);
+            return false;
+        }
+
+        SNAP_LOG(DEBUG) << "Block commit of size: " << snapuserd_->GetTotalBlocksToMerge();
+        // Mark the block as merge complete
+        snapuserd_->SetMergeCompleted(block_index);
+
+        // Notify RA thread that the merge thread is ready to merge the next
+        // window
+        snapuserd_->NotifyRAForMergeReady();
+
+        // Get the next block
+        block_index += 1;
+    }
+
+    return true;
+}
+
 bool Worker::MergeOrderedOps(const std::unique_ptr<ICowOpIter>& cowop_iter) {
     void* mapped_addr = snapuserd_->GetMappedAddr();
     void* read_ahead_buffer =
@@ -260,15 +427,23 @@
 bool Worker::Merge() {
     std::unique_ptr<ICowOpIter> cowop_iter = reader_->GetMergeOpIter();
 
-    // Start with Copy and Xor ops
-    if (!MergeOrderedOps(cowop_iter)) {
-        SNAP_LOG(ERROR) << "Merge failed for ordered ops";
-        snapuserd_->MergeFailed();
-        return false;
+    if (merge_async_) {
+        if (!MergeOrderedOpsAsync(cowop_iter)) {
+            SNAP_LOG(ERROR) << "Merge failed for ordered ops";
+            snapuserd_->MergeFailed();
+            return false;
+        }
+        SNAP_LOG(INFO) << "MergeOrderedOpsAsync completed.....";
+    } else {
+        // Start with Copy and Xor ops
+        if (!MergeOrderedOps(cowop_iter)) {
+            SNAP_LOG(ERROR) << "Merge failed for ordered ops";
+            snapuserd_->MergeFailed();
+            return false;
+        }
+        SNAP_LOG(INFO) << "MergeOrderedOps completed.....";
     }
 
-    SNAP_LOG(INFO) << "MergeOrderedOps completed...";
-
     // Replace and Zero ops
     if (!MergeReplaceZeroOps(cowop_iter)) {
         SNAP_LOG(ERROR) << "Merge failed for replace/zero ops";
@@ -281,6 +456,31 @@
     return true;
 }
 
+bool Worker::InitializeIouring() {
+    if (!snapuserd_->IsIouringSupported()) {
+        return false;
+    }
+
+    ring_ = std::make_unique<struct io_uring>();
+
+    int ret = io_uring_queue_init(queue_depth_, ring_.get(), 0);
+    if (ret) {
+        LOG(ERROR) << "Merge: io_uring_queue_init failed with ret: " << ret;
+        return false;
+    }
+
+    merge_async_ = true;
+
+    LOG(INFO) << "Merge: io_uring initialized with queue depth: " << queue_depth_;
+    return true;
+}
+
+void Worker::FinalizeIouring() {
+    if (merge_async_) {
+        io_uring_queue_exit(ring_.get());
+    }
+}
+
 bool Worker::RunMergeThread() {
     SNAP_LOG(DEBUG) << "Waiting for merge begin...";
     if (!snapuserd_->WaitForMergeBegin()) {
@@ -296,10 +496,13 @@
         return false;
     }
 
+    InitializeIouring();
+
     if (!Merge()) {
         return false;
     }
 
+    FinalizeIouring();
     CloseFds();
     reader_->CloseCowFd();
 
diff --git a/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_readahead.cpp b/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_readahead.cpp
index 9e8ccfb..26c5f19 100644
--- a/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_readahead.cpp
+++ b/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_readahead.cpp
@@ -183,25 +183,311 @@
     return true;
 }
 
-bool ReadAhead::ReadAheadIOStart() {
-    // Check if the data has to be constructed from the COW file.
-    // This will be true only once during boot up after a crash
-    // during merge.
-    if (snapuserd_->ShouldReconstructDataFromCow()) {
-        return ReconstructDataFromCow();
-    }
+/*
+ * With io_uring, the data flow is slightly different.
+ *
+ * The data flow is as follows:
+ *
+ * 1: Queue the I/O requests to be read from backing source device.
+ * This is done by retrieving the SQE entry from ring and populating
+ * the SQE entry. Note that the I/O is not submitted yet.
+ *
+ * 2: Once the ring is full (aka queue_depth), we will submit all
+ * the queued I/O request with a single system call. This essentially
+ * cuts down "queue_depth" number of system calls to a single system call.
+ *
+ * 3: Once the I/O is submitted, user-space thread will now work
+ * on processing the XOR Operations. This happens in parallel when
+ * I/O requests are submitted to the kernel. This is ok because, for XOR
+ * operations, we first need to retrieve the compressed data form COW block
+ * device. Thus, we have offloaded the backing source I/O to the kernel
+ * and user-space is parallely working on fetching the data for XOR operations.
+ *
+ * 4: After the XOR operations are read from COW device, poll the completion
+ * queue for all the I/O submitted. If the I/O's were already completed,
+ * then user-space thread will just read the CQE requests from the ring
+ * without doing any system call. If none of the I/O were completed yet,
+ * user-space thread will do a system call and wait for I/O completions.
+ *
+ * Flow diagram:
+ *                                                    SQ-RING
+ *  SQE1 <----------- Fetch SQE1 Entry ---------- |SQE1||SQE2|SQE3|
+ *
+ *  SQE1  ------------ Populate SQE1 Entry ------> |SQE1-X||SQE2|SQE3|
+ *
+ *  SQE2 <----------- Fetch SQE2 Entry ---------- |SQE1-X||SQE2|SQE3|
+ *
+ *  SQE2  ------------ Populate SQE2 Entry ------> |SQE1-X||SQE2-X|SQE3|
+ *
+ *  SQE3 <----------- Fetch SQE3 Entry ---------- |SQE1-X||SQE2-X|SQE3|
+ *
+ *  SQE3  ------------ Populate SQE3 Entry ------> |SQE1-X||SQE2-X|SQE3-X|
+ *
+ *  Submit-IO ---------------------------------> |SQE1-X||SQE2-X|SQE3-X|
+ *     |                                                  |
+ *     |                                        Process I/O entries in kernel
+ *     |                                                  |
+ *  Retrieve XOR                                          |
+ *  data from COW                                         |
+ *     |                                                  |
+ *     |                                                  |
+ *  Fetch CQ completions
+ *     |                                              CQ-RING
+ *                                               |CQE1-X||CQE2-X|CQE3-X|
+ *                                                        |
+ *   CQE1 <------------Fetch CQE1 Entry          |CQE1||CQE2-X|CQE3-X|
+ *   CQE2 <------------Fetch CQE2 Entry          |CQE1||CQE2-|CQE3-X|
+ *   CQE3 <------------Fetch CQE3 Entry          |CQE1||CQE2-|CQE3-|
+ *    |
+ *    |
+ *  Continue Next set of operations in the RING
+ */
 
-    std::vector<uint64_t> blocks;
-
+bool ReadAhead::ReadAheadAsyncIO() {
     int num_ops = (snapuserd_->GetBufferDataSize()) / BLOCK_SZ;
     loff_t buffer_offset = 0;
-    int total_blocks_merged = 0;
+    total_blocks_merged_ = 0;
     overlap_ = false;
     dest_blocks_.clear();
     source_blocks_.clear();
+    blocks_.clear();
     std::vector<const CowOperation*> xor_op_vec;
 
-    auto ra_temp_buffer = std::make_unique<uint8_t[]>(snapuserd_->GetBufferDataSize());
+    int pending_sqe = queue_depth_;
+    int pending_ios_to_submit = 0;
+
+    size_t xor_op_index = 0;
+    size_t block_index = 0;
+
+    loff_t offset = 0;
+
+    bufsink_.ResetBufferOffset();
+
+    // Number of ops to be merged in this window. This is a fixed size
+    // except for the last window wherein the number of ops can be less
+    // than the size of the RA window.
+    while (num_ops) {
+        uint64_t source_offset;
+        struct io_uring_sqe* sqe;
+
+        int linear_blocks = PrepareNextReadAhead(&source_offset, &num_ops, blocks_, xor_op_vec);
+
+        if (linear_blocks != 0) {
+            size_t io_size = (linear_blocks * BLOCK_SZ);
+
+            // Get an SQE entry from the ring and populate the I/O variables
+            sqe = io_uring_get_sqe(ring_.get());
+            if (!sqe) {
+                SNAP_PLOG(ERROR) << "io_uring_get_sqe failed during read-ahead";
+                snapuserd_->ReadAheadIOFailed();
+                return false;
+            }
+
+            io_uring_prep_read(sqe, backing_store_fd_.get(),
+                               (char*)ra_temp_buffer_.get() + buffer_offset, io_size,
+                               source_offset);
+
+            buffer_offset += io_size;
+            num_ops -= linear_blocks;
+            total_blocks_merged_ += linear_blocks;
+
+            pending_sqe -= 1;
+            pending_ios_to_submit += 1;
+            sqe->flags |= IOSQE_ASYNC;
+        }
+
+        // pending_sqe == 0 : Ring is full
+        //
+        // num_ops == 0 : All the COW ops in this batch are processed - Submit
+        // pending I/O requests in the ring
+        //
+        // linear_blocks == 0 : All the COW ops processing is done. Submit
+        // pending I/O requests in the ring
+        if (pending_sqe == 0 || num_ops == 0 || (linear_blocks == 0 && pending_ios_to_submit)) {
+            // Submit the IO for all the COW ops in a single syscall
+            int ret = io_uring_submit(ring_.get());
+            if (ret != pending_ios_to_submit) {
+                SNAP_PLOG(ERROR) << "io_uring_submit failed for read-ahead: "
+                                 << " io submit: " << ret << " expected: " << pending_ios_to_submit;
+                snapuserd_->ReadAheadIOFailed();
+                return false;
+            }
+
+            int pending_ios_to_complete = pending_ios_to_submit;
+            pending_ios_to_submit = 0;
+
+            bool xor_processing_required = (xor_op_vec.size() > 0);
+
+            // Read XOR data from COW file in parallel when I/O's are in-flight
+            if (xor_processing_required && !ReadXorData(block_index, xor_op_index, xor_op_vec)) {
+                SNAP_LOG(ERROR) << "ReadXorData failed";
+                snapuserd_->ReadAheadIOFailed();
+                return false;
+            }
+
+            // Fetch I/O completions
+            if (!ReapIoCompletions(pending_ios_to_complete)) {
+                SNAP_LOG(ERROR) << "ReapIoCompletions failed";
+                snapuserd_->ReadAheadIOFailed();
+                return false;
+            }
+
+            // Retrieve XOR'ed data
+            if (xor_processing_required) {
+                ProcessXorData(block_index, xor_op_index, xor_op_vec, ra_temp_buffer_.get(),
+                               offset);
+            }
+
+            // All the I/O in the ring is processed.
+            pending_sqe = queue_depth_;
+        }
+
+        if (linear_blocks == 0) {
+            break;
+        }
+    }
+
+    // Done with merging ordered ops
+    if (RAIterDone() && total_blocks_merged_ == 0) {
+        return true;
+    }
+
+    CHECK(blocks_.size() == total_blocks_merged_);
+
+    UpdateScratchMetadata();
+
+    return true;
+}
+
+void ReadAhead::UpdateScratchMetadata() {
+    loff_t metadata_offset = 0;
+
+    struct ScratchMetadata* bm = reinterpret_cast<struct ScratchMetadata*>(
+            (char*)ra_temp_meta_buffer_.get() + metadata_offset);
+
+    bm->new_block = 0;
+    bm->file_offset = 0;
+
+    loff_t file_offset = snapuserd_->GetBufferDataOffset();
+
+    for (size_t block_index = 0; block_index < blocks_.size(); block_index++) {
+        uint64_t new_block = blocks_[block_index];
+        // Track the metadata blocks which are stored in scratch space
+        bm = reinterpret_cast<struct ScratchMetadata*>((char*)ra_temp_meta_buffer_.get() +
+                                                       metadata_offset);
+
+        bm->new_block = new_block;
+        bm->file_offset = file_offset;
+
+        metadata_offset += sizeof(struct ScratchMetadata);
+        file_offset += BLOCK_SZ;
+    }
+
+    // This is important - explicitly set the contents to zero. This is used
+    // when re-constructing the data after crash. This indicates end of
+    // reading metadata contents when re-constructing the data
+    bm = reinterpret_cast<struct ScratchMetadata*>((char*)ra_temp_meta_buffer_.get() +
+                                                   metadata_offset);
+    bm->new_block = 0;
+    bm->file_offset = 0;
+}
+
+bool ReadAhead::ReapIoCompletions(int pending_ios_to_complete) {
+    // Reap I/O completions
+    while (pending_ios_to_complete) {
+        struct io_uring_cqe* cqe;
+
+        int ret = io_uring_wait_cqe(ring_.get(), &cqe);
+        if (ret) {
+            SNAP_LOG(ERROR) << "Read-ahead - io_uring_wait_cqe failed: " << ret;
+            return false;
+        }
+
+        if (cqe->res < 0) {
+            SNAP_LOG(ERROR) << "Read-ahead - io_uring_Wait_cqe failed with res: " << cqe->res;
+            return false;
+        }
+
+        io_uring_cqe_seen(ring_.get(), cqe);
+        pending_ios_to_complete -= 1;
+    }
+
+    return true;
+}
+
+void ReadAhead::ProcessXorData(size_t& block_xor_index, size_t& xor_index,
+                               std::vector<const CowOperation*>& xor_op_vec, void* buffer,
+                               loff_t& buffer_offset) {
+    loff_t xor_buf_offset = 0;
+
+    while (block_xor_index < blocks_.size()) {
+        void* bufptr = static_cast<void*>((char*)buffer + buffer_offset);
+        uint64_t new_block = blocks_[block_xor_index];
+
+        if (xor_index < xor_op_vec.size()) {
+            const CowOperation* xor_op = xor_op_vec[xor_index];
+
+            // Check if this block is an XOR op
+            if (xor_op->new_block == new_block) {
+                // Pointer to the data read from base device
+                uint8_t* buffer = reinterpret_cast<uint8_t*>(bufptr);
+                // Get the xor'ed data read from COW device
+                uint8_t* xor_data = reinterpret_cast<uint8_t*>((char*)bufsink_.GetPayloadBufPtr() +
+                                                               xor_buf_offset);
+
+                for (size_t byte_offset = 0; byte_offset < BLOCK_SZ; byte_offset++) {
+                    buffer[byte_offset] ^= xor_data[byte_offset];
+                }
+
+                // Move to next XOR op
+                xor_index += 1;
+                xor_buf_offset += BLOCK_SZ;
+            }
+        }
+
+        buffer_offset += BLOCK_SZ;
+        block_xor_index += 1;
+    }
+
+    bufsink_.ResetBufferOffset();
+}
+
+bool ReadAhead::ReadXorData(size_t block_index, size_t xor_op_index,
+                            std::vector<const CowOperation*>& xor_op_vec) {
+    // Process the XOR ops in parallel - We will be reading data
+    // from COW file for XOR ops processing.
+    while (block_index < blocks_.size()) {
+        uint64_t new_block = blocks_[block_index];
+
+        if (xor_op_index < xor_op_vec.size()) {
+            const CowOperation* xor_op = xor_op_vec[xor_op_index];
+            if (xor_op->new_block == new_block) {
+                if (!reader_->ReadData(*xor_op, &bufsink_)) {
+                    SNAP_LOG(ERROR)
+                            << " ReadAhead - XorOp Read failed for block: " << xor_op->new_block;
+                    return false;
+                }
+
+                xor_op_index += 1;
+                bufsink_.UpdateBufferOffset(BLOCK_SZ);
+            }
+        }
+        block_index += 1;
+    }
+    return true;
+}
+
+bool ReadAhead::ReadAheadSyncIO() {
+    int num_ops = (snapuserd_->GetBufferDataSize()) / BLOCK_SZ;
+    loff_t buffer_offset = 0;
+    total_blocks_merged_ = 0;
+    overlap_ = false;
+    dest_blocks_.clear();
+    source_blocks_.clear();
+    blocks_.clear();
+    std::vector<const CowOperation*> xor_op_vec;
+
+    bufsink_.ResetBufferOffset();
 
     // Number of ops to be merged in this window. This is a fixed size
     // except for the last window wherein the number of ops can be less
@@ -209,7 +495,7 @@
     while (num_ops) {
         uint64_t source_offset;
 
-        int linear_blocks = PrepareNextReadAhead(&source_offset, &num_ops, blocks, xor_op_vec);
+        int linear_blocks = PrepareNextReadAhead(&source_offset, &num_ops, blocks_, xor_op_vec);
         if (linear_blocks == 0) {
             // No more blocks to read
             SNAP_LOG(DEBUG) << " Read-ahead completed....";
@@ -220,7 +506,7 @@
 
         // Read from the base device consecutive set of blocks in one shot
         if (!android::base::ReadFullyAtOffset(backing_store_fd_,
-                                              (char*)ra_temp_buffer.get() + buffer_offset, io_size,
+                                              (char*)ra_temp_buffer_.get() + buffer_offset, io_size,
                                               source_offset)) {
             SNAP_PLOG(ERROR) << "Ordered-op failed. Read from backing store: "
                              << backing_store_device_ << "at block :" << source_offset / BLOCK_SZ
@@ -233,21 +519,19 @@
         }
 
         buffer_offset += io_size;
-        total_blocks_merged += linear_blocks;
+        total_blocks_merged_ += linear_blocks;
         num_ops -= linear_blocks;
     }
 
     // Done with merging ordered ops
-    if (RAIterDone() && total_blocks_merged == 0) {
+    if (RAIterDone() && total_blocks_merged_ == 0) {
         return true;
     }
 
     loff_t metadata_offset = 0;
 
-    auto ra_temp_meta_buffer = std::make_unique<uint8_t[]>(snapuserd_->GetBufferMetadataSize());
-
     struct ScratchMetadata* bm = reinterpret_cast<struct ScratchMetadata*>(
-            (char*)ra_temp_meta_buffer.get() + metadata_offset);
+            (char*)ra_temp_meta_buffer_.get() + metadata_offset);
 
     bm->new_block = 0;
     bm->file_offset = 0;
@@ -255,12 +539,15 @@
     loff_t file_offset = snapuserd_->GetBufferDataOffset();
 
     loff_t offset = 0;
-    CHECK(blocks.size() == total_blocks_merged);
+    CHECK(blocks_.size() == total_blocks_merged_);
 
     size_t xor_index = 0;
-    for (size_t block_index = 0; block_index < blocks.size(); block_index++) {
-        void* bufptr = static_cast<void*>((char*)ra_temp_buffer.get() + offset);
-        uint64_t new_block = blocks[block_index];
+    BufferSink bufsink;
+    bufsink.Initialize(BLOCK_SZ * 2);
+
+    for (size_t block_index = 0; block_index < blocks_.size(); block_index++) {
+        void* bufptr = static_cast<void*>((char*)ra_temp_buffer_.get() + offset);
+        uint64_t new_block = blocks_[block_index];
 
         if (xor_index < xor_op_vec.size()) {
             const CowOperation* xor_op = xor_op_vec[xor_index];
@@ -268,17 +555,16 @@
             // Check if this block is an XOR op
             if (xor_op->new_block == new_block) {
                 // Read the xor'ed data from COW
-                if (!reader_->ReadData(*xor_op, &bufsink_)) {
+                if (!reader_->ReadData(*xor_op, &bufsink)) {
                     SNAP_LOG(ERROR)
                             << " ReadAhead - XorOp Read failed for block: " << xor_op->new_block;
                     snapuserd_->ReadAheadIOFailed();
                     return false;
                 }
-
                 // Pointer to the data read from base device
                 uint8_t* buffer = reinterpret_cast<uint8_t*>(bufptr);
                 // Get the xor'ed data read from COW device
-                uint8_t* xor_data = reinterpret_cast<uint8_t*>(bufsink_.GetPayloadBufPtr());
+                uint8_t* xor_data = reinterpret_cast<uint8_t*>(bufsink.GetPayloadBufPtr());
 
                 // Retrieve the original data
                 for (size_t byte_offset = 0; byte_offset < BLOCK_SZ; byte_offset++) {
@@ -292,7 +578,7 @@
 
         offset += BLOCK_SZ;
         // Track the metadata blocks which are stored in scratch space
-        bm = reinterpret_cast<struct ScratchMetadata*>((char*)ra_temp_meta_buffer.get() +
+        bm = reinterpret_cast<struct ScratchMetadata*>((char*)ra_temp_meta_buffer_.get() +
                                                        metadata_offset);
 
         bm->new_block = new_block;
@@ -308,11 +594,34 @@
     // This is important - explicitly set the contents to zero. This is used
     // when re-constructing the data after crash. This indicates end of
     // reading metadata contents when re-constructing the data
-    bm = reinterpret_cast<struct ScratchMetadata*>((char*)ra_temp_meta_buffer.get() +
+    bm = reinterpret_cast<struct ScratchMetadata*>((char*)ra_temp_meta_buffer_.get() +
                                                    metadata_offset);
     bm->new_block = 0;
     bm->file_offset = 0;
 
+    return true;
+}
+
+bool ReadAhead::ReadAheadIOStart() {
+    // Check if the data has to be constructed from the COW file.
+    // This will be true only once during boot up after a crash
+    // during merge.
+    if (snapuserd_->ShouldReconstructDataFromCow()) {
+        return ReconstructDataFromCow();
+    }
+
+    if (read_ahead_async_) {
+        if (!ReadAheadAsyncIO()) {
+            SNAP_LOG(ERROR) << "ReadAheadAsyncIO failed - io_uring processing failure.";
+            return false;
+        }
+    } else {
+        if (!ReadAheadSyncIO()) {
+            SNAP_LOG(ERROR) << "ReadAheadSyncIO failed";
+            return false;
+        }
+    }
+
     // Wait for the merge to finish for the previous RA window. We shouldn't
     // be touching the scratch space until merge is complete of previous RA
     // window. If there is a crash during this time frame, merge should resume
@@ -322,22 +631,22 @@
     }
 
     // Copy the data to scratch space
-    memcpy(metadata_buffer_, ra_temp_meta_buffer.get(), snapuserd_->GetBufferMetadataSize());
-    memcpy(read_ahead_buffer_, ra_temp_buffer.get(), total_blocks_merged * BLOCK_SZ);
+    memcpy(metadata_buffer_, ra_temp_meta_buffer_.get(), snapuserd_->GetBufferMetadataSize());
+    memcpy(read_ahead_buffer_, ra_temp_buffer_.get(), total_blocks_merged_ * BLOCK_SZ);
 
-    offset = 0;
+    loff_t offset = 0;
     std::unordered_map<uint64_t, void*>& read_ahead_buffer_map = snapuserd_->GetReadAheadMap();
     read_ahead_buffer_map.clear();
 
-    for (size_t block_index = 0; block_index < blocks.size(); block_index++) {
+    for (size_t block_index = 0; block_index < blocks_.size(); block_index++) {
         void* bufptr = static_cast<void*>((char*)read_ahead_buffer_ + offset);
-        uint64_t new_block = blocks[block_index];
+        uint64_t new_block = blocks_[block_index];
 
         read_ahead_buffer_map[new_block] = bufptr;
         offset += BLOCK_SZ;
     }
 
-    snapuserd_->SetMergedBlockCountForNextCommit(total_blocks_merged);
+    snapuserd_->SetMergedBlockCountForNextCommit(total_blocks_merged_);
 
     // Flush the data only if we have a overlapping blocks in the region
     // Notify the Merge thread to resume merging this window
@@ -350,6 +659,33 @@
     return true;
 }
 
+bool ReadAhead::InitializeIouring() {
+    if (!snapuserd_->IsIouringSupported()) {
+        return false;
+    }
+
+    ring_ = std::make_unique<struct io_uring>();
+
+    int ret = io_uring_queue_init(queue_depth_, ring_.get(), 0);
+    if (ret) {
+        SNAP_LOG(ERROR) << "io_uring_queue_init failed with ret: " << ret;
+        return false;
+    }
+
+    // For xor ops processing
+    bufsink_.Initialize(PAYLOAD_BUFFER_SZ * 2);
+    read_ahead_async_ = true;
+
+    SNAP_LOG(INFO) << "Read-ahead: io_uring initialized with queue depth: " << queue_depth_;
+    return true;
+}
+
+void ReadAhead::FinalizeIouring() {
+    if (read_ahead_async_) {
+        io_uring_queue_exit(ring_.get());
+    }
+}
+
 bool ReadAhead::RunThread() {
     if (!InitializeFds()) {
         return false;
@@ -363,14 +699,18 @@
 
     InitializeRAIter();
 
+    InitializeIouring();
+
     while (!RAIterDone()) {
         if (!ReadAheadIOStart()) {
             break;
         }
     }
 
+    FinalizeIouring();
     CloseFds();
     reader_->CloseCowFd();
+
     SNAP_LOG(INFO) << " ReadAhead thread terminating....";
     return true;
 }
@@ -434,8 +774,9 @@
     metadata_buffer_ =
             static_cast<void*>((char*)mapped_addr + snapuserd_->GetBufferMetadataOffset());
     read_ahead_buffer_ = static_cast<void*>((char*)mapped_addr + snapuserd_->GetBufferDataOffset());
-    // For xor ops
-    bufsink_.Initialize(PAYLOAD_BUFFER_SZ);
+
+    ra_temp_buffer_ = std::make_unique<uint8_t[]>(snapuserd_->GetBufferDataSize());
+    ra_temp_meta_buffer_ = std::make_unique<uint8_t[]>(snapuserd_->GetBufferMetadataSize());
 }
 
 }  // namespace snapshot
diff --git a/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_test.cpp b/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_test.cpp
index 1c3e04b..d670f1e 100644
--- a/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_test.cpp
+++ b/fs_mgr/libsnapshot/snapuserd/user-space-merge/snapuserd_test.cpp
@@ -12,6 +12,9 @@
 // See the License for the specific language governing permissions and
 // limitations under the License.
 
+#include <android-base/strings.h>
+#include <gflags/gflags.h>
+
 #include <fcntl.h>
 #include <linux/fs.h>
 #include <linux/memfd.h>
@@ -27,6 +30,7 @@
 #include <string_view>
 
 #include <android-base/file.h>
+#include <android-base/properties.h>
 #include <android-base/unique_fd.h>
 #include <fs_mgr/file_wait.h>
 #include <gtest/gtest.h>
@@ -38,6 +42,8 @@
 
 #include "snapuserd_core.h"
 
+DEFINE_string(force_config, "", "Force testing mode with iouring disabled");
+
 namespace android {
 namespace snapshot {
 
@@ -857,5 +863,23 @@
 
 int main(int argc, char** argv) {
     ::testing::InitGoogleTest(&argc, argv);
-    return RUN_ALL_TESTS();
+
+    gflags::ParseCommandLineFlags(&argc, &argv, false);
+
+    android::base::SetProperty("ctl.stop", "snapuserd");
+
+    if (FLAGS_force_config == "iouring_disabled") {
+        if (!android::base::SetProperty("snapuserd.test.io_uring.force_disable", "1")) {
+            return testing::AssertionFailure()
+                   << "Failed to disable property: snapuserd.test.io_uring.disabled";
+        }
+    }
+
+    int ret = RUN_ALL_TESTS();
+
+    if (FLAGS_force_config == "iouring_disabled") {
+        android::base::SetProperty("snapuserd.test.io_uring.force_disable", "0");
+    }
+
+    return ret;
 }
diff --git a/init/init.cpp b/init/init.cpp
index e3596cb..1df4c44 100644
--- a/init/init.cpp
+++ b/init/init.cpp
@@ -33,6 +33,7 @@
 #define _REALLY_INCLUDE_SYS__SYSTEM_PROPERTIES_H_
 #include <sys/_system_properties.h>
 
+#include <filesystem>
 #include <functional>
 #include <map>
 #include <memory>
@@ -46,6 +47,7 @@
 #include <android-base/logging.h>
 #include <android-base/parseint.h>
 #include <android-base/properties.h>
+#include <android-base/scopeguard.h>
 #include <android-base/stringprintf.h>
 #include <android-base/strings.h>
 #include <backtrace/Backtrace.h>
@@ -773,6 +775,82 @@
     return {};
 }
 
+static bool SystemReadSmokeTest() {
+    std::string dev = "/dev/block/mapper/system"s + fs_mgr_get_slot_suffix();
+    android::base::unique_fd fd(open(dev.c_str(), O_RDONLY));
+    if (fd < 0) {
+        PLOG(ERROR) << "open " << dev << " failed, will not diangose snapuserd hangs";
+        return false;
+    }
+
+    for (size_t i = 1; i <= 100; i++) {
+        // Skip around the partition a bit.
+        size_t offset = i * 4096 * 512;
+
+        char b;
+        ssize_t n = TEMP_FAILURE_RETRY(pread(fd.get(), &b, 1, offset));
+        if (n < 0) {
+            PLOG(ERROR) << "snapuserd smoke test read failed";
+            return false;
+        }
+    }
+    return true;
+}
+
+static void DiagnoseSnapuserdHang(pid_t pid) {
+    bool succeeded = false;
+
+    std::mutex m;
+    std::condition_variable cv;
+
+    // Enforce an ordering between this and the thread startup, by taking the
+    // lock before we lanuch the thread.
+    std::unique_lock<std::mutex> cv_lock(m);
+
+    std::thread t([&]() -> void {
+        std::lock_guard<std::mutex> lock(m);
+        succeeded = SystemReadSmokeTest();
+        cv.notify_all();
+    });
+
+    auto join = android::base::make_scope_guard([&]() -> void {
+        // If the smoke test is hung, then this will too. We expect the device to
+        // automatically reboot once the watchdog kicks in.
+        t.join();
+    });
+
+    auto now = std::chrono::system_clock::now();
+    auto deadline = now + 10s;
+    auto status = cv.wait_until(cv_lock, deadline);
+    if (status == std::cv_status::timeout) {
+        LOG(ERROR) << "snapuserd smoke test timed out";
+    } else if (!succeeded) {
+        LOG(ERROR) << "snapuserd smoke test failed";
+    }
+
+    if (succeeded) {
+        LOG(INFO) << "snapuserd smoke test succeeded";
+        return;
+    }
+
+    while (true) {
+        LOG(ERROR) << "snapuserd problem detected, printing open fds";
+
+        std::error_code ec;
+        std::string proc_dir = "/proc/" + std::to_string(pid) + "/fd";
+        for (const auto& entry : std::filesystem::directory_iterator(proc_dir)) {
+            std::string target;
+            if (android::base::Readlink(entry.path(), &target)) {
+                LOG(ERROR) << "snapuserd opened: " << target;
+            } else {
+                LOG(ERROR) << "snapuserd opened: " << entry.path();
+            }
+        }
+
+        std::this_thread::sleep_for(10s);
+    }
+}
+
 int SecondStageMain(int argc, char** argv) {
     if (REBOOT_BOOTLOADER_ON_PANIC) {
         InstallRebootSignalHandlers();
@@ -786,6 +864,11 @@
     InitKernelLogging(argv);
     LOG(INFO) << "init second stage started!";
 
+    if (auto pid = GetSnapuserdFirstStagePid()) {
+        std::thread t(DiagnoseSnapuserdHang, *pid);
+        t.detach();
+    }
+
     // Update $PATH in the case the second stage init is newer than first stage init, where it is
     // first set.
     if (setenv("PATH", _PATH_DEFPATH, 1) != 0) {
diff --git a/init/perfboot.py b/init/perfboot.py
index 4b23ad2..968df38 100755
--- a/init/perfboot.py
+++ b/init/perfboot.py
@@ -1,4 +1,4 @@
-#!/usr/bin/env python
+#!/usr/bin/env python3
 # Copyright (C) 2015 The Android Open Source Project
 #
 # Licensed under the Apache License, Version 2.0 (the "License");
@@ -39,7 +39,7 @@
 
 import argparse
 import atexit
-import cStringIO
+import io
 import glob
 import inspect
 import logging
@@ -102,7 +102,7 @@
             self._wait_cpu_cool_down(self._product, self._temp_paths)
         else:
             if self._waited:
-                print 'Waiting for %d seconds' % self._interval
+                print('Waiting for %d seconds' % self._interval)
                 time.sleep(self._interval)
         self._waited = True
 
@@ -119,9 +119,9 @@
         threshold = IntervalAdjuster._CPU_COOL_DOWN_THRESHOLDS.get(
             self._product)
         if threshold is None:
-            print 'No CPU temperature threshold is set for ' + self._product
-            print ('Just wait %d seconds' %
-                   IntervalAdjuster._CPU_COOL_DOWN_WAIT_TIME_DEFAULT)
+            print('No CPU temperature threshold is set for ' + self._product)
+            print(('Just wait %d seconds' %
+                   IntervalAdjuster._CPU_COOL_DOWN_WAIT_TIME_DEFAULT))
             time.sleep(IntervalAdjuster._CPU_COOL_DOWN_WAIT_TIME_DEFAULT)
             return
         while True:
@@ -129,8 +129,8 @@
             if temp < threshold:
                 logging.info('Current CPU temperature %s' % temp)
                 return
-            print 'Waiting until CPU temperature (%d) falls below %d' % (
-                temp, threshold)
+            print('Waiting until CPU temperature (%d) falls below %d' % (
+                temp, threshold))
             time.sleep(IntervalAdjuster._CPU_COOL_DOWN_WAIT_INTERVAL)
 
 
@@ -260,7 +260,7 @@
 
 def get_values(record, tag):
     """Gets values that matches |tag| from |record|."""
-    keys = [key for key in record.keys() if key[0] == tag]
+    keys = [key for key in list(record.keys()) if key[0] == tag]
     return [record[k] for k in sorted(keys)]
 
 
@@ -304,7 +304,7 @@
     with open(filename, 'w') as f:
         f.write('\t'.join(labels) + '\n')
         for record in record_list:
-            line = cStringIO.StringIO()
+            line = io.StringIO()
             invalid_line = False
             for i, tag in enumerate(tags):
                 if i != 0:
@@ -319,7 +319,7 @@
                 logging.error('Invalid record found: ' + line.getvalue())
             line.write('\n')
             f.write(line.getvalue())
-    print 'Wrote: ' + filename
+    print(('Wrote: ' + filename))
 
 
 def median(data):
@@ -349,9 +349,9 @@
     # Filter out invalid data.
     end_times = [get_last_value(record, end_tag) for record in record_list
                  if get_last_value(record, end_tag) != 0]
-    print 'mean:', int(round(mean(end_times))), 'ms'
-    print 'median:', int(round(median(end_times))), 'ms'
-    print 'standard deviation:', int(round(stddev(end_times))), 'ms'
+    print(('mean:', int(round(mean(end_times))), 'ms'))
+    print(('median:', int(round(median(end_times))), 'ms'))
+    print(('standard deviation:', int(round(stddev(end_times))), 'ms'))
 
 
 def do_iteration(device, interval_adjuster, event_tags_re, end_tag):
@@ -359,7 +359,7 @@
     device.wait()
     interval_adjuster.wait()
     device.reboot()
-    print 'Rebooted the device'
+    print('Rebooted the device, waiting for tag', end_tag)
     record = {}
     booted = False
     while not booted:
@@ -372,7 +372,7 @@
                 stdout=subprocess.PIPE)
         for line in readlines_unbuffered(p):
             if t.is_timedout():
-                print '*** Timed out ***'
+                print('*** Timed out ***')
                 return record
             m = event_tags_re.search(line)
             if not m:
@@ -381,8 +381,8 @@
             event_time = int(m.group('time'))
             pid = m.group('pid')
             record[(tag, pid)] = event_time
-            print 'Event log recorded: %s (%s) - %d ms' % (
-                tag, pid, event_time)
+            print(('Event log recorded: %s (%s) - %d ms' % (
+                tag, pid, event_time)))
             if tag == end_tag:
                 booted = True
                 t.cancel()
@@ -420,7 +420,7 @@
 
 def install_apks(device, apk_dir):
     for apk in glob.glob(os.path.join(apk_dir, '*.apk')):
-        print 'Installing: ' + apk
+        print('Installing: ' + apk)
         device.install(apk, replace=True)
 
 
@@ -452,7 +452,7 @@
     event_tags_re = make_event_tags_re(event_tags)
 
     for i in range(args.iterations):
-        print 'Run #%d ' % i
+        print('Run #%d ' % i)
         record = do_iteration(
             device, interval_adjuster, event_tags_re, end_tag)
         record_list.append(record)
diff --git a/libsparse/Android.bp b/libsparse/Android.bp
index 3f9aeb2..02bfee6 100644
--- a/libsparse/Android.bp
+++ b/libsparse/Android.bp
@@ -96,12 +96,14 @@
 
 cc_fuzz {
     name: "sparse_fuzzer",
-    host_supported: false,
+    host_supported: true,
     srcs: [
         "sparse_fuzzer.cpp",
     ],
     static_libs: [
         "libsparse",
+        "libbase",
+        "libz",
         "liblog",
     ],
 }
diff --git a/libsparse/include/sparse/sparse.h b/libsparse/include/sparse/sparse.h
index 2f75349..9f91269 100644
--- a/libsparse/include/sparse/sparse.h
+++ b/libsparse/include/sparse/sparse.h
@@ -244,21 +244,6 @@
 int sparse_file_read(struct sparse_file *s, int fd, bool sparse, bool crc);
 
 /**
- * sparse_file_read_buf - read a buffer into a sparse file cookie
- *
- * @s - sparse file cookie
- * @buf - buffer to read from
- * @crc - verify the crc of a file in the Android sparse file format
- *
- * Reads a buffer into a sparse file cookie. The buffer must remain
- * valid until the sparse file cookie is freed. If crc is true, the
- * crc of the sparse file will be verified.
- *
- * Returns 0 on success, negative errno on error.
- */
-int sparse_file_read_buf(struct sparse_file *s, char *buf, bool crc);
-
-/**
  * sparse_file_import - import an existing sparse file
  *
  * @fd - file descriptor to read from
@@ -277,6 +262,7 @@
  * sparse_file_import_buf - import an existing sparse file from a buffer
  *
  * @buf - buffer to read from
+ * @len - length of buffer
  * @verbose - print verbose errors while reading the sparse file
  * @crc - verify the crc of a file in the Android sparse file format
  *
@@ -286,7 +272,7 @@
  *
  * Returns a new sparse file cookie on success, NULL on error.
  */
-struct sparse_file *sparse_file_import_buf(char* buf, bool verbose, bool crc);
+struct sparse_file* sparse_file_import_buf(char* buf, size_t len, bool verbose, bool crc);
 
 /**
  * sparse_file_import_auto - import an existing sparse or normal file
diff --git a/libsparse/output_file.cpp b/libsparse/output_file.cpp
index b2c5407..cb5d730 100644
--- a/libsparse/output_file.cpp
+++ b/libsparse/output_file.cpp
@@ -54,6 +54,8 @@
 #define SPARSE_HEADER_LEN (sizeof(sparse_header_t))
 #define CHUNK_HEADER_LEN (sizeof(chunk_header_t))
 
+#define FILL_ZERO_BUFSIZE (2 * 1024 * 1024)
+
 #define container_of(inner, outer_t, elem) ((outer_t*)((char*)(inner)-offsetof(outer_t, elem)))
 
 struct output_file_ops {
@@ -391,13 +393,29 @@
   ret = out->ops->write(out, data, len);
   if (ret < 0) return -1;
   if (zero_len) {
-    ret = out->ops->write(out, out->zero_buf, zero_len);
-    if (ret < 0) return -1;
+    uint64_t len = zero_len;
+    uint64_t write_len;
+    while (len) {
+      write_len = std::min(len, (uint64_t)FILL_ZERO_BUFSIZE);
+      ret = out->ops->write(out, out->zero_buf, write_len);
+      if (ret < 0) {
+        return ret;
+      }
+      len -= write_len;
+    }
   }
 
   if (out->use_crc) {
     out->crc32 = sparse_crc32(out->crc32, data, len);
-    if (zero_len) out->crc32 = sparse_crc32(out->crc32, out->zero_buf, zero_len);
+    if (zero_len) {
+      uint64_t len = zero_len;
+      uint64_t write_len;
+      while (len) {
+        write_len = std::min(len, (uint64_t)FILL_ZERO_BUFSIZE);
+        out->crc32 = sparse_crc32(out->crc32, out->zero_buf, write_len);
+        len -= write_len;
+      }
+    }
   }
 
   out->cur_out_ptr += rnd_up_len;
@@ -460,12 +478,12 @@
   uint64_t write_len;
 
   /* Initialize fill_buf with the fill_val */
-  for (i = 0; i < out->block_size / sizeof(uint32_t); i++) {
+  for (i = 0; i < FILL_ZERO_BUFSIZE / sizeof(uint32_t); i++) {
     out->fill_buf[i] = fill_val;
   }
 
   while (len) {
-    write_len = std::min(len, (uint64_t)out->block_size);
+    write_len = std::min(len, (uint64_t)FILL_ZERO_BUFSIZE);
     ret = out->ops->write(out, out->fill_buf, write_len);
     if (ret < 0) {
       return ret;
@@ -512,13 +530,15 @@
   out->crc32 = 0;
   out->use_crc = crc;
 
-  out->zero_buf = reinterpret_cast<char*>(calloc(block_size, 1));
+  // don't use sparse format block size as it can takes up to 32GB
+  out->zero_buf = reinterpret_cast<char*>(calloc(FILL_ZERO_BUFSIZE, 1));
   if (!out->zero_buf) {
     error_errno("malloc zero_buf");
     return -ENOMEM;
   }
 
-  out->fill_buf = reinterpret_cast<uint32_t*>(calloc(block_size, 1));
+  // don't use sparse format block size as it can takes up to 32GB
+  out->fill_buf = reinterpret_cast<uint32_t*>(calloc(FILL_ZERO_BUFSIZE, 1));
   if (!out->fill_buf) {
     error_errno("malloc fill_buf");
     ret = -ENOMEM;
diff --git a/libsparse/simg_dump.py b/libsparse/simg_dump.py
index b0b3b22..8811a52 100755
--- a/libsparse/simg_dump.py
+++ b/libsparse/simg_dump.py
@@ -120,7 +120,7 @@
                           "output offset", "output blocks", "type", "hash"])
 
     offset = 0
-    for i in xrange(1, total_chunks + 1):
+    for i in range(1, total_chunks + 1):
       header_bin = FH.read(12)
       header = struct.unpack("<2H2I", header_bin)
       chunk_type = header[0]
@@ -159,7 +159,7 @@
           if showhash:
             h = hashlib.sha1()
             data = fill_bin * (blk_sz / 4);
-            for block in xrange(chunk_sz):
+            for block in range(chunk_sz):
               h.update(data)
             curhash = h.hexdigest()
       elif chunk_type == 0xCAC3:
diff --git a/libsparse/sparse_fuzzer.cpp b/libsparse/sparse_fuzzer.cpp
index 42f331f..235d15d 100644
--- a/libsparse/sparse_fuzzer.cpp
+++ b/libsparse/sparse_fuzzer.cpp
@@ -1,16 +1,27 @@
 #include "include/sparse/sparse.h"
 
-extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) {
-  if (size < 2 * sizeof(wchar_t)) return 0;
+static volatile int count;
 
-  int64_t blocksize = 4096;
-  struct sparse_file* file = sparse_file_new(size, blocksize);
-  if (!file) {
+int WriteCallback(void* priv  __attribute__((__unused__)), const void* data, size_t len) {
+  if (!data) {
+    return 0;
+  }
+  if (len == 0) {
     return 0;
   }
 
-  unsigned int block = 1;
-  sparse_file_add_data(file, &data, size, block);
-  sparse_file_destroy(file);
+  const char* p = (const char*)data;
+  // Just to make sure the data is accessible
+  // We only check the head and tail to save time
+  count += *p;
+  count += *(p+len-1);
   return 0;
 }
+
+extern "C" int LLVMFuzzerTestOneInput(const uint8_t* data, size_t size) {
+  struct sparse_file* file = sparse_file_import_buf((char*)data, size, true, false);
+  if (!file) {
+      return 0;
+  }
+  return sparse_file_callback(file, false, false, WriteCallback, nullptr);
+}
diff --git a/libsparse/sparse_read.cpp b/libsparse/sparse_read.cpp
index c4c1823..0f39172 100644
--- a/libsparse/sparse_read.cpp
+++ b/libsparse/sparse_read.cpp
@@ -58,14 +58,15 @@
 
 class SparseFileSource {
  public:
-  /* Seeks the source ahead by the given offset. */
-  virtual void Seek(int64_t offset) = 0;
+  /* Seeks the source ahead by the given offset.
+   * Return 0 if successful. */
+  virtual int Seek(int64_t offset) = 0;
 
   /* Return the current offset. */
   virtual int64_t GetOffset() = 0;
 
-  /* Set the current offset. Return 0 if successful. */
-  virtual int SetOffset(int64_t offset) = 0;
+  /* Rewind to beginning. Return 0 if successful. */
+  virtual int Rewind() = 0;
 
   /* Adds the given length from the current offset of the source to the file at the given block.
    * Return 0 if successful. */
@@ -88,12 +89,14 @@
   SparseFileFdSource(int fd) : fd(fd) {}
   ~SparseFileFdSource() override {}
 
-  void Seek(int64_t off) override { lseek64(fd, off, SEEK_CUR); }
+  int Seek(int64_t off) override {
+    return lseek64(fd, off, SEEK_CUR) != -1 ? 0 : -errno;
+  }
 
   int64_t GetOffset() override { return lseek64(fd, 0, SEEK_CUR); }
 
-  int SetOffset(int64_t offset) override {
-    return lseek64(fd, offset, SEEK_SET) == offset ? 0 : -errno;
+  int Rewind() override {
+    return lseek64(fd, 0, SEEK_SET) == 0 ? 0 : -errno;
   }
 
   int AddToSparseFile(struct sparse_file* s, int64_t len, unsigned int block) override {
@@ -120,39 +123,74 @@
 
 class SparseFileBufSource : public SparseFileSource {
  private:
+  char* buf_start;
+  char* buf_end;
   char* buf;
   int64_t offset;
 
+  int AccessOkay(int64_t len) {
+    if (len <= 0) return -EINVAL;
+    if (buf < buf_start) return -EOVERFLOW;
+    if (buf >= buf_end) return -EOVERFLOW;
+    if (len > buf_end - buf) return -EOVERFLOW;
+
+    return 0;
+  }
+
  public:
-  SparseFileBufSource(char* buf) : buf(buf), offset(0) {}
+  SparseFileBufSource(char* buf, uint64_t len) {
+    this->buf = buf;
+    this->offset = 0;
+    this->buf_start = buf;
+    this->buf_end = buf + len;
+  }
   ~SparseFileBufSource() override {}
 
-  void Seek(int64_t off) override {
+  int Seek(int64_t off) override {
+    int ret = AccessOkay(off);
+    if (ret < 0) {
+      return ret;
+    }
     buf += off;
     offset += off;
+    return 0;
   }
 
   int64_t GetOffset() override { return offset; }
 
-  int SetOffset(int64_t off) override {
-    buf += off - offset;
-    offset = off;
+  int Rewind() override {
+    buf = buf_start;
+    offset = 0;
     return 0;
   }
 
   int AddToSparseFile(struct sparse_file* s, int64_t len, unsigned int block) override {
+    int ret = AccessOkay(len);
+    if (ret < 0) {
+      return ret;
+    }
     return sparse_file_add_data(s, buf, len, block);
   }
 
   int ReadValue(void* ptr, int len) override {
+    int ret = AccessOkay(len);
+    if (ret < 0) {
+      return ret;
+    }
     memcpy(ptr, buf, len);
-    Seek(len);
+    buf += len;
+    offset += len;
     return 0;
   }
 
   int GetCrc32(uint32_t* crc32, int64_t len) override {
+    int ret = AccessOkay(len);
+    if (ret < 0) {
+      return ret;
+    }
     *crc32 = sparse_crc32(*crc32, buf, len);
-    Seek(len);
+    buf += len;
+    offset += len;
     return 0;
   }
 };
@@ -175,7 +213,7 @@
                              SparseFileSource* source, unsigned int blocks, unsigned int block,
                              uint32_t* crc32) {
   int ret;
-  int64_t len = blocks * s->block_size;
+  int64_t len = (int64_t)blocks * s->block_size;
 
   if (chunk_size % s->block_size != 0) {
     return -EINVAL;
@@ -196,7 +234,10 @@
       return ret;
     }
   } else {
-    source->Seek(len);
+    ret = source->Seek(len);
+    if (ret < 0) {
+      return ret;
+    }
   }
 
   return 0;
@@ -379,7 +420,10 @@
     /* Skip the remaining bytes in a header that is longer than
      * we expected.
      */
-    source->Seek(sparse_header.file_hdr_sz - SPARSE_HEADER_LEN);
+    ret = source->Seek(sparse_header.file_hdr_sz - SPARSE_HEADER_LEN);
+    if (ret < 0) {
+      return ret;
+    }
   }
 
   for (i = 0; i < sparse_header.total_chunks; i++) {
@@ -392,7 +436,10 @@
       /* Skip the remaining bytes in a header that is longer than
        * we expected.
        */
-      source->Seek(sparse_header.chunk_hdr_sz - CHUNK_HEADER_LEN);
+      ret = source->Seek(sparse_header.chunk_hdr_sz - CHUNK_HEADER_LEN);
+      if (ret < 0) {
+        return ret;
+      }
     }
 
     ret = process_chunk(s, source, sparse_header.chunk_hdr_sz, &chunk_header, cur_block, crc_ptr);
@@ -474,11 +521,6 @@
   }
 }
 
-int sparse_file_read_buf(struct sparse_file* s, char* buf, bool crc) {
-  SparseFileBufSource source(buf);
-  return sparse_file_read_sparse(s, &source, crc);
-}
-
 static struct sparse_file* sparse_file_import_source(SparseFileSource* source, bool verbose,
                                                      bool crc) {
   int ret;
@@ -510,6 +552,14 @@
     return nullptr;
   }
 
+  if (!sparse_header.blk_sz || (sparse_header.blk_sz % 4)) {
+    return nullptr;
+  }
+
+  if (!sparse_header.total_blks) {
+    return nullptr;
+  }
+
   len = (int64_t)sparse_header.total_blks * sparse_header.blk_sz;
   s = sparse_file_new(sparse_header.blk_sz, len);
   if (!s) {
@@ -517,7 +567,7 @@
     return nullptr;
   }
 
-  ret = source->SetOffset(0);
+  ret = source->Rewind();
   if (ret < 0) {
     verbose_error(verbose, ret, "seeking");
     sparse_file_destroy(s);
@@ -540,8 +590,8 @@
   return sparse_file_import_source(&source, verbose, crc);
 }
 
-struct sparse_file* sparse_file_import_buf(char* buf, bool verbose, bool crc) {
-  SparseFileBufSource source(buf);
+struct sparse_file* sparse_file_import_buf(char* buf, size_t len, bool verbose, bool crc) {
+  SparseFileBufSource source(buf, len);
   return sparse_file_import_source(&source, verbose, crc);
 }
 
diff --git a/libstats/pull_rust/Android.bp b/libstats/pull_rust/Android.bp
index f07e32b..4ffa98d 100644
--- a/libstats/pull_rust/Android.bp
+++ b/libstats/pull_rust/Android.bp
@@ -44,6 +44,11 @@
             ],
         },
     },
+    min_sdk_version: "apex_inherit",
+    apex_available: [
+        "//apex_available:platform",
+        "com.android.virt",
+    ]
 }
 
 rust_library {
diff --git a/libutils/Unicode_test.cpp b/libutils/Unicode_test.cpp
index b92eef8..8b994d9 100644
--- a/libutils/Unicode_test.cpp
+++ b/libutils/Unicode_test.cpp
@@ -100,7 +100,7 @@
         0xF0, 0x90, 0x80, 0x80, // U+10000, 2 UTF-16 character
     };
 
-    char16_t output[1 + 1 + 1 + 2 + 1]; // Room for NULL
+    char16_t output[1 + 1 + 1 + 2 + 1];  // Room for null
 
     utf8_to_utf16(str, sizeof(str), output, sizeof(output) / sizeof(output[0]));
 
@@ -114,8 +114,7 @@
             << "should be first half of surrogate U+10000";
     EXPECT_EQ(0xDC00, output[4])
             << "should be second half of surrogate U+10000";
-    EXPECT_EQ(NULL, output[5])
-            << "should be NULL terminated";
+    EXPECT_EQ(0, output[5]) << "should be null terminated";
 }
 
 TEST_F(UnicodeTest, strstr16EmptyTarget) {
diff --git a/rootdir/Android.mk b/rootdir/Android.mk
index 2ed9eec..ce2ec0e 100644
--- a/rootdir/Android.mk
+++ b/rootdir/Android.mk
@@ -75,14 +75,9 @@
   EXPORT_GLOBAL_GCOV_OPTIONS := export GCOV_PREFIX /data/misc/trace
 endif
 
+EXPORT_GLOBAL_CLANG_COVERAGE_OPTIONS :=
 ifeq ($(CLANG_COVERAGE),true)
-  ifeq ($(BIONIC_COVERAGE),false)
-    # http://b/210012154 Disable continuous coverage if instrumentation is on
-    # for bionic/libc
-    EXPORT_GLOBAL_CLANG_COVERAGE_OPTIONS := export LLVM_PROFILE_FILE /data/misc/trace/clang%c-%20m.profraw
-  else
-    EXPORT_GLOBAL_CLANG_COVERAGE_OPTIONS := export LLVM_PROFILE_FILE /data/misc/trace/clang-%20m.profraw
-  endif
+  EXPORT_GLOBAL_CLANG_COVERAGE_OPTIONS := export LLVM_PROFILE_FILE /data/misc/trace/clang-%20m.profraw
 endif
 
 # Put it here instead of in init.rc module definition,
diff --git a/trusty/storage/proxy/Android.bp b/trusty/storage/proxy/Android.bp
index 38d8685..94f26d8 100644
--- a/trusty/storage/proxy/Android.bp
+++ b/trusty/storage/proxy/Android.bp
@@ -35,7 +35,10 @@
         "liblog",
         "libhardware_legacy",
     ],
-    header_libs: ["libcutils_headers"],
+    header_libs: [
+        "libcutils_headers",
+        "libgsi_headers",
+    ],
 
     static_libs: [
         "libfstab",
diff --git a/trusty/storage/proxy/checkpoint_handling.cpp b/trusty/storage/proxy/checkpoint_handling.cpp
index 6c2fd36..3305d8d 100644
--- a/trusty/storage/proxy/checkpoint_handling.cpp
+++ b/trusty/storage/proxy/checkpoint_handling.cpp
@@ -18,9 +18,12 @@
 #include "log.h"
 
 #include <fstab/fstab.h>
+#include <unistd.h>
 #include <cstring>
 #include <string>
 
+#include <libgsi/libgsi.h>
+
 namespace {
 
 bool checkpointingDoneForever = false;
@@ -75,3 +78,15 @@
 
     return 0;
 }
+
+/**
+ * is_gsi_running() - Check if a GSI image is running via DSU.
+ *
+ * This function is equivalent to android::gsi::IsGsiRunning(), but this API is
+ * not yet vendor-accessible although the underlying metadata file is.
+ *
+ */
+bool is_gsi_running() {
+    /* TODO(b/210501710): Expose GSI image running state to vendor storageproxyd */
+    return !access(android::gsi::kGsiBootedIndicatorFile, F_OK);
+}
diff --git a/trusty/storage/proxy/checkpoint_handling.h b/trusty/storage/proxy/checkpoint_handling.h
index f1bf27c..dfe2947 100644
--- a/trusty/storage/proxy/checkpoint_handling.h
+++ b/trusty/storage/proxy/checkpoint_handling.h
@@ -32,6 +32,8 @@
  */
 int is_data_checkpoint_active(bool* active);
 
+bool is_gsi_running();
+
 #ifdef __cplusplus
 }
 #endif
diff --git a/trusty/storage/proxy/proxy.c b/trusty/storage/proxy/proxy.c
index c690a28..2620034 100644
--- a/trusty/storage/proxy/proxy.c
+++ b/trusty/storage/proxy/proxy.c
@@ -104,8 +104,11 @@
         return -1;
     }
 
-    /* no-execute for user, no access for group and other */
-    umask(S_IXUSR | S_IRWXG | S_IRWXO);
+    /*
+     * No access for group and other. We need execute access for user to create
+     * an accessible directory.
+     */
+    umask(S_IRWXG | S_IRWXO);
 
     return 0;
 }
diff --git a/trusty/storage/proxy/storage.c b/trusty/storage/proxy/storage.c
index 2fde30f..d74a708 100644
--- a/trusty/storage/proxy/storage.c
+++ b/trusty/storage/proxy/storage.c
@@ -16,6 +16,7 @@
 #include <errno.h>
 #include <fcntl.h>
 #include <inttypes.h>
+#include <libgen.h>
 #include <stdbool.h>
 #include <stdlib.h>
 #include <string.h>
@@ -24,13 +25,16 @@
 #include <sys/types.h>
 #include <unistd.h>
 
-#include "log.h"
+#include "checkpoint_handling.h"
 #include "ipc.h"
+#include "log.h"
 #include "storage.h"
 
 #define FD_TBL_SIZE 64
 #define MAX_READ_SIZE 4096
 
+#define ALTERNATE_DATA_DIR "alternate/"
+
 enum sync_state {
     SS_UNUSED = -1,
     SS_CLEAN =  0,
@@ -44,6 +48,8 @@
 static enum sync_state dir_state;
 static enum sync_state fd_state[FD_TBL_SIZE];
 
+static bool alternate_mode;
+
 static struct {
    struct storage_file_read_resp hdr;
    uint8_t data[MAX_READ_SIZE];
@@ -216,6 +222,7 @@
                       const void *r, size_t req_len)
 {
     char *path = NULL;
+    char* parent_path;
     const struct storage_file_open_req *req = r;
     struct storage_file_open_resp resp = {0};
 
@@ -234,6 +241,24 @@
         goto err_response;
     }
 
+    /*
+     * TODO(b/210501710): Expose GSI image running state to vendor
+     * storageproxyd. We want to control data file paths in vendor_init, but we
+     * don't have access to the necessary property there yet. When we have
+     * access to that property we can set the root data path read-only and only
+     * allow creation of files in alternate/. Checking paths here temporarily
+     * until that is fixed.
+     *
+     * We are just checking for "/" instead of "alternate/" because we still
+     * want to still allow access to "persist/" in alternate mode (for now, this
+     * may change in the future).
+     */
+    if (alternate_mode && !strchr(req->name, '/')) {
+        ALOGE("%s: Cannot open root data file \"%s\" in alternate mode\n", __func__, req->name);
+        msg->result = STORAGE_ERR_ACCESS;
+        goto err_response;
+    }
+
     int rc = asprintf(&path, "%s/%s", ssdir_name, req->name);
     if (rc < 0) {
         ALOGE("%s: asprintf failed\n", __func__);
@@ -246,7 +271,23 @@
     if (req->flags & STORAGE_FILE_OPEN_TRUNCATE)
         open_flags |= O_TRUNC;
 
+    parent_path = dirname(path);
     if (req->flags & STORAGE_FILE_OPEN_CREATE) {
+        /*
+         * Create the alternate parent dir if needed & allowed.
+         *
+         * TODO(b/210501710): Expose GSI image running state to vendor
+         * storageproxyd. This directory should be created by vendor_init, once
+         * it has access to the necessary bit of information.
+         */
+        if (strstr(req->name, ALTERNATE_DATA_DIR) == req->name) {
+            rc = mkdir(parent_path, S_IRWXU);
+            if (rc && errno != EEXIST) {
+                ALOGE("%s: Could not create parent directory \"%s\": %s\n", __func__, parent_path,
+                      strerror(errno));
+            }
+        }
+
         /* open or create */
         if (req->flags & STORAGE_FILE_OPEN_CREATE_EXCLUSIVE) {
             /* create exclusive */
@@ -467,6 +508,9 @@
 
 int storage_init(const char *dirname)
 {
+    /* If there is an active DSU image, use the alternate fs mode. */
+    alternate_mode = is_gsi_running();
+
     fs_state = SS_CLEAN;
     dir_state = SS_CLEAN;
     for (uint i = 0; i < FD_TBL_SIZE; i++) {