41 files changed, 1126 insertions, 627 deletions

diff --git a/src/audio_core/codec.cpp b/src/audio_core/codec.cpp
index 7a3bd7eb3..6fba9fdae 100644
--- a/src/audio_core/codec.cpp
+++ b/src/audio_core/codec.cpp
@@ -117,7 +117,9 @@ StereoBuffer16 DecodePCM16(const unsigned num_channels, const u8* const data,
             ret[i].fill(sample);
         }
     } else {
-        std::memcpy(ret.data(), data, sample_count * 2 * sizeof(u16));
+        for (size_t i = 0; i < sample_count; ++i) {
+            std::memcpy(&ret[i], data + i * sizeof(s16) * 2, 2 * sizeof(s16));
+        }
     }
 
     return ret;
diff --git a/src/audio_core/codec.h b/src/audio_core/codec.h
index 2b0c395e6..877b2202d 100644
--- a/src/audio_core/codec.h
+++ b/src/audio_core/codec.h
@@ -5,13 +5,13 @@
 #pragma once
 
 #include <array>
-#include <vector>
+#include <deque>
 #include "common/common_types.h"
 
 namespace Codec {
 
 /// A variable length buffer of signed PCM16 stereo samples.
-using StereoBuffer16 = std::vector<std::array<s16, 2>>;
+using StereoBuffer16 = std::deque<std::array<s16, 2>>;
 
 /// See: Codec::DecodeADPCM
 struct ADPCMState {
diff --git a/src/audio_core/hle/source.h b/src/audio_core/hle/source.h
index ccb7f064f..c4d2debc2 100644
--- a/src/audio_core/hle/source.h
+++ b/src/audio_core/hle/source.h
@@ -108,7 +108,7 @@ private:
 
         u32 current_sample_number = 0;
         u32 next_sample_number = 0;
-        std::vector<std::array<s16, 2>> current_buffer;
+        AudioInterp::StereoBuffer16 current_buffer;
 
         // buffer_id state
 
diff --git a/src/audio_core/interpolate.cpp b/src/audio_core/interpolate.cpp
index 16e68bc5c..83573d772 100644
--- a/src/audio_core/interpolate.cpp
+++ b/src/audio_core/interpolate.cpp
@@ -47,7 +47,7 @@ static void StepOverSamples(State& state, StereoBuffer16& input, float rate,
     state.xn1 = input[inputi + 1];
     state.fposition = fposition - inputi * scale_factor;
 
-    input.erase(input.begin(), input.begin() + inputi + 2);
+    input.erase(input.begin(), std::next(input.begin(), inputi + 2));
 }
 
 void None(State& state, StereoBuffer16& input, float rate, DSP::HLE::StereoFrame16& output,
diff --git a/src/audio_core/interpolate.h b/src/audio_core/interpolate.h
index 59f59bc14..8dff6111a 100644
--- a/src/audio_core/interpolate.h
+++ b/src/audio_core/interpolate.h
@@ -5,14 +5,14 @@
 #pragma once
 
 #include <array>
-#include <vector>
+#include <deque>
 #include "audio_core/hle/common.h"
 #include "common/common_types.h"
 
 namespace AudioInterp {
 
 /// A variable length buffer of signed PCM16 stereo samples.
-using StereoBuffer16 = std::vector<std::array<s16, 2>>;
+using StereoBuffer16 = std::deque<std::array<s16, 2>>;
 
 struct State {
     /// Two historical samples.
diff --git a/src/citra_qt/configuration/configure_graphics.ui b/src/citra_qt/configuration/configure_graphics.ui
index b340149d5..5667b14b6 100644
--- a/src/citra_qt/configuration/configure_graphics.ui
+++ b/src/citra_qt/configuration/configure_graphics.ui
@@ -63,57 +63,57 @@
               <widget class="QComboBox" name="resolution_factor_combobox">
                 <item>
                   <property name="text">
-                    <string notr="true">Auto (Window Size)</string>
+                    <string>Auto (Window Size)</string>
                   </property>
                 </item>
                 <item>
                   <property name="text">
-                    <string notr="true">Native (400x240)</string>
+                    <string>Native (400x240)</string>
                   </property>
                 </item>
                 <item>
                   <property name="text">
-                    <string notr="true">2x Native (800x480)</string>
+                    <string>2x Native (800x480)</string>
                   </property>
                 </item>
                 <item>
                   <property name="text">
-                    <string notr="true">3x Native (1200x720)</string>
+                    <string>3x Native (1200x720)</string>
                   </property>
                 </item>
                 <item>
                   <property name="text">
-                    <string notr="true">4x Native (1600x960)</string>
+                    <string>4x Native (1600x960)</string>
                   </property>
                 </item>
                 <item>
                   <property name="text">
-                    <string notr="true">5x Native (2000x1200)</string>
+                    <string>5x Native (2000x1200)</string>
                   </property>
                 </item>
                 <item>
                   <property name="text">
-                    <string notr="true">6x Native (2400x1440)</string>
+                    <string>6x Native (2400x1440)</string>
                   </property>
                 </item>
                 <item>
                   <property name="text">
-                    <string notr="true">7x Native (2800x1680)</string>
+                    <string>7x Native (2800x1680)</string>
                   </property>
                 </item>
                 <item>
                   <property name="text">
-                    <string notr="true">8x Native (3200x1920)</string>
+                    <string>8x Native (3200x1920)</string>
                   </property>
                 </item>
                 <item>
                   <property name="text">
-                    <string notr="true">9x Native (3600x2160)</string>
+                    <string>9x Native (3600x2160)</string>
                   </property>
                 </item>
                 <item>
                   <property name="text">
-                    <string notr="true">10x Native (4000x2400)</string>
+                    <string>10x Native (4000x2400)</string>
                   </property>
                 </item>
               </widget>
diff --git a/src/citra_qt/configuration/configure_system.cpp b/src/citra_qt/configuration/configure_system.cpp
index 9b1e6711d..88a067c12 100644
--- a/src/citra_qt/configuration/configure_system.cpp
+++ b/src/citra_qt/configuration/configure_system.cpp
@@ -78,7 +78,8 @@ void ConfigureSystem::ReadSystemSettings() {
 
     // set the console id
     u64 console_id = Service::CFG::GetConsoleUniqueId();
-    ui->label_console_id->setText("Console ID: 0x" + QString::number(console_id, 16).toUpper());
+    ui->label_console_id->setText(
+        tr("Console ID: 0x%1").arg(QString::number(console_id, 16).toUpper()));
 }
 
 void ConfigureSystem::applyConfiguration() {
diff --git a/src/citra_qt/configuration/configure_web.cpp b/src/citra_qt/configuration/configure_web.cpp
index 38ce19c0f..bf8c21ac7 100644
--- a/src/citra_qt/configuration/configure_web.cpp
+++ b/src/citra_qt/configuration/configure_web.cpp
@@ -24,15 +24,15 @@ ConfigureWeb::~ConfigureWeb() {}
 void ConfigureWeb::setConfiguration() {
     ui->web_credentials_disclaimer->setWordWrap(true);
     ui->telemetry_learn_more->setOpenExternalLinks(true);
-    ui->telemetry_learn_more->setText("<a "
-                                      "href='https://citra-emu.org/entry/"
-                                      "telemetry-and-why-thats-a-good-thing/'>Learn more</a>");
+    ui->telemetry_learn_more->setText(tr("<a "
+                                         "href='https://citra-emu.org/entry/"
+                                         "telemetry-and-why-thats-a-good-thing/'>Learn more</a>"));
 
     ui->web_signup_link->setOpenExternalLinks(true);
-    ui->web_signup_link->setText("<a href='https://services.citra-emu.org/'>Sign up</a>");
+    ui->web_signup_link->setText(tr("<a href='https://services.citra-emu.org/'>Sign up</a>"));
     ui->web_token_info_link->setOpenExternalLinks(true);
     ui->web_token_info_link->setText(
-        "<a href='https://citra-emu.org/wiki/citra-web-service/'>What is my token?</a>");
+        tr("<a href='https://citra-emu.org/wiki/citra-web-service/'>What is my token?</a>"));
 
     ui->toggle_telemetry->setChecked(Settings::values.enable_telemetry);
     ui->edit_username->setText(QString::fromStdString(Settings::values.citra_username));
@@ -40,8 +40,8 @@ void ConfigureWeb::setConfiguration() {
     // Connect after setting the values, to avoid calling OnLoginChanged now
     connect(ui->edit_token, &QLineEdit::textChanged, this, &ConfigureWeb::OnLoginChanged);
     connect(ui->edit_username, &QLineEdit::textChanged, this, &ConfigureWeb::OnLoginChanged);
-    ui->label_telemetry_id->setText("Telemetry ID: 0x" +
-                                    QString::number(Core::GetTelemetryId(), 16).toUpper());
+    ui->label_telemetry_id->setText(
+        tr("Telemetry ID: 0x%1").arg(QString::number(Core::GetTelemetryId(), 16).toUpper()));
     user_verified = true;
 }
 
@@ -60,8 +60,8 @@ void ConfigureWeb::applyConfiguration() {
 
 void ConfigureWeb::RefreshTelemetryID() {
     const u64 new_telemetry_id{Core::RegenerateTelemetryId()};
-    ui->label_telemetry_id->setText("Telemetry ID: 0x" +
-                                    QString::number(new_telemetry_id, 16).toUpper());
+    ui->label_telemetry_id->setText(
+        tr("Telemetry ID: 0x%1").arg(QString::number(new_telemetry_id, 16).toUpper()));
 }
 
 void ConfigureWeb::OnLoginChanged() {
diff --git a/src/core/CMakeLists.txt b/src/core/CMakeLists.txt
index cd1a8de2d..3ed619991 100644
--- a/src/core/CMakeLists.txt
+++ b/src/core/CMakeLists.txt
@@ -26,6 +26,7 @@ set(SRCS
             file_sys/archive_systemsavedata.cpp
             file_sys/disk_archive.cpp
             file_sys/ivfc_archive.cpp
+            file_sys/ncch_container.cpp
             file_sys/path_parser.cpp
             file_sys/savedata_archive.cpp
             frontend/camera/blank_camera.cpp
diff --git a/src/core/arm/arm_interface.h b/src/core/arm/arm_interface.h
index ccd43f431..2aa017a54 100644
--- a/src/core/arm/arm_interface.h
+++ b/src/core/arm/arm_interface.h
@@ -41,6 +41,9 @@ public:
     /// Clear all instruction cache
     virtual void ClearInstructionCache() = 0;
 
+    /// Notify CPU emulation that page tables have changed
+    virtual void PageTableChanged() = 0;
+
     /**
      * Set the Program Counter to an address
      * @param addr Address to set PC to
diff --git a/src/core/arm/dynarmic/arm_dynarmic.cpp b/src/core/arm/dynarmic/arm_dynarmic.cpp
index 34c5aa381..42ae93ae8 100644
--- a/src/core/arm/dynarmic/arm_dynarmic.cpp
+++ b/src/core/arm/dynarmic/arm_dynarmic.cpp
@@ -41,7 +41,7 @@ static bool IsReadOnlyMemory(u32 vaddr) {
 }
 
 static Dynarmic::UserCallbacks GetUserCallbacks(
-    const std::shared_ptr<ARMul_State>& interpeter_state) {
+    const std::shared_ptr<ARMul_State>& interpeter_state, Memory::PageTable* current_page_table) {
     Dynarmic::UserCallbacks user_callbacks{};
     user_callbacks.InterpreterFallback = &InterpreterFallback;
     user_callbacks.user_arg = static_cast<void*>(interpeter_state.get());
@@ -56,16 +56,14 @@ static Dynarmic::UserCallbacks GetUserCallbacks(
     user_callbacks.memory.Write16 = &Memory::Write16;
     user_callbacks.memory.Write32 = &Memory::Write32;
     user_callbacks.memory.Write64 = &Memory::Write64;
-    // TODO(Subv): Re-add the page table pointers once dynarmic supports switching page tables at
-    // runtime.
-    user_callbacks.page_table = nullptr;
+    user_callbacks.page_table = &current_page_table->pointers;
     user_callbacks.coprocessors[15] = std::make_shared<DynarmicCP15>(interpeter_state);
     return user_callbacks;
 }
 
 ARM_Dynarmic::ARM_Dynarmic(PrivilegeMode initial_mode) {
     interpreter_state = std::make_shared<ARMul_State>(initial_mode);
-    jit = std::make_unique<Dynarmic::Jit>(GetUserCallbacks(interpreter_state));
+    PageTableChanged();
 }
 
 void ARM_Dynarmic::SetPC(u32 pc) {
@@ -136,6 +134,7 @@ void ARM_Dynarmic::AddTicks(u64 ticks) {
 MICROPROFILE_DEFINE(ARM_Jit, "ARM JIT", "ARM JIT", MP_RGB(255, 64, 64));
 
 void ARM_Dynarmic::ExecuteInstructions(int num_instructions) {
+    ASSERT(Memory::GetCurrentPageTable() == current_page_table);
     MICROPROFILE_SCOPE(ARM_Jit);
 
     std::size_t ticks_executed = jit->Run(static_cast<unsigned>(num_instructions));
@@ -178,3 +177,16 @@ void ARM_Dynarmic::PrepareReschedule() {
 void ARM_Dynarmic::ClearInstructionCache() {
     jit->ClearCache();
 }
+
+void ARM_Dynarmic::PageTableChanged() {
+    current_page_table = Memory::GetCurrentPageTable();
+
+    auto iter = jits.find(current_page_table);
+    if (iter != jits.end()) {
+        jit = iter->second.get();
+        return;
+    }
+
+    jit = new Dynarmic::Jit(GetUserCallbacks(interpreter_state, current_page_table));
+    jits.emplace(current_page_table, std::unique_ptr<Dynarmic::Jit>(jit));
+}
diff --git a/src/core/arm/dynarmic/arm_dynarmic.h b/src/core/arm/dynarmic/arm_dynarmic.h
index 834dc989e..96148a1a5 100644
--- a/src/core/arm/dynarmic/arm_dynarmic.h
+++ b/src/core/arm/dynarmic/arm_dynarmic.h
@@ -4,12 +4,17 @@
 
 #pragma once
 
+#include <map>
 #include <memory>
 #include <dynarmic/dynarmic.h>
 #include "common/common_types.h"
 #include "core/arm/arm_interface.h"
 #include "core/arm/skyeye_common/armstate.h"
 
+namespace Memory {
+struct PageTable;
+} // namespace Memory
+
 class ARM_Dynarmic final : public ARM_Interface {
 public:
     ARM_Dynarmic(PrivilegeMode initial_mode);
@@ -36,8 +41,11 @@ public:
     void ExecuteInstructions(int num_instructions) override;
 
     void ClearInstructionCache() override;
+    void PageTableChanged() override;
 
 private:
-    std::unique_ptr<Dynarmic::Jit> jit;
+    Dynarmic::Jit* jit = nullptr;
+    Memory::PageTable* current_page_table = nullptr;
+    std::map<Memory::PageTable*, std::unique_ptr<Dynarmic::Jit>> jits;
     std::shared_ptr<ARMul_State> interpreter_state;
 };
diff --git a/src/core/arm/dyncom/arm_dyncom.cpp b/src/core/arm/dyncom/arm_dyncom.cpp
index 81f9bf99e..da955c9b9 100644
--- a/src/core/arm/dyncom/arm_dyncom.cpp
+++ b/src/core/arm/dyncom/arm_dyncom.cpp
@@ -25,6 +25,10 @@ void ARM_DynCom::ClearInstructionCache() {
     trans_cache_buf_top = 0;
 }
 
+void ARM_DynCom::PageTableChanged() {
+    ClearInstructionCache();
+}
+
 void ARM_DynCom::SetPC(u32 pc) {
     state->Reg[15] = pc;
 }
diff --git a/src/core/arm/dyncom/arm_dyncom.h b/src/core/arm/dyncom/arm_dyncom.h
index 62c174f3c..0ae535671 100644
--- a/src/core/arm/dyncom/arm_dyncom.h
+++ b/src/core/arm/dyncom/arm_dyncom.h
@@ -16,6 +16,7 @@ public:
     ~ARM_DynCom();
 
     void ClearInstructionCache() override;
+    void PageTableChanged() override;
 
     void SetPC(u32 pc) override;
     u32 GetPC() const override;
diff --git a/src/core/file_sys/archive_selfncch.cpp b/src/core/file_sys/archive_selfncch.cpp
index 298a37a44..a16941c70 100644
--- a/src/core/file_sys/archive_selfncch.cpp
+++ b/src/core/file_sys/archive_selfncch.cpp
@@ -3,12 +3,14 @@
 // Refer to the license.txt file included.
 
 #include <array>
+#include <cinttypes>
 #include "common/common_types.h"
 #include "common/logging/log.h"
 #include "common/swap.h"
 #include "core/file_sys/archive_selfncch.h"
 #include "core/file_sys/errors.h"
 #include "core/file_sys/ivfc_archive.h"
+#include "core/hle/kernel/process.h"
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
 // FileSys namespace
@@ -102,8 +104,7 @@ public:
 
         switch (static_cast<SelfNCCHFilePathType>(file_path.type)) {
         case SelfNCCHFilePathType::UpdateRomFS:
-            LOG_WARNING(Service_FS, "(STUBBED) open update RomFS");
-            return OpenRomFS();
+            return OpenUpdateRomFS();
 
         case SelfNCCHFilePathType::RomFS:
             return OpenRomFS();
@@ -179,6 +180,17 @@ private:
         }
     }
 
+    ResultVal<std::unique_ptr<FileBackend>> OpenUpdateRomFS() const {
+        if (ncch_data.update_romfs_file) {
+            return MakeResult<std::unique_ptr<FileBackend>>(std::make_unique<IVFCFile>(
+                ncch_data.update_romfs_file, ncch_data.update_romfs_offset,
+                ncch_data.update_romfs_size));
+        } else {
+            LOG_INFO(Service_FS, "Unable to read update RomFS");
+            return ERROR_ROMFS_NOT_FOUND;
+        }
+    }
+
     ResultVal<std::unique_ptr<FileBackend>> OpenExeFS(const std::string& filename) const {
         if (filename == "icon") {
             if (ncch_data.icon) {
@@ -217,30 +229,57 @@ private:
     NCCHData ncch_data;
 };
 
-ArchiveFactory_SelfNCCH::ArchiveFactory_SelfNCCH(Loader::AppLoader& app_loader) {
+void ArchiveFactory_SelfNCCH::Register(Loader::AppLoader& app_loader) {
+    u64 program_id = 0;
+    if (app_loader.ReadProgramId(program_id) != Loader::ResultStatus::Success) {
+        LOG_WARNING(
+            Service_FS,
+            "Could not read program id when registering with SelfNCCH, this might be a 3dsx file");
+    }
+
+    LOG_DEBUG(Service_FS, "Registering program %016" PRIX64 " with the SelfNCCH archive factory",
+              program_id);
+
+    if (ncch_data.find(program_id) != ncch_data.end()) {
+        LOG_WARNING(Service_FS, "Registering program %016" PRIX64
+                                " with SelfNCCH will override existing mapping",
+                    program_id);
+    }
+
+    NCCHData& data = ncch_data[program_id];
+
     std::shared_ptr<FileUtil::IOFile> romfs_file_;
     if (Loader::ResultStatus::Success ==
-        app_loader.ReadRomFS(romfs_file_, ncch_data.romfs_offset, ncch_data.romfs_size)) {
+        app_loader.ReadRomFS(romfs_file_, data.romfs_offset, data.romfs_size)) {
+
+        data.romfs_file = std::move(romfs_file_);
+    }
+
+    std::shared_ptr<FileUtil::IOFile> update_romfs_file;
+    if (Loader::ResultStatus::Success ==
+        app_loader.ReadUpdateRomFS(update_romfs_file, data.update_romfs_offset,
+                                   data.update_romfs_size)) {
 
-        ncch_data.romfs_file = std::move(romfs_file_);
+        data.update_romfs_file = std::move(update_romfs_file);
     }
 
     std::vector<u8> buffer;
 
     if (Loader::ResultStatus::Success == app_loader.ReadIcon(buffer))
-        ncch_data.icon = std::make_shared<std::vector<u8>>(std::move(buffer));
+        data.icon = std::make_shared<std::vector<u8>>(std::move(buffer));
 
     buffer.clear();
     if (Loader::ResultStatus::Success == app_loader.ReadLogo(buffer))
-        ncch_data.logo = std::make_shared<std::vector<u8>>(std::move(buffer));
+        data.logo = std::make_shared<std::vector<u8>>(std::move(buffer));
 
     buffer.clear();
     if (Loader::ResultStatus::Success == app_loader.ReadBanner(buffer))
-        ncch_data.banner = std::make_shared<std::vector<u8>>(std::move(buffer));
+        data.banner = std::make_shared<std::vector<u8>>(std::move(buffer));
 }
 
 ResultVal<std::unique_ptr<ArchiveBackend>> ArchiveFactory_SelfNCCH::Open(const Path& path) {
-    auto archive = std::make_unique<SelfNCCHArchive>(ncch_data);
+    auto archive = std::make_unique<SelfNCCHArchive>(
+        ncch_data[Kernel::g_current_process->codeset->program_id]);
     return MakeResult<std::unique_ptr<ArchiveBackend>>(std::move(archive));
 }
 
diff --git a/src/core/file_sys/archive_selfncch.h b/src/core/file_sys/archive_selfncch.h
index f1b971296..0d6d6766e 100644
--- a/src/core/file_sys/archive_selfncch.h
+++ b/src/core/file_sys/archive_selfncch.h
@@ -6,6 +6,7 @@
 
 #include <memory>
 #include <string>
+#include <unordered_map>
 #include <vector>
 #include "common/common_types.h"
 #include "core/file_sys/archive_backend.h"
@@ -24,12 +25,19 @@ struct NCCHData {
     std::shared_ptr<FileUtil::IOFile> romfs_file;
     u64 romfs_offset = 0;
     u64 romfs_size = 0;
+
+    std::shared_ptr<FileUtil::IOFile> update_romfs_file;
+    u64 update_romfs_offset = 0;
+    u64 update_romfs_size = 0;
 };
 
 /// File system interface to the SelfNCCH archive
 class ArchiveFactory_SelfNCCH final : public ArchiveFactory {
 public:
-    explicit ArchiveFactory_SelfNCCH(Loader::AppLoader& app_loader);
+    ArchiveFactory_SelfNCCH() = default;
+
+    /// Registers a loaded application so that we can open its SelfNCCH archive when requested.
+    void Register(Loader::AppLoader& app_loader);
 
     std::string GetName() const override {
         return "SelfNCCH";
@@ -39,7 +47,8 @@ public:
     ResultVal<ArchiveFormatInfo> GetFormatInfo(const Path& path) const override;
 
 private:
-    NCCHData ncch_data;
+    /// Mapping of ProgramId -> NCCHData
+    std::unordered_map<u64, NCCHData> ncch_data;
 };
 
 } // namespace FileSys
diff --git a/src/core/file_sys/ncch_container.cpp b/src/core/file_sys/ncch_container.cpp
new file mode 100644
index 000000000..59c72f3e9
--- /dev/null
+++ b/src/core/file_sys/ncch_container.cpp
@@ -0,0 +1,316 @@
+// Copyright 2017 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#include <cinttypes>
+#include <cstring>
+#include <memory>
+#include "common/common_types.h"
+#include "common/logging/log.h"
+#include "core/core.h"
+#include "core/file_sys/ncch_container.h"
+#include "core/loader/loader.h"
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// FileSys namespace
+
+namespace FileSys {
+
+static const int kMaxSections = 8;   ///< Maximum number of sections (files) in an ExeFs
+static const int kBlockSize = 0x200; ///< Size of ExeFS blocks (in bytes)
+
+/**
+ * Get the decompressed size of an LZSS compressed ExeFS file
+ * @param buffer Buffer of compressed file
+ * @param size Size of compressed buffer
+ * @return Size of decompressed buffer
+ */
+static u32 LZSS_GetDecompressedSize(const u8* buffer, u32 size) {
+    u32 offset_size = *(u32*)(buffer + size - 4);
+    return offset_size + size;
+}
+
+/**
+ * Decompress ExeFS file (compressed with LZSS)
+ * @param compressed Compressed buffer
+ * @param compressed_size Size of compressed buffer
+ * @param decompressed Decompressed buffer
+ * @param decompressed_size Size of decompressed buffer
+ * @return True on success, otherwise false
+ */
+static bool LZSS_Decompress(const u8* compressed, u32 compressed_size, u8* decompressed,
+                            u32 decompressed_size) {
+    const u8* footer = compressed + compressed_size - 8;
+    u32 buffer_top_and_bottom = *reinterpret_cast<const u32*>(footer);
+    u32 out = decompressed_size;
+    u32 index = compressed_size - ((buffer_top_and_bottom >> 24) & 0xFF);
+    u32 stop_index = compressed_size - (buffer_top_and_bottom & 0xFFFFFF);
+
+    memset(decompressed, 0, decompressed_size);
+    memcpy(decompressed, compressed, compressed_size);
+
+    while (index > stop_index) {
+        u8 control = compressed[--index];
+
+        for (unsigned i = 0; i < 8; i++) {
+            if (index <= stop_index)
+                break;
+            if (index <= 0)
+                break;
+            if (out <= 0)
+                break;
+
+            if (control & 0x80) {
+                // Check if compression is out of bounds
+                if (index < 2)
+                    return false;
+                index -= 2;
+
+                u32 segment_offset = compressed[index] | (compressed[index + 1] << 8);
+                u32 segment_size = ((segment_offset >> 12) & 15) + 3;
+                segment_offset &= 0x0FFF;
+                segment_offset += 2;
+
+                // Check if compression is out of bounds
+                if (out < segment_size)
+                    return false;
+
+                for (unsigned j = 0; j < segment_size; j++) {
+                    // Check if compression is out of bounds
+                    if (out + segment_offset >= decompressed_size)
+                        return false;
+
+                    u8 data = decompressed[out + segment_offset];
+                    decompressed[--out] = data;
+                }
+            } else {
+                // Check if compression is out of bounds
+                if (out < 1)
+                    return false;
+                decompressed[--out] = compressed[--index];
+            }
+            control <<= 1;
+        }
+    }
+    return true;
+}
+
+NCCHContainer::NCCHContainer(const std::string& filepath) : filepath(filepath) {
+    file = FileUtil::IOFile(filepath, "rb");
+}
+
+Loader::ResultStatus NCCHContainer::OpenFile(const std::string& filepath) {
+    this->filepath = filepath;
+    file = FileUtil::IOFile(filepath, "rb");
+
+    if (!file.IsOpen()) {
+        LOG_WARNING(Service_FS, "Failed to open %s", filepath.c_str());
+        return Loader::ResultStatus::Error;
+    }
+
+    LOG_DEBUG(Service_FS, "Opened %s", filepath.c_str());
+    return Loader::ResultStatus::Success;
+}
+
+Loader::ResultStatus NCCHContainer::Load() {
+    if (is_loaded)
+        return Loader::ResultStatus::Success;
+
+    // Reset read pointer in case this file has been read before.
+    file.Seek(0, SEEK_SET);
+
+    if (file.ReadBytes(&ncch_header, sizeof(NCCH_Header)) != sizeof(NCCH_Header))
+        return Loader::ResultStatus::Error;
+
+    // Skip NCSD header and load first NCCH (NCSD is just a container of NCCH files)...
+    if (Loader::MakeMagic('N', 'C', 'S', 'D') == ncch_header.magic) {
+        LOG_DEBUG(Service_FS, "Only loading the first (bootable) NCCH within the NCSD file!");
+        ncch_offset = 0x4000;
+        file.Seek(ncch_offset, SEEK_SET);
+        file.ReadBytes(&ncch_header, sizeof(NCCH_Header));
+    }
+
+    // Verify we are loading the correct file type...
+    if (Loader::MakeMagic('N', 'C', 'C', 'H') != ncch_header.magic)
+        return Loader::ResultStatus::ErrorInvalidFormat;
+
+    // System archives and DLC don't have an extended header but have RomFS
+    if (ncch_header.extended_header_size) {
+        if (file.ReadBytes(&exheader_header, sizeof(ExHeader_Header)) != sizeof(ExHeader_Header))
+            return Loader::ResultStatus::Error;
+
+        is_compressed = (exheader_header.codeset_info.flags.flag & 1) == 1;
+        u32 entry_point = exheader_header.codeset_info.text.address;
+        u32 code_size = exheader_header.codeset_info.text.code_size;
+        u32 stack_size = exheader_header.codeset_info.stack_size;
+        u32 bss_size = exheader_header.codeset_info.bss_size;
+        u32 core_version = exheader_header.arm11_system_local_caps.core_version;
+        u8 priority = exheader_header.arm11_system_local_caps.priority;
+        u8 resource_limit_category =
+            exheader_header.arm11_system_local_caps.resource_limit_category;
+
+        LOG_DEBUG(Service_FS, "Name:                        %s", exheader_header.codeset_info.name);
+        LOG_DEBUG(Service_FS, "Program ID:                  %016" PRIX64, ncch_header.program_id);
+        LOG_DEBUG(Service_FS, "Code compressed:             %s", is_compressed ? "yes" : "no");
+        LOG_DEBUG(Service_FS, "Entry point:                 0x%08X", entry_point);
+        LOG_DEBUG(Service_FS, "Code size:                   0x%08X", code_size);
+        LOG_DEBUG(Service_FS, "Stack size:                  0x%08X", stack_size);
+        LOG_DEBUG(Service_FS, "Bss size:                    0x%08X", bss_size);
+        LOG_DEBUG(Service_FS, "Core version:                %d", core_version);
+        LOG_DEBUG(Service_FS, "Thread priority:             0x%X", priority);
+        LOG_DEBUG(Service_FS, "Resource limit category:     %d", resource_limit_category);
+        LOG_DEBUG(Service_FS, "System Mode:                 %d",
+                  static_cast<int>(exheader_header.arm11_system_local_caps.system_mode));
+
+        if (exheader_header.system_info.jump_id != ncch_header.program_id) {
+            LOG_ERROR(Service_FS, "ExHeader Program ID mismatch: the ROM is probably encrypted.");
+            return Loader::ResultStatus::ErrorEncrypted;
+        }
+
+        has_exheader = true;
+    }
+
+    // DLC can have an ExeFS and a RomFS but no extended header
+    if (ncch_header.exefs_size) {
+        exefs_offset = ncch_header.exefs_offset * kBlockSize;
+        u32 exefs_size = ncch_header.exefs_size * kBlockSize;
+
+        LOG_DEBUG(Service_FS, "ExeFS offset:                0x%08X", exefs_offset);
+        LOG_DEBUG(Service_FS, "ExeFS size:                  0x%08X", exefs_size);
+
+        file.Seek(exefs_offset + ncch_offset, SEEK_SET);
+        if (file.ReadBytes(&exefs_header, sizeof(ExeFs_Header)) != sizeof(ExeFs_Header))
+            return Loader::ResultStatus::Error;
+
+        has_exefs = true;
+    }
+
+    if (ncch_header.romfs_offset != 0 && ncch_header.romfs_size != 0)
+        has_romfs = true;
+
+    is_loaded = true;
+    return Loader::ResultStatus::Success;
+}
+
+Loader::ResultStatus NCCHContainer::LoadSectionExeFS(const char* name, std::vector<u8>& buffer) {
+    if (!file.IsOpen())
+        return Loader::ResultStatus::Error;
+
+    Loader::ResultStatus result = Load();
+    if (result != Loader::ResultStatus::Success)
+        return result;
+
+    if (!has_exefs)
+        return Loader::ResultStatus::ErrorNotUsed;
+
+    LOG_DEBUG(Service_FS, "%d sections:", kMaxSections);
+    // Iterate through the ExeFs archive until we find a section with the specified name...
+    for (unsigned section_number = 0; section_number < kMaxSections; section_number++) {
+        const auto& section = exefs_header.section[section_number];
+
+        // Load the specified section...
+        if (strcmp(section.name, name) == 0) {
+            LOG_DEBUG(Service_FS, "%d - offset: 0x%08X, size: 0x%08X, name: %s", section_number,
+                      section.offset, section.size, section.name);
+
+            s64 section_offset =
+                (section.offset + exefs_offset + sizeof(ExeFs_Header) + ncch_offset);
+            file.Seek(section_offset, SEEK_SET);
+
+            if (strcmp(section.name, ".code") == 0 && is_compressed) {
+                // Section is compressed, read compressed .code section...
+                std::unique_ptr<u8[]> temp_buffer;
+                try {
+                    temp_buffer.reset(new u8[section.size]);
+                } catch (std::bad_alloc&) {
+                    return Loader::ResultStatus::ErrorMemoryAllocationFailed;
+                }
+
+                if (file.ReadBytes(&temp_buffer[0], section.size) != section.size)
+                    return Loader::ResultStatus::Error;
+
+                // Decompress .code section...
+                u32 decompressed_size = LZSS_GetDecompressedSize(&temp_buffer[0], section.size);
+                buffer.resize(decompressed_size);
+                if (!LZSS_Decompress(&temp_buffer[0], section.size, &buffer[0], decompressed_size))
+                    return Loader::ResultStatus::ErrorInvalidFormat;
+            } else {
+                // Section is uncompressed...
+                buffer.resize(section.size);
+                if (file.ReadBytes(&buffer[0], section.size) != section.size)
+                    return Loader::ResultStatus::Error;
+            }
+            return Loader::ResultStatus::Success;
+        }
+    }
+    return Loader::ResultStatus::ErrorNotUsed;
+}
+
+Loader::ResultStatus NCCHContainer::ReadRomFS(std::shared_ptr<FileUtil::IOFile>& romfs_file,
+                                              u64& offset, u64& size) {
+    if (!file.IsOpen())
+        return Loader::ResultStatus::Error;
+
+    Loader::ResultStatus result = Load();
+    if (result != Loader::ResultStatus::Success)
+        return result;
+
+    if (!has_romfs) {
+        LOG_DEBUG(Service_FS, "RomFS requested from NCCH which has no RomFS");
+        return Loader::ResultStatus::ErrorNotUsed;
+    }
+
+    u32 romfs_offset = ncch_offset + (ncch_header.romfs_offset * kBlockSize) + 0x1000;
+    u32 romfs_size = (ncch_header.romfs_size * kBlockSize) - 0x1000;
+
+    LOG_DEBUG(Service_FS, "RomFS offset:           0x%08X", romfs_offset);
+    LOG_DEBUG(Service_FS, "RomFS size:             0x%08X", romfs_size);
+
+    if (file.GetSize() < romfs_offset + romfs_size)
+        return Loader::ResultStatus::Error;
+
+    // We reopen the file, to allow its position to be independent from file's
+    romfs_file = std::make_shared<FileUtil::IOFile>(filepath, "rb");
+    if (!romfs_file->IsOpen())
+        return Loader::ResultStatus::Error;
+
+    offset = romfs_offset;
+    size = romfs_size;
+
+    return Loader::ResultStatus::Success;
+}
+
+Loader::ResultStatus NCCHContainer::ReadProgramId(u64_le& program_id) {
+    Loader::ResultStatus result = Load();
+    if (result != Loader::ResultStatus::Success)
+        return result;
+
+    program_id = ncch_header.program_id;
+    return Loader::ResultStatus::Success;
+}
+
+bool NCCHContainer::HasExeFS() {
+    Loader::ResultStatus result = Load();
+    if (result != Loader::ResultStatus::Success)
+        return false;
+
+    return has_exefs;
+}
+
+bool NCCHContainer::HasRomFS() {
+    Loader::ResultStatus result = Load();
+    if (result != Loader::ResultStatus::Success)
+        return false;
+
+    return has_romfs;
+}
+
+bool NCCHContainer::HasExHeader() {
+    Loader::ResultStatus result = Load();
+    if (result != Loader::ResultStatus::Success)
+        return false;
+
+    return has_exheader;
+}
+
+} // namespace FileSys
diff --git a/src/core/file_sys/ncch_container.h b/src/core/file_sys/ncch_container.h
new file mode 100644
index 000000000..8af9032b4
--- /dev/null
+++ b/src/core/file_sys/ncch_container.h
@@ -0,0 +1,244 @@
+// Copyright 2017 Citra Emulator Project
+// Licensed under GPLv2 or any later version
+// Refer to the license.txt file included.
+
+#pragma once
+
+#include <cstddef>
+#include <memory>
+#include <string>
+#include <vector>
+#include "common/bit_field.h"
+#include "common/common_types.h"
+#include "common/file_util.h"
+#include "common/swap.h"
+#include "core/core.h"
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+/// NCCH header (Note: "NCCH" appears to be a publicly unknown acronym)
+
+struct NCCH_Header {
+    u8 signature[0x100];
+    u32_le magic;
+    u32_le content_size;
+    u8 partition_id[8];
+    u16_le maker_code;
+    u16_le version;
+    u8 reserved_0[4];
+    u64_le program_id;
+    u8 reserved_1[0x10];
+    u8 logo_region_hash[0x20];
+    u8 product_code[0x10];
+    u8 extended_header_hash[0x20];
+    u32_le extended_header_size;
+    u8 reserved_2[4];
+    u8 flags[8];
+    u32_le plain_region_offset;
+    u32_le plain_region_size;
+    u32_le logo_region_offset;
+    u32_le logo_region_size;
+    u32_le exefs_offset;
+    u32_le exefs_size;
+    u32_le exefs_hash_region_size;
+    u8 reserved_3[4];
+    u32_le romfs_offset;
+    u32_le romfs_size;
+    u32_le romfs_hash_region_size;
+    u8 reserved_4[4];
+    u8 exefs_super_block_hash[0x20];
+    u8 romfs_super_block_hash[0x20];
+};
+
+static_assert(sizeof(NCCH_Header) == 0x200, "NCCH header structure size is wrong");
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// ExeFS (executable file system) headers
+
+struct ExeFs_SectionHeader {
+    char name[8];
+    u32 offset;
+    u32 size;
+};
+
+struct ExeFs_Header {
+    ExeFs_SectionHeader section[8];
+    u8 reserved[0x80];
+    u8 hashes[8][0x20];
+};
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// ExHeader (executable file system header) headers
+
+struct ExHeader_SystemInfoFlags {
+    u8 reserved[5];
+    u8 flag;
+    u8 remaster_version[2];
+};
+
+struct ExHeader_CodeSegmentInfo {
+    u32 address;
+    u32 num_max_pages;
+    u32 code_size;
+};
+
+struct ExHeader_CodeSetInfo {
+    u8 name[8];
+    ExHeader_SystemInfoFlags flags;
+    ExHeader_CodeSegmentInfo text;
+    u32 stack_size;
+    ExHeader_CodeSegmentInfo ro;
+    u8 reserved[4];
+    ExHeader_CodeSegmentInfo data;
+    u32 bss_size;
+};
+
+struct ExHeader_DependencyList {
+    u8 program_id[0x30][8];
+};
+
+struct ExHeader_SystemInfo {
+    u64 save_data_size;
+    u64_le jump_id;
+    u8 reserved_2[0x30];
+};
+
+struct ExHeader_StorageInfo {
+    u8 ext_save_data_id[8];
+    u8 system_save_data_id[8];
+    u8 reserved[8];
+    u8 access_info[7];
+    u8 other_attributes;
+};
+
+struct ExHeader_ARM11_SystemLocalCaps {
+    u64_le program_id;
+    u32_le core_version;
+    u8 reserved_flags[2];
+    union {
+        u8 flags0;
+        BitField<0, 2, u8> ideal_processor;
+        BitField<2, 2, u8> affinity_mask;
+        BitField<4, 4, u8> system_mode;
+    };
+    u8 priority;
+    u8 resource_limit_descriptor[0x10][2];
+    ExHeader_StorageInfo storage_info;
+    u8 service_access_control[0x20][8];
+    u8 ex_service_access_control[0x2][8];
+    u8 reserved[0xf];
+    u8 resource_limit_category;
+};
+
+struct ExHeader_ARM11_KernelCaps {
+    u32_le descriptors[28];
+    u8 reserved[0x10];
+};
+
+struct ExHeader_ARM9_AccessControl {
+    u8 descriptors[15];
+    u8 descversion;
+};
+
+struct ExHeader_Header {
+    ExHeader_CodeSetInfo codeset_info;
+    ExHeader_DependencyList dependency_list;
+    ExHeader_SystemInfo system_info;
+    ExHeader_ARM11_SystemLocalCaps arm11_system_local_caps;
+    ExHeader_ARM11_KernelCaps arm11_kernel_caps;
+    ExHeader_ARM9_AccessControl arm9_access_control;
+    struct {
+        u8 signature[0x100];
+        u8 ncch_public_key_modulus[0x100];
+        ExHeader_ARM11_SystemLocalCaps arm11_system_local_caps;
+        ExHeader_ARM11_KernelCaps arm11_kernel_caps;
+        ExHeader_ARM9_AccessControl arm9_access_control;
+    } access_desc;
+};
+
+static_assert(sizeof(ExHeader_Header) == 0x800, "ExHeader structure size is wrong");
+
+////////////////////////////////////////////////////////////////////////////////////////////////////
+// FileSys namespace
+
+namespace FileSys {
+
+/**
+ * Helper which implements an interface to deal with NCCH containers which can
+ * contain ExeFS archives or RomFS archives for games or other applications.
+ */
+class NCCHContainer {
+public:
+    NCCHContainer(const std::string& filepath);
+    NCCHContainer() {}
+
+    Loader::ResultStatus OpenFile(const std::string& filepath);
+
+    /**
+     * Ensure ExeFS and exheader is loaded and ready for reading sections
+     * @return ResultStatus result of function
+     */
+    Loader::ResultStatus Load();
+
+    /**
+     * Reads an application ExeFS section of an NCCH file (e.g. .code, .logo, etc.)
+     * @param name Name of section to read out of NCCH file
+     * @param buffer Vector to read data into
+     * @return ResultStatus result of function
+     */
+    Loader::ResultStatus LoadSectionExeFS(const char* name, std::vector<u8>& buffer);
+
+    /**
+     * Get the RomFS of the NCCH container
+     * Since the RomFS can be huge, we return a file reference instead of copying to a buffer
+     * @param romfs_file The file containing the RomFS
+     * @param offset The offset the romfs begins on
+     * @param size The size of the romfs
+     * @return ResultStatus result of function
+     */
+    Loader::ResultStatus ReadRomFS(std::shared_ptr<FileUtil::IOFile>& romfs_file, u64& offset,
+                                   u64& size);
+
+    /**
+     * Get the Program ID of the NCCH container
+     * @return ResultStatus result of function
+     */
+    Loader::ResultStatus ReadProgramId(u64_le& program_id);
+
+    /**
+     * Checks whether the NCCH container contains an ExeFS
+     * @return bool check result
+     */
+    bool HasExeFS();
+
+    /**
+     * Checks whether the NCCH container contains a RomFS
+     * @return bool check result
+     */
+    bool HasRomFS();
+
+    /**
+     * Checks whether the NCCH container contains an ExHeader
+     * @return bool check result
+     */
+    bool HasExHeader();
+
+    NCCH_Header ncch_header;
+    ExeFs_Header exefs_header;
+    ExHeader_Header exheader_header;
+
+private:
+    bool has_exheader = false;
+    bool has_exefs = false;
+    bool has_romfs = false;
+
+    bool is_loaded = false;
+    bool is_compressed = false;
+
+    u32 ncch_offset = 0; // Offset to NCCH header, can be 0 or after NCSD header
+    u32 exefs_offset = 0;
+
+    std::string filepath;
+    FileUtil::IOFile file;
+};
+
+} // namespace FileSys
diff --git a/src/core/hle/kernel/thread.cpp b/src/core/hle/kernel/thread.cpp
index 324415a36..61378211f 100644
--- a/src/core/hle/kernel/thread.cpp
+++ b/src/core/hle/kernel/thread.cpp
@@ -178,16 +178,13 @@ static void SwitchContext(Thread* new_thread) {
         ready_queue.remove(new_thread->current_priority, new_thread);
         new_thread->status = THREADSTATUS_RUNNING;
 
-        Core::CPU().LoadContext(new_thread->context);
-        Core::CPU().SetCP15Register(CP15_THREAD_URO, new_thread->GetTLSAddress());
-
         if (previous_process != current_thread->owner_process) {
             Kernel::g_current_process = current_thread->owner_process;
-            Memory::current_page_table = &Kernel::g_current_process->vm_manager.page_table;
-            // We have switched processes and thus, page tables, clear the instruction cache so we
-            // don't keep stale data from the previous process.
-            Core::CPU().ClearInstructionCache();
+            SetCurrentPageTable(&Kernel::g_current_process->vm_manager.page_table);
         }
+
+        Core::CPU().LoadContext(new_thread->context);
+        Core::CPU().SetCP15Register(CP15_THREAD_URO, new_thread->GetTLSAddress());
     } else {
         current_thread = nullptr;
         // Note: We do not reset the current process and current page table when idling because
diff --git a/src/core/hle/service/fs/archive.cpp b/src/core/hle/service/fs/archive.cpp
index 033fbc9aa..4ccb3cd32 100644
--- a/src/core/hle/service/fs/archive.cpp
+++ b/src/core/hle/service/fs/archive.cpp
@@ -20,6 +20,7 @@
 #include "core/file_sys/archive_savedata.h"
 #include "core/file_sys/archive_sdmc.h"
 #include "core/file_sys/archive_sdmcwriteonly.h"
+#include "core/file_sys/archive_selfncch.h"
 #include "core/file_sys/archive_systemsavedata.h"
 #include "core/file_sys/directory_backend.h"
 #include "core/file_sys/errors.h"
@@ -48,7 +49,7 @@ struct hash<Service::FS::ArchiveIdCode> {
         return std::hash<Type>()(static_cast<Type>(id_code));
     }
 };
-}
+} // namespace std
 
 static constexpr Kernel::Handle INVALID_HANDLE{};
 
@@ -564,6 +565,21 @@ void RegisterArchiveTypes() {
     auto systemsavedata_factory =
         std::make_unique<FileSys::ArchiveFactory_SystemSaveData>(nand_directory);
     RegisterArchiveType(std::move(systemsavedata_factory), ArchiveIdCode::SystemSaveData);
+
+    auto selfncch_factory = std::make_unique<FileSys::ArchiveFactory_SelfNCCH>();
+    RegisterArchiveType(std::move(selfncch_factory), ArchiveIdCode::SelfNCCH);
+}
+
+void RegisterSelfNCCH(Loader::AppLoader& app_loader) {
+    auto itr = id_code_map.find(ArchiveIdCode::SelfNCCH);
+    if (itr == id_code_map.end()) {
+        LOG_ERROR(Service_FS,
+                  "Could not register a new NCCH because the SelfNCCH archive hasn't been created");
+        return;
+    }
+
+    auto* factory = static_cast<FileSys::ArchiveFactory_SelfNCCH*>(itr->second.get());
+    factory->Register(app_loader);
 }
 
 void UnregisterArchiveTypes() {
diff --git a/src/core/hle/service/fs/archive.h b/src/core/hle/service/fs/archive.h
index 3a3371c88..e3c8fc2ef 100644
--- a/src/core/hle/service/fs/archive.h
+++ b/src/core/hle/service/fs/archive.h
@@ -21,6 +21,10 @@ static constexpr char SYSTEM_ID[]{"00000000000000000000000000000000"};
 /// The scrambled SD card CID, also known as ID1
 static constexpr char SDCARD_ID[]{"00000000000000000000000000000000"};
 
+namespace Loader {
+class AppLoader;
+}
+
 namespace Service {
 namespace FS {
 
@@ -259,6 +263,9 @@ void ArchiveInit();
 /// Shutdown archives
 void ArchiveShutdown();
 
+/// Registers a new NCCH file with the SelfNCCH archive factory
+void RegisterSelfNCCH(Loader::AppLoader& app_loader);
+
 /// Register all archive types
 void RegisterArchiveTypes();
 
diff --git a/src/core/hle/service/nwm/nwm_uds.cpp b/src/core/hle/service/nwm/nwm_uds.cpp
index 893bbb1e7..4e2af9ae6 100644
--- a/src/core/hle/service/nwm/nwm_uds.cpp
+++ b/src/core/hle/service/nwm/nwm_uds.cpp
@@ -2,8 +2,10 @@
 // Licensed under GPLv2 or any later version
 // Refer to the license.txt file included.
 
+#include <algorithm>
 #include <array>
 #include <cstring>
+#include <list>
 #include <mutex>
 #include <unordered_map>
 #include <vector>
@@ -37,9 +39,12 @@ static ConnectionStatus connection_status{};
 /* Node information about the current network.
  * The amount of elements in this vector is always the maximum number
  * of nodes specified in the network configuration.
- * The first node is always the host, so this always contains at least 1 entry.
+ * The first node is always the host.
  */
-static NodeList node_info(1);
+static NodeList node_info;
+
+// Node information about our own system.
+static NodeInfo current_node;
 
 // Mapping of bind node ids to their respective events.
 static std::unordered_map<u32, Kernel::SharedPtr<Kernel::Event>> bind_node_events;
@@ -54,6 +59,10 @@ static NetworkInfo network_info;
 // Event that will generate and send the 802.11 beacon frames.
 static int beacon_broadcast_event;
 
+// Mutex to synchronize access to the connection status between the emulation thread and the
+// network thread.
+static std::mutex connection_status_mutex;
+
 // Mutex to synchronize access to the list of received beacons between the emulation thread and the
 // network thread.
 static std::mutex beacon_mutex;
@@ -63,14 +72,26 @@ static std::mutex beacon_mutex;
 constexpr size_t MaxBeaconFrames = 15;
 
 // List of the last <MaxBeaconFrames> beacons received from the network.
-static std::deque<Network::WifiPacket> received_beacons;
+static std::list<Network::WifiPacket> received_beacons;
 
 /**
  * Returns a list of received 802.11 beacon frames from the specified sender since the last call.
  */
-std::deque<Network::WifiPacket> GetReceivedBeacons(const MacAddress& sender) {
+std::list<Network::WifiPacket> GetReceivedBeacons(const MacAddress& sender) {
     std::lock_guard<std::mutex> lock(beacon_mutex);
-    // TODO(Subv): Filter by sender.
+    if (sender != Network::BroadcastMac) {
+        std::list<Network::WifiPacket> filtered_list;
+        const auto beacon = std::find_if(received_beacons.begin(), received_beacons.end(),
+                                         [&sender](const Network::WifiPacket& packet) {
+                                             return packet.transmitter_address == sender;
+                                         });
+        if (beacon != received_beacons.end()) {
+            filtered_list.push_back(*beacon);
+            // TODO(B3N30): Check if the complete deque is cleared or just the fetched entries
+            received_beacons.erase(beacon);
+        }
+        return filtered_list;
+    }
     return std::move(received_beacons);
 }
 
@@ -83,6 +104,15 @@ void SendPacket(Network::WifiPacket& packet) {
 // limit is exceeded.
 void HandleBeaconFrame(const Network::WifiPacket& packet) {
     std::lock_guard<std::mutex> lock(beacon_mutex);
+    const auto unique_beacon =
+        std::find_if(received_beacons.begin(), received_beacons.end(),
+                     [&packet](const Network::WifiPacket& new_packet) {
+                         return new_packet.transmitter_address == packet.transmitter_address;
+                     });
+    if (unique_beacon != received_beacons.end()) {
+        // We already have a beacon from the same mac in the deque, remove the old one;
+        received_beacons.erase(unique_beacon);
+    }
 
     received_beacons.emplace_back(packet);
 
@@ -91,14 +121,33 @@ void HandleBeaconFrame(const Network::WifiPacket& packet) {
         received_beacons.pop_front();
 }
 
+void HandleAssociationResponseFrame(const Network::WifiPacket& packet) {
+    auto assoc_result = GetAssociationResult(packet.data);
+
+    ASSERT_MSG(std::get<AssocStatus>(assoc_result) == AssocStatus::Successful,
+               "Could not join network");
+    {
+        std::lock_guard<std::mutex> lock(connection_status_mutex);
+        ASSERT(connection_status.status == static_cast<u32>(NetworkStatus::Connecting));
+    }
+
+    // Send the EAPoL-Start packet to the server.
+    using Network::WifiPacket;
+    WifiPacket eapol_start;
+    eapol_start.channel = network_channel;
+    eapol_start.data = GenerateEAPoLStartFrame(std::get<u16>(assoc_result), current_node);
+    // TODO(B3N30): Encrypt the packet.
+    eapol_start.destination_address = packet.transmitter_address;
+    eapol_start.type = WifiPacket::PacketType::Data;
+
+    SendPacket(eapol_start);
+}
+
 /*
  * Returns an available index in the nodes array for the
  * currently-hosted UDS network.
  */
 static u16 GetNextAvailableNodeId() {
-    ASSERT_MSG(connection_status.status == static_cast<u32>(NetworkStatus::ConnectedAsHost),
-               "Can not accept clients if we're not hosting a network");
-
     for (u16 index = 0; index < connection_status.max_nodes; ++index) {
         if ((connection_status.node_bitmask & (1 << index)) == 0)
             return index;
@@ -113,35 +162,46 @@ static u16 GetNextAvailableNodeId() {
  * authentication frame with SEQ1.
  */
 void StartConnectionSequence(const MacAddress& server) {
-    ASSERT(connection_status.status == static_cast<u32>(NetworkStatus::NotConnected));
-
-    // TODO(Subv): Handle timeout.
-
-    // Send an authentication frame with SEQ1
     using Network::WifiPacket;
     WifiPacket auth_request;
-    auth_request.channel = network_channel;
-    auth_request.data = GenerateAuthenticationFrame(AuthenticationSeq::SEQ1);
-    auth_request.destination_address = server;
-    auth_request.type = WifiPacket::PacketType::Authentication;
+    {
+        std::lock_guard<std::mutex> lock(connection_status_mutex);
+        ASSERT(connection_status.status == static_cast<u32>(NetworkStatus::NotConnected));
+
+        // TODO(Subv): Handle timeout.
+
+        // Send an authentication frame with SEQ1
+        auth_request.channel = network_channel;
+        auth_request.data = GenerateAuthenticationFrame(AuthenticationSeq::SEQ1);
+        auth_request.destination_address = server;
+        auth_request.type = WifiPacket::PacketType::Authentication;
+    }
 
     SendPacket(auth_request);
 }
 
 /// Sends an Association Response frame to the specified mac address
 void SendAssociationResponseFrame(const MacAddress& address) {
-    ASSERT_MSG(connection_status.status == static_cast<u32>(NetworkStatus::ConnectedAsHost));
-
     using Network::WifiPacket;
     WifiPacket assoc_response;
-    assoc_response.channel = network_channel;
-    // TODO(Subv): This will cause multiple clients to end up with the same association id, but
-    // we're not using that for anything.
-    u16 association_id = 1;
-    assoc_response.data = GenerateAssocResponseFrame(AssocStatus::Successful, association_id,
-                                                     network_info.network_id);
-    assoc_response.destination_address = address;
-    assoc_response.type = WifiPacket::PacketType::AssociationResponse;
+
+    {
+        std::lock_guard<std::mutex> lock(connection_status_mutex);
+        if (connection_status.status != static_cast<u32>(NetworkStatus::ConnectedAsHost)) {
+            LOG_ERROR(Service_NWM, "Connection sequence aborted, because connection status is %u",
+                      connection_status.status);
+            return;
+        }
+
+        assoc_response.channel = network_channel;
+        // TODO(Subv): This will cause multiple clients to end up with the same association id, but
+        // we're not using that for anything.
+        u16 association_id = 1;
+        assoc_response.data = GenerateAssocResponseFrame(AssocStatus::Successful, association_id,
+                                                         network_info.network_id);
+        assoc_response.destination_address = address;
+        assoc_response.type = WifiPacket::PacketType::AssociationResponse;
+    }
 
     SendPacket(assoc_response);
 }
@@ -155,16 +215,23 @@ void SendAssociationResponseFrame(const MacAddress& address) {
 void HandleAuthenticationFrame(const Network::WifiPacket& packet) {
     // Only the SEQ1 auth frame is handled here, the SEQ2 frame doesn't need any special behavior
     if (GetAuthenticationSeqNumber(packet.data) == AuthenticationSeq::SEQ1) {
-        ASSERT_MSG(connection_status.status == static_cast<u32>(NetworkStatus::ConnectedAsHost));
-
-        // Respond with an authentication response frame with SEQ2
         using Network::WifiPacket;
         WifiPacket auth_request;
-        auth_request.channel = network_channel;
-        auth_request.data = GenerateAuthenticationFrame(AuthenticationSeq::SEQ2);
-        auth_request.destination_address = packet.transmitter_address;
-        auth_request.type = WifiPacket::PacketType::Authentication;
-
+        {
+            std::lock_guard<std::mutex> lock(connection_status_mutex);
+            if (connection_status.status != static_cast<u32>(NetworkStatus::ConnectedAsHost)) {
+                LOG_ERROR(Service_NWM,
+                          "Connection sequence aborted, because connection status is %u",
+                          connection_status.status);
+                return;
+            }
+
+            // Respond with an authentication response frame with SEQ2
+            auth_request.channel = network_channel;
+            auth_request.data = GenerateAuthenticationFrame(AuthenticationSeq::SEQ2);
+            auth_request.destination_address = packet.transmitter_address;
+            auth_request.type = WifiPacket::PacketType::Authentication;
+        }
         SendPacket(auth_request);
 
         SendAssociationResponseFrame(packet.transmitter_address);
@@ -180,6 +247,9 @@ void OnWifiPacketReceived(const Network::WifiPacket& packet) {
     case Network::WifiPacket::PacketType::Authentication:
         HandleAuthenticationFrame(packet);
         break;
+    case Network::WifiPacket::PacketType::AssociationResponse:
+        HandleAssociationResponseFrame(packet);
+        break;
     }
 }
 
@@ -305,7 +375,7 @@ static void InitializeWithVersion(Interface* self) {
     u32 sharedmem_size = rp.Pop<u32>();
 
     // Update the node information with the data the game gave us.
-    rp.PopRaw(node_info[0]);
+    rp.PopRaw(current_node);
 
     u16 version = rp.Pop<u16>();
 
@@ -315,10 +385,14 @@ static void InitializeWithVersion(Interface* self) {
 
     ASSERT_MSG(recv_buffer_memory->size == sharedmem_size, "Invalid shared memory size.");
 
-    // Reset the connection status, it contains all zeros after initialization,
-    // except for the actual status value.
-    connection_status = {};
-    connection_status.status = static_cast<u32>(NetworkStatus::NotConnected);
+    {
+        std::lock_guard<std::mutex> lock(connection_status_mutex);
+
+        // Reset the connection status, it contains all zeros after initialization,
+        // except for the actual status value.
+        connection_status = {};
+        connection_status.status = static_cast<u32>(NetworkStatus::NotConnected);
+    }
 
     IPC::RequestBuilder rb = rp.MakeBuilder(1, 2);
     rb.Push(RESULT_SUCCESS);
@@ -348,12 +422,16 @@ static void GetConnectionStatus(Interface* self) {
     IPC::RequestBuilder rb = rp.MakeBuilder(13, 0);
 
     rb.Push(RESULT_SUCCESS);
-    rb.PushRaw(connection_status);
-
-    // Reset the bitmask of changed nodes after each call to this
-    // function to prevent falsely informing games of outstanding
-    // changes in subsequent calls.
-    connection_status.changed_nodes = 0;
+    {
+        std::lock_guard<std::mutex> lock(connection_status_mutex);
+        rb.PushRaw(connection_status);
+
+        // Reset the bitmask of changed nodes after each call to this
+        // function to prevent falsely informing games of outstanding
+        // changes in subsequent calls.
+        // TODO(Subv): Find exactly where the NWM module resets this value.
+        connection_status.changed_nodes = 0;
+    }
 
     LOG_DEBUG(Service_NWM, "called");
 }
@@ -434,31 +512,36 @@ static void BeginHostingNetwork(Interface* self) {
     // The real UDS module throws a fatal error if this assert fails.
     ASSERT_MSG(network_info.max_nodes > 1, "Trying to host a network of only one member.");
 
-    connection_status.status = static_cast<u32>(NetworkStatus::ConnectedAsHost);
-
-    // Ensure the application data size is less than the maximum value.
-    ASSERT_MSG(network_info.application_data_size <= ApplicationDataSize, "Data size is too big.");
-
-    // Set up basic information for this network.
-    network_info.oui_value = NintendoOUI;
-    network_info.oui_type = static_cast<u8>(NintendoTagId::NetworkInfo);
-
-    connection_status.max_nodes = network_info.max_nodes;
-
-    // Resize the nodes list to hold max_nodes.
-    node_info.resize(network_info.max_nodes);
-
-    // There's currently only one node in the network (the host).
-    connection_status.total_nodes = 1;
-    network_info.total_nodes = 1;
-    // The host is always the first node
-    connection_status.network_node_id = 1;
-    node_info[0].network_node_id = 1;
-    connection_status.nodes[0] = connection_status.network_node_id;
-    // Set the bit 0 in the nodes bitmask to indicate that node 1 is already taken.
-    connection_status.node_bitmask |= 1;
-    // Notify the application that the first node was set.
-    connection_status.changed_nodes |= 1;
+    {
+        std::lock_guard<std::mutex> lock(connection_status_mutex);
+        connection_status.status = static_cast<u32>(NetworkStatus::ConnectedAsHost);
+
+        // Ensure the application data size is less than the maximum value.
+        ASSERT_MSG(network_info.application_data_size <= ApplicationDataSize,
+                   "Data size is too big.");
+
+        // Set up basic information for this network.
+        network_info.oui_value = NintendoOUI;
+        network_info.oui_type = static_cast<u8>(NintendoTagId::NetworkInfo);
+
+        connection_status.max_nodes = network_info.max_nodes;
+
+        // Resize the nodes list to hold max_nodes.
+        node_info.resize(network_info.max_nodes);
+
+        // There's currently only one node in the network (the host).
+        connection_status.total_nodes = 1;
+        network_info.total_nodes = 1;
+        // The host is always the first node
+        connection_status.network_node_id = 1;
+        current_node.network_node_id = 1;
+        connection_status.nodes[0] = connection_status.network_node_id;
+        // Set the bit 0 in the nodes bitmask to indicate that node 1 is already taken.
+        connection_status.node_bitmask |= 1;
+        // Notify the application that the first node was set.
+        connection_status.changed_nodes |= 1;
+        node_info[0] = current_node;
+    }
 
     // If the game has a preferred channel, use that instead.
     if (network_info.channel != 0)
@@ -495,9 +578,13 @@ static void DestroyNetwork(Interface* self) {
     // Unschedule the beacon broadcast event.
     CoreTiming::UnscheduleEvent(beacon_broadcast_event, 0);
 
-    // TODO(Subv): Check if connection_status is indeed reset after this call.
-    connection_status = {};
-    connection_status.status = static_cast<u8>(NetworkStatus::NotConnected);
+    {
+        std::lock_guard<std::mutex> lock(connection_status_mutex);
+
+        // TODO(Subv): Check if connection_status is indeed reset after this call.
+        connection_status = {};
+        connection_status.status = static_cast<u8>(NetworkStatus::NotConnected);
+    }
     connection_status_event->Signal();
 
     IPC::RequestBuilder rb = rp.MakeBuilder(1, 0);
@@ -540,17 +627,24 @@ static void SendTo(Interface* self) {
 
     IPC::RequestBuilder rb = rp.MakeBuilder(1, 0);
 
-    if (connection_status.status != static_cast<u32>(NetworkStatus::ConnectedAsClient) &&
-        connection_status.status != static_cast<u32>(NetworkStatus::ConnectedAsHost)) {
-        rb.Push(ResultCode(ErrorDescription::NotAuthorized, ErrorModule::UDS,
-                           ErrorSummary::InvalidState, ErrorLevel::Status));
-        return;
-    }
+    u16 network_node_id;
 
-    if (dest_node_id == connection_status.network_node_id) {
-        rb.Push(ResultCode(ErrorDescription::NotFound, ErrorModule::UDS,
-                           ErrorSummary::WrongArgument, ErrorLevel::Status));
-        return;
+    {
+        std::lock_guard<std::mutex> lock(connection_status_mutex);
+        if (connection_status.status != static_cast<u32>(NetworkStatus::ConnectedAsClient) &&
+            connection_status.status != static_cast<u32>(NetworkStatus::ConnectedAsHost)) {
+            rb.Push(ResultCode(ErrorDescription::NotAuthorized, ErrorModule::UDS,
+                               ErrorSummary::InvalidState, ErrorLevel::Status));
+            return;
+        }
+
+        if (dest_node_id == connection_status.network_node_id) {
+            rb.Push(ResultCode(ErrorDescription::NotFound, ErrorModule::UDS,
+                               ErrorSummary::WrongArgument, ErrorLevel::Status));
+            return;
+        }
+
+        network_node_id = connection_status.network_node_id;
     }
 
     // TODO(Subv): Do something with the flags.
@@ -567,8 +661,8 @@ static void SendTo(Interface* self) {
 
     // TODO(Subv): Increment the sequence number after each sent packet.
     u16 sequence_number = 0;
-    std::vector<u8> data_payload = GenerateDataPayload(
-        data, data_channel, dest_node_id, connection_status.network_node_id, sequence_number);
+    std::vector<u8> data_payload =
+        GenerateDataPayload(data, data_channel, dest_node_id, network_node_id, sequence_number);
 
     // TODO(Subv): Retrieve the MAC address of the dest_node_id and our own to encrypt
     // and encapsulate the payload.
@@ -595,6 +689,7 @@ static void GetChannel(Interface* self) {
     IPC::RequestParser rp(Kernel::GetCommandBuffer(), 0x1A, 0, 0);
     IPC::RequestBuilder rb = rp.MakeBuilder(2, 0);
 
+    std::lock_guard<std::mutex> lock(connection_status_mutex);
     bool is_connected = connection_status.status != static_cast<u32>(NetworkStatus::NotConnected);
 
     u8 channel = is_connected ? network_channel : 0;
@@ -766,6 +861,7 @@ static void BeaconBroadcastCallback(u64 userdata, int cycles_late) {
  * @param network_node_id Network Node Id of the connecting client.
  */
 void OnClientConnected(u16 network_node_id) {
+    std::lock_guard<std::mutex> lock(connection_status_mutex);
     ASSERT_MSG(connection_status.status == static_cast<u32>(NetworkStatus::ConnectedAsHost),
                "Can not accept clients if we're not hosting a network");
     ASSERT_MSG(connection_status.total_nodes < connection_status.max_nodes,
@@ -827,8 +923,11 @@ NWM_UDS::~NWM_UDS() {
     connection_status_event = nullptr;
     recv_buffer_memory = nullptr;
 
-    connection_status = {};
-    connection_status.status = static_cast<u32>(NetworkStatus::NotConnected);
+    {
+        std::lock_guard<std::mutex> lock(connection_status_mutex);
+        connection_status = {};
+        connection_status.status = static_cast<u32>(NetworkStatus::NotConnected);
+    }
 
     CoreTiming::UnscheduleEvent(beacon_broadcast_event, 0);
 }
diff --git a/src/core/hle/service/nwm/uds_connection.cpp b/src/core/hle/service/nwm/uds_connection.cpp
index c8a76ec2a..c74f51253 100644
--- a/src/core/hle/service/nwm/uds_connection.cpp
+++ b/src/core/hle/service/nwm/uds_connection.cpp
@@ -75,5 +75,14 @@ std::vector<u8> GenerateAssocResponseFrame(AssocStatus status, u16 association_i
     return data;
 }
 
+std::tuple<AssocStatus, u16> GetAssociationResult(const std::vector<u8>& body) {
+    AssociationResponseFrame frame;
+    memcpy(&frame, body.data(), sizeof(frame));
+
+    constexpr u16 AssociationIdMask = 0x3FFF;
+    return std::make_tuple(static_cast<AssocStatus>(frame.status_code),
+                           frame.assoc_id & AssociationIdMask);
+}
+
 } // namespace NWM
 } // namespace Service
diff --git a/src/core/hle/service/nwm/uds_connection.h b/src/core/hle/service/nwm/uds_connection.h
index 73f55a4fd..a664f8471 100644
--- a/src/core/hle/service/nwm/uds_connection.h
+++ b/src/core/hle/service/nwm/uds_connection.h
@@ -4,6 +4,7 @@
 
 #pragma once
 
+#include <tuple>
 #include <vector>
 #include "common/common_types.h"
 #include "common/swap.h"
@@ -47,5 +48,9 @@ AuthenticationSeq GetAuthenticationSeqNumber(const std::vector<u8>& body);
 /// network id, starting at the frame body.
 std::vector<u8> GenerateAssocResponseFrame(AssocStatus status, u16 association_id, u32 network_id);
 
+/// Returns a tuple of (association status, association id) from the body of an AssociationResponse
+/// frame.
+std::tuple<AssocStatus, u16> GetAssociationResult(const std::vector<u8>& body);
+
 } // namespace NWM
 } // namespace Service
diff --git a/src/core/hle/service/nwm/uds_data.cpp b/src/core/hle/service/nwm/uds_data.cpp
index 8c6742dba..0fd9b8b8c 100644
--- a/src/core/hle/service/nwm/uds_data.cpp
+++ b/src/core/hle/service/nwm/uds_data.cpp
@@ -274,5 +274,26 @@ std::vector<u8> GenerateDataPayload(const std::vector<u8>& data, u8 channel, u16
     return buffer;
 }
 
+std::vector<u8> GenerateEAPoLStartFrame(u16 association_id, const NodeInfo& node_info) {
+    EAPoLStartPacket eapol_start{};
+    eapol_start.association_id = association_id;
+    eapol_start.friend_code_seed = node_info.friend_code_seed;
+
+    for (int i = 0; i < node_info.username.size(); ++i)
+        eapol_start.username[i] = node_info.username[i];
+
+    // Note: The network_node_id and unknown bytes seem to be uninitialized in the NWM module.
+    // TODO(B3N30): The last 8 bytes seem to have a fixed value of 07 88 15 00 04 e9 13 00 in
+    // EAPoL-Start packets from different 3DSs to the same host during a Super Smash Bros. 4 game.
+    // Find out what that means.
+
+    std::vector<u8> eapol_buffer(sizeof(EAPoLStartPacket));
+    std::memcpy(eapol_buffer.data(), &eapol_start, sizeof(eapol_start));
+
+    std::vector<u8> buffer = GenerateLLCHeader(EtherType::EAPoL);
+    buffer.insert(buffer.end(), eapol_buffer.begin(), eapol_buffer.end());
+    return buffer;
+}
+
 } // namespace NWM
 } // namespace Service
diff --git a/src/core/hle/service/nwm/uds_data.h b/src/core/hle/service/nwm/uds_data.h
index a23520a41..76e8f546b 100644
--- a/src/core/hle/service/nwm/uds_data.h
+++ b/src/core/hle/service/nwm/uds_data.h
@@ -67,6 +67,27 @@ struct DataFrameCryptoCTR {
 
 static_assert(sizeof(DataFrameCryptoCTR) == 16, "DataFrameCryptoCTR has the wrong size");
 
+constexpr u16 EAPoLStartMagic = 0x201;
+
+/*
+ * Nintendo EAPoLStartPacket, is used to initaliaze a connection between client and host
+ */
+struct EAPoLStartPacket {
+    u16_be magic = EAPoLStartMagic;
+    u16_be association_id;
+    // This value is hardcoded to 1 in the NWM module.
+    u16_be unknown = 1;
+    INSERT_PADDING_BYTES(2);
+
+    u64_be friend_code_seed;
+    std::array<u16_be, 10> username;
+    INSERT_PADDING_BYTES(4);
+    u16_be network_node_id;
+    INSERT_PADDING_BYTES(6);
+};
+
+static_assert(sizeof(EAPoLStartPacket) == 0x30, "EAPoLStartPacket has the wrong size");
+
 /**
  * Generates an unencrypted 802.11 data payload.
  * @returns The generated frame payload.
@@ -74,5 +95,12 @@ static_assert(sizeof(DataFrameCryptoCTR) == 16, "DataFrameCryptoCTR has the wron
 std::vector<u8> GenerateDataPayload(const std::vector<u8>& data, u8 channel, u16 dest_node,
                                     u16 src_node, u16 sequence_number);
 
+/*
+ * Generates an unencrypted 802.11 data frame body with the EAPoL-Start format for UDS
+ * communication.
+ * @returns The generated frame body.
+ */
+std::vector<u8> GenerateEAPoLStartFrame(u16 association_id, const NodeInfo& node_info);
+
 } // namespace NWM
 } // namespace Service
diff --git a/src/core/hle/service/sm/sm.cpp b/src/core/hle/service/sm/sm.cpp
index 5e7fc68f9..854ab9a05 100644
--- a/src/core/hle/service/sm/sm.cpp
+++ b/src/core/hle/service/sm/sm.cpp
@@ -36,6 +36,10 @@ ResultVal<Kernel::SharedPtr<Kernel::ServerPort>> ServiceManager::RegisterService
     std::string name, unsigned int max_sessions) {
 
     CASCADE_CODE(ValidateServiceName(name));
+
+    if (registered_services.find(name) != registered_services.end())
+        return ERR_ALREADY_REGISTERED;
+
     Kernel::SharedPtr<Kernel::ServerPort> server_port;
     Kernel::SharedPtr<Kernel::ClientPort> client_port;
     std::tie(server_port, client_port) = Kernel::ServerPort::CreatePortPair(max_sessions, name);
diff --git a/src/core/hle/service/sm/sm.h b/src/core/hle/service/sm/sm.h
index 8f0dbf2db..9f60a7965 100644
--- a/src/core/hle/service/sm/sm.h
+++ b/src/core/hle/service/sm/sm.h
@@ -32,6 +32,9 @@ constexpr ResultCode ERR_ACCESS_DENIED(6, ErrorModule::SRV, ErrorSummary::Invali
                                        ErrorLevel::Permanent); // 0xD8E06406
 constexpr ResultCode ERR_NAME_CONTAINS_NUL(7, ErrorModule::SRV, ErrorSummary::WrongArgument,
                                            ErrorLevel::Permanent); // 0xD9006407
+constexpr ResultCode ERR_ALREADY_REGISTERED(ErrorDescription::AlreadyExists, ErrorModule::OS,
+                                            ErrorSummary::WrongArgument,
+                                            ErrorLevel::Permanent); // 0xD9001BFC
 
 class ServiceManager {
 public:
diff --git a/src/core/hle/service/sm/srv.cpp b/src/core/hle/service/sm/srv.cpp
index 352941e69..5c955cf54 100644
--- a/src/core/hle/service/sm/srv.cpp
+++ b/src/core/hle/service/sm/srv.cpp
@@ -13,6 +13,7 @@
 #include "core/hle/kernel/errors.h"
 #include "core/hle/kernel/hle_ipc.h"
 #include "core/hle/kernel/semaphore.h"
+#include "core/hle/kernel/server_port.h"
 #include "core/hle/kernel/server_session.h"
 #include "core/hle/service/sm/sm.h"
 #include "core/hle/service/sm/srv.h"
@@ -184,12 +185,35 @@ void SRV::PublishToSubscriber(Kernel::HLERequestContext& ctx) {
                 flags);
 }
 
+void SRV::RegisterService(Kernel::HLERequestContext& ctx) {
+    IPC::RequestParser rp(ctx, 0x3, 4, 0);
+
+    auto name_buf = rp.PopRaw<std::array<char, 8>>();
+    size_t name_len = rp.Pop<u32>();
+    u32 max_sessions = rp.Pop<u32>();
+
+    std::string name(name_buf.data(), std::min(name_len, name_buf.size()));
+
+    auto port = service_manager->RegisterService(name, max_sessions);
+
+    if (port.Failed()) {
+        IPC::RequestBuilder rb = rp.MakeBuilder(1, 0);
+        rb.Push(port.Code());
+        LOG_ERROR(Service_SRV, "called service=%s -> error 0x%08X", name.c_str(), port.Code().raw);
+        return;
+    }
+
+    IPC::RequestBuilder rb = rp.MakeBuilder(1, 2);
+    rb.Push(RESULT_SUCCESS);
+    rb.PushObjects(port.Unwrap());
+}
+
 SRV::SRV(std::shared_ptr<ServiceManager> service_manager)
     : ServiceFramework("srv:", 4), service_manager(std::move(service_manager)) {
     static const FunctionInfo functions[] = {
         {0x00010002, &SRV::RegisterClient, "RegisterClient"},
         {0x00020000, &SRV::EnableNotification, "EnableNotification"},
-        {0x00030100, nullptr, "RegisterService"},
+        {0x00030100, &SRV::RegisterService, "RegisterService"},
         {0x000400C0, nullptr, "UnregisterService"},
         {0x00050100, &SRV::GetServiceHandle, "GetServiceHandle"},
         {0x000600C2, nullptr, "RegisterPort"},
diff --git a/src/core/hle/service/sm/srv.h b/src/core/hle/service/sm/srv.h
index 75cca5184..aad839563 100644
--- a/src/core/hle/service/sm/srv.h
+++ b/src/core/hle/service/sm/srv.h
@@ -28,6 +28,7 @@ private:
     void Subscribe(Kernel::HLERequestContext& ctx);
     void Unsubscribe(Kernel::HLERequestContext& ctx);
     void PublishToSubscriber(Kernel::HLERequestContext& ctx);
+    void RegisterService(Kernel::HLERequestContext& ctx);
 
     std::shared_ptr<ServiceManager> service_manager;
     Kernel::SharedPtr<Kernel::Semaphore> notification_semaphore;
diff --git a/src/core/loader/3dsx.cpp b/src/core/loader/3dsx.cpp
index 69cdc0867..5ad5c5287 100644
--- a/src/core/loader/3dsx.cpp
+++ b/src/core/loader/3dsx.cpp
@@ -270,7 +270,7 @@ ResultStatus AppLoader_THREEDSX::Load() {
     Kernel::g_current_process = Kernel::Process::Create(std::move(codeset));
     Kernel::g_current_process->svc_access_mask.set();
     Kernel::g_current_process->address_mappings = default_address_mappings;
-    Memory::current_page_table = &Kernel::g_current_process->vm_manager.page_table;
+    Memory::SetCurrentPageTable(&Kernel::g_current_process->vm_manager.page_table);
 
     // Attach the default resource limit (APPLICATION) to the process
     Kernel::g_current_process->resource_limit =
@@ -278,8 +278,7 @@ ResultStatus AppLoader_THREEDSX::Load() {
 
     Kernel::g_current_process->Run(48, Kernel::DEFAULT_STACK_SIZE);
 
-    Service::FS::RegisterArchiveType(std::make_unique<FileSys::ArchiveFactory_SelfNCCH>(*this),
-                                     Service::FS::ArchiveIdCode::SelfNCCH);
+    Service::FS::RegisterSelfNCCH(*this);
 
     is_loaded = true;
     return ResultStatus::Success;
diff --git a/src/core/loader/elf.cpp b/src/core/loader/elf.cpp
index 2f27606a1..2de1f4e81 100644
--- a/src/core/loader/elf.cpp
+++ b/src/core/loader/elf.cpp
@@ -397,7 +397,7 @@ ResultStatus AppLoader_ELF::Load() {
     Kernel::g_current_process = Kernel::Process::Create(std::move(codeset));
     Kernel::g_current_process->svc_access_mask.set();
     Kernel::g_current_process->address_mappings = default_address_mappings;
-    Memory::current_page_table = &Kernel::g_current_process->vm_manager.page_table;
+    Memory::SetCurrentPageTable(&Kernel::g_current_process->vm_manager.page_table);
 
     // Attach the default resource limit (APPLICATION) to the process
     Kernel::g_current_process->resource_limit =
diff --git a/src/core/loader/loader.h b/src/core/loader/loader.h
index e731888a2..3160fd2fd 100644
--- a/src/core/loader/loader.h
+++ b/src/core/loader/loader.h
@@ -167,6 +167,19 @@ public:
     }
 
     /**
+     * Get the update RomFS of the application
+     * Since the RomFS can be huge, we return a file reference instead of copying to a buffer
+     * @param romfs_file The file containing the RomFS
+     * @param offset The offset the romfs begins on
+     * @param size The size of the romfs
+     * @return ResultStatus result of function
+     */
+    virtual ResultStatus ReadUpdateRomFS(std::shared_ptr<FileUtil::IOFile>& romfs_file, u64& offset,
+                                         u64& size) {
+        return ResultStatus::ErrorNotImplemented;
+    }
+
+    /**
      * Get the title of the application
      * @param title Reference to store the application title into
      * @return ResultStatus result of function
diff --git a/src/core/loader/ncch.cpp b/src/core/loader/ncch.cpp
index 79ea50147..5107135f9 100644
--- a/src/core/loader/ncch.cpp
+++ b/src/core/loader/ncch.cpp
@@ -13,6 +13,7 @@
 #include "common/swap.h"
 #include "core/core.h"
 #include "core/file_sys/archive_selfncch.h"
+#include "core/file_sys/ncch_container.h"
 #include "core/hle/kernel/process.h"
 #include "core/hle/kernel/resource_limit.h"
 #include "core/hle/service/cfg/cfg.h"
@@ -27,87 +28,7 @@
 
 namespace Loader {
 
-static const int kMaxSections = 8;   ///< Maximum number of sections (files) in an ExeFs
-static const int kBlockSize = 0x200; ///< Size of ExeFS blocks (in bytes)
-
-/**
- * Get the decompressed size of an LZSS compressed ExeFS file
- * @param buffer Buffer of compressed file
- * @param size Size of compressed buffer
- * @return Size of decompressed buffer
- */
-static u32 LZSS_GetDecompressedSize(const u8* buffer, u32 size) {
-    u32 offset_size = *(u32*)(buffer + size - 4);
-    return offset_size + size;
-}
-
-/**
- * Decompress ExeFS file (compressed with LZSS)
- * @param compressed Compressed buffer
- * @param compressed_size Size of compressed buffer
- * @param decompressed Decompressed buffer
- * @param decompressed_size Size of decompressed buffer
- * @return True on success, otherwise false
- */
-static bool LZSS_Decompress(const u8* compressed, u32 compressed_size, u8* decompressed,
-                            u32 decompressed_size) {
-    const u8* footer = compressed + compressed_size - 8;
-    u32 buffer_top_and_bottom = *reinterpret_cast<const u32*>(footer);
-    u32 out = decompressed_size;
-    u32 index = compressed_size - ((buffer_top_and_bottom >> 24) & 0xFF);
-    u32 stop_index = compressed_size - (buffer_top_and_bottom & 0xFFFFFF);
-
-    memset(decompressed, 0, decompressed_size);
-    memcpy(decompressed, compressed, compressed_size);
-
-    while (index > stop_index) {
-        u8 control = compressed[--index];
-
-        for (unsigned i = 0; i < 8; i++) {
-            if (index <= stop_index)
-                break;
-            if (index <= 0)
-                break;
-            if (out <= 0)
-                break;
-
-            if (control & 0x80) {
-                // Check if compression is out of bounds
-                if (index < 2)
-                    return false;
-                index -= 2;
-
-                u32 segment_offset = compressed[index] | (compressed[index + 1] << 8);
-                u32 segment_size = ((segment_offset >> 12) & 15) + 3;
-                segment_offset &= 0x0FFF;
-                segment_offset += 2;
-
-                // Check if compression is out of bounds
-                if (out < segment_size)
-                    return false;
-
-                for (unsigned j = 0; j < segment_size; j++) {
-                    // Check if compression is out of bounds
-                    if (out + segment_offset >= decompressed_size)
-                        return false;
-
-                    u8 data = decompressed[out + segment_offset];
-                    decompressed[--out] = data;
-                }
-            } else {
-                // Check if compression is out of bounds
-                if (out < 1)
-                    return false;
-                decompressed[--out] = compressed[--index];
-            }
-            control <<= 1;
-        }
-    }
-    return true;
-}
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// AppLoader_NCCH class
+static const u64 UPDATE_MASK = 0x0000000e00000000;
 
 FileType AppLoader_NCCH::IdentifyType(FileUtil::IOFile& file) {
     u32 magic;
@@ -124,15 +45,25 @@ FileType AppLoader_NCCH::IdentifyType(FileUtil::IOFile& file) {
     return FileType::Error;
 }
 
+static std::string GetUpdateNCCHPath(u64_le program_id) {
+    u32 high = static_cast<u32>((program_id | UPDATE_MASK) >> 32);
+    u32 low = static_cast<u32>((program_id | UPDATE_MASK) & 0xFFFFFFFF);
+
+    return Common::StringFromFormat("%sNintendo 3DS/%s/%s/title/%08x/%08x/content/00000000.app",
+                                    FileUtil::GetUserPath(D_SDMC_IDX).c_str(), SYSTEM_ID, SDCARD_ID,
+                                    high, low);
+}
+
 std::pair<boost::optional<u32>, ResultStatus> AppLoader_NCCH::LoadKernelSystemMode() {
     if (!is_loaded) {
-        ResultStatus res = LoadExeFS();
+        ResultStatus res = base_ncch.Load();
         if (res != ResultStatus::Success) {
             return std::make_pair(boost::none, res);
         }
     }
+
     // Set the system mode as the one from the exheader.
-    return std::make_pair(exheader_header.arm11_system_local_caps.system_mode.Value(),
+    return std::make_pair(overlay_ncch->exheader_header.arm11_system_local_caps.system_mode.Value(),
                           ResultStatus::Success);
 }
 
@@ -144,183 +75,65 @@ ResultStatus AppLoader_NCCH::LoadExec() {
         return ResultStatus::ErrorNotLoaded;
 
     std::vector<u8> code;
-    if (ResultStatus::Success == ReadCode(code)) {
+    u64_le program_id;
+    if (ResultStatus::Success == ReadCode(code) &&
+        ResultStatus::Success == ReadProgramId(program_id)) {
         std::string process_name = Common::StringFromFixedZeroTerminatedBuffer(
-            (const char*)exheader_header.codeset_info.name, 8);
+            (const char*)overlay_ncch->exheader_header.codeset_info.name, 8);
 
-        SharedPtr<CodeSet> codeset = CodeSet::Create(process_name, ncch_header.program_id);
+        SharedPtr<CodeSet> codeset = CodeSet::Create(process_name, program_id);
 
         codeset->code.offset = 0;
-        codeset->code.addr = exheader_header.codeset_info.text.address;
-        codeset->code.size = exheader_header.codeset_info.text.num_max_pages * Memory::PAGE_SIZE;
+        codeset->code.addr = overlay_ncch->exheader_header.codeset_info.text.address;
+        codeset->code.size =
+            overlay_ncch->exheader_header.codeset_info.text.num_max_pages * Memory::PAGE_SIZE;
 
         codeset->rodata.offset = codeset->code.offset + codeset->code.size;
-        codeset->rodata.addr = exheader_header.codeset_info.ro.address;
-        codeset->rodata.size = exheader_header.codeset_info.ro.num_max_pages * Memory::PAGE_SIZE;
+        codeset->rodata.addr = overlay_ncch->exheader_header.codeset_info.ro.address;
+        codeset->rodata.size =
+            overlay_ncch->exheader_header.codeset_info.ro.num_max_pages * Memory::PAGE_SIZE;
 
         // TODO(yuriks): Not sure if the bss size is added to the page-aligned .data size or just
         //               to the regular size. Playing it safe for now.
-        u32 bss_page_size = (exheader_header.codeset_info.bss_size + 0xFFF) & ~0xFFF;
+        u32 bss_page_size = (overlay_ncch->exheader_header.codeset_info.bss_size + 0xFFF) & ~0xFFF;
         code.resize(code.size() + bss_page_size, 0);
 
         codeset->data.offset = codeset->rodata.offset + codeset->rodata.size;
-        codeset->data.addr = exheader_header.codeset_info.data.address;
+        codeset->data.addr = overlay_ncch->exheader_header.codeset_info.data.address;
         codeset->data.size =
-            exheader_header.codeset_info.data.num_max_pages * Memory::PAGE_SIZE + bss_page_size;
+            overlay_ncch->exheader_header.codeset_info.data.num_max_pages * Memory::PAGE_SIZE +
+            bss_page_size;
 
         codeset->entrypoint = codeset->code.addr;
         codeset->memory = std::make_shared<std::vector<u8>>(std::move(code));
 
         Kernel::g_current_process = Kernel::Process::Create(std::move(codeset));
-        Memory::current_page_table = &Kernel::g_current_process->vm_manager.page_table;
+        Memory::SetCurrentPageTable(&Kernel::g_current_process->vm_manager.page_table);
 
         // Attach a resource limit to the process based on the resource limit category
         Kernel::g_current_process->resource_limit =
             Kernel::ResourceLimit::GetForCategory(static_cast<Kernel::ResourceLimitCategory>(
-                exheader_header.arm11_system_local_caps.resource_limit_category));
+                overlay_ncch->exheader_header.arm11_system_local_caps.resource_limit_category));
 
         // Set the default CPU core for this process
         Kernel::g_current_process->ideal_processor =
-            exheader_header.arm11_system_local_caps.ideal_processor;
+            overlay_ncch->exheader_header.arm11_system_local_caps.ideal_processor;
 
         // Copy data while converting endianness
-        std::array<u32, ARRAY_SIZE(exheader_header.arm11_kernel_caps.descriptors)> kernel_caps;
-        std::copy_n(exheader_header.arm11_kernel_caps.descriptors, kernel_caps.size(),
+        std::array<u32, ARRAY_SIZE(overlay_ncch->exheader_header.arm11_kernel_caps.descriptors)>
+            kernel_caps;
+        std::copy_n(overlay_ncch->exheader_header.arm11_kernel_caps.descriptors, kernel_caps.size(),
                     begin(kernel_caps));
         Kernel::g_current_process->ParseKernelCaps(kernel_caps.data(), kernel_caps.size());
 
-        s32 priority = exheader_header.arm11_system_local_caps.priority;
-        u32 stack_size = exheader_header.codeset_info.stack_size;
+        s32 priority = overlay_ncch->exheader_header.arm11_system_local_caps.priority;
+        u32 stack_size = overlay_ncch->exheader_header.codeset_info.stack_size;
         Kernel::g_current_process->Run(priority, stack_size);
         return ResultStatus::Success;
     }
     return ResultStatus::Error;
 }
 
-ResultStatus AppLoader_NCCH::LoadSectionExeFS(const char* name, std::vector<u8>& buffer) {
-    if (!file.IsOpen())
-        return ResultStatus::Error;
-
-    ResultStatus result = LoadExeFS();
-    if (result != ResultStatus::Success)
-        return result;
-
-    LOG_DEBUG(Loader, "%d sections:", kMaxSections);
-    // Iterate through the ExeFs archive until we find a section with the specified name...
-    for (unsigned section_number = 0; section_number < kMaxSections; section_number++) {
-        const auto& section = exefs_header.section[section_number];
-
-        // Load the specified section...
-        if (strcmp(section.name, name) == 0) {
-            LOG_DEBUG(Loader, "%d - offset: 0x%08X, size: 0x%08X, name: %s", section_number,
-                      section.offset, section.size, section.name);
-
-            s64 section_offset =
-                (section.offset + exefs_offset + sizeof(ExeFs_Header) + ncch_offset);
-            file.Seek(section_offset, SEEK_SET);
-
-            if (strcmp(section.name, ".code") == 0 && is_compressed) {
-                // Section is compressed, read compressed .code section...
-                std::unique_ptr<u8[]> temp_buffer;
-                try {
-                    temp_buffer.reset(new u8[section.size]);
-                } catch (std::bad_alloc&) {
-                    return ResultStatus::ErrorMemoryAllocationFailed;
-                }
-
-                if (file.ReadBytes(&temp_buffer[0], section.size) != section.size)
-                    return ResultStatus::Error;
-
-                // Decompress .code section...
-                u32 decompressed_size = LZSS_GetDecompressedSize(&temp_buffer[0], section.size);
-                buffer.resize(decompressed_size);
-                if (!LZSS_Decompress(&temp_buffer[0], section.size, &buffer[0], decompressed_size))
-                    return ResultStatus::ErrorInvalidFormat;
-            } else {
-                // Section is uncompressed...
-                buffer.resize(section.size);
-                if (file.ReadBytes(&buffer[0], section.size) != section.size)
-                    return ResultStatus::Error;
-            }
-            return ResultStatus::Success;
-        }
-    }
-    return ResultStatus::ErrorNotUsed;
-}
-
-ResultStatus AppLoader_NCCH::LoadExeFS() {
-    if (is_exefs_loaded)
-        return ResultStatus::Success;
-
-    if (!file.IsOpen())
-        return ResultStatus::Error;
-
-    // Reset read pointer in case this file has been read before.
-    file.Seek(0, SEEK_SET);
-
-    if (file.ReadBytes(&ncch_header, sizeof(NCCH_Header)) != sizeof(NCCH_Header))
-        return ResultStatus::Error;
-
-    // Skip NCSD header and load first NCCH (NCSD is just a container of NCCH files)...
-    if (MakeMagic('N', 'C', 'S', 'D') == ncch_header.magic) {
-        LOG_DEBUG(Loader, "Only loading the first (bootable) NCCH within the NCSD file!");
-        ncch_offset = 0x4000;
-        file.Seek(ncch_offset, SEEK_SET);
-        file.ReadBytes(&ncch_header, sizeof(NCCH_Header));
-    }
-
-    // Verify we are loading the correct file type...
-    if (MakeMagic('N', 'C', 'C', 'H') != ncch_header.magic)
-        return ResultStatus::ErrorInvalidFormat;
-
-    // Read ExHeader...
-
-    if (file.ReadBytes(&exheader_header, sizeof(ExHeader_Header)) != sizeof(ExHeader_Header))
-        return ResultStatus::Error;
-
-    is_compressed = (exheader_header.codeset_info.flags.flag & 1) == 1;
-    entry_point = exheader_header.codeset_info.text.address;
-    code_size = exheader_header.codeset_info.text.code_size;
-    stack_size = exheader_header.codeset_info.stack_size;
-    bss_size = exheader_header.codeset_info.bss_size;
-    core_version = exheader_header.arm11_system_local_caps.core_version;
-    priority = exheader_header.arm11_system_local_caps.priority;
-    resource_limit_category = exheader_header.arm11_system_local_caps.resource_limit_category;
-
-    LOG_DEBUG(Loader, "Name:                        %s", exheader_header.codeset_info.name);
-    LOG_DEBUG(Loader, "Program ID:                  %016" PRIX64, ncch_header.program_id);
-    LOG_DEBUG(Loader, "Code compressed:             %s", is_compressed ? "yes" : "no");
-    LOG_DEBUG(Loader, "Entry point:                 0x%08X", entry_point);
-    LOG_DEBUG(Loader, "Code size:                   0x%08X", code_size);
-    LOG_DEBUG(Loader, "Stack size:                  0x%08X", stack_size);
-    LOG_DEBUG(Loader, "Bss size:                    0x%08X", bss_size);
-    LOG_DEBUG(Loader, "Core version:                %d", core_version);
-    LOG_DEBUG(Loader, "Thread priority:             0x%X", priority);
-    LOG_DEBUG(Loader, "Resource limit category:     %d", resource_limit_category);
-    LOG_DEBUG(Loader, "System Mode:                 %d",
-              static_cast<int>(exheader_header.arm11_system_local_caps.system_mode));
-
-    if (exheader_header.arm11_system_local_caps.program_id != ncch_header.program_id) {
-        LOG_ERROR(Loader, "ExHeader Program ID mismatch: the ROM is probably encrypted.");
-        return ResultStatus::ErrorEncrypted;
-    }
-
-    // Read ExeFS...
-
-    exefs_offset = ncch_header.exefs_offset * kBlockSize;
-    u32 exefs_size = ncch_header.exefs_size * kBlockSize;
-
-    LOG_DEBUG(Loader, "ExeFS offset:                0x%08X", exefs_offset);
-    LOG_DEBUG(Loader, "ExeFS size:                  0x%08X", exefs_size);
-
-    file.Seek(exefs_offset + ncch_offset, SEEK_SET);
-    if (file.ReadBytes(&exefs_header, sizeof(ExeFs_Header)) != sizeof(ExeFs_Header))
-        return ResultStatus::Error;
-
-    is_exefs_loaded = true;
-    return ResultStatus::Success;
-}
-
 void AppLoader_NCCH::ParseRegionLockoutInfo() {
     std::vector<u8> smdh_buffer;
     if (ReadIcon(smdh_buffer) == ResultStatus::Success && smdh_buffer.size() >= sizeof(SMDH)) {
@@ -339,23 +152,32 @@ void AppLoader_NCCH::ParseRegionLockoutInfo() {
 }
 
 ResultStatus AppLoader_NCCH::Load() {
+    u64_le ncch_program_id;
+
     if (is_loaded)
         return ResultStatus::ErrorAlreadyLoaded;
 
-    ResultStatus result = LoadExeFS();
+    ResultStatus result = base_ncch.Load();
     if (result != ResultStatus::Success)
         return result;
 
-    std::string program_id{Common::StringFromFormat("%016" PRIX64, ncch_header.program_id)};
+    ReadProgramId(ncch_program_id);
+    std::string program_id{Common::StringFromFormat("%016" PRIX64, ncch_program_id)};
 
     LOG_INFO(Loader, "Program ID: %s", program_id.c_str());
 
+    update_ncch.OpenFile(GetUpdateNCCHPath(ncch_program_id));
+    result = update_ncch.Load();
+    if (result == ResultStatus::Success) {
+        overlay_ncch = &update_ncch;
+    }
+
     Core::Telemetry().AddField(Telemetry::FieldType::Session, "ProgramId", program_id);
 
     if (auto room_member = Network::GetRoomMember().lock()) {
         Network::GameInfo game_info;
         ReadTitle(game_info.name);
-        game_info.id = ncch_header.program_id;
+        game_info.id = ncch_program_id;
         room_member->SendGameInfo(game_info);
     }
 
@@ -365,8 +187,7 @@ ResultStatus AppLoader_NCCH::Load() {
     if (ResultStatus::Success != result)
         return result;
 
-    Service::FS::RegisterArchiveType(std::make_unique<FileSys::ArchiveFactory_SelfNCCH>(*this),
-                                     Service::FS::ArchiveIdCode::SelfNCCH);
+    Service::FS::RegisterSelfNCCH(*this);
 
     ParseRegionLockoutInfo();
 
@@ -374,61 +195,40 @@ ResultStatus AppLoader_NCCH::Load() {
 }
 
 ResultStatus AppLoader_NCCH::ReadCode(std::vector<u8>& buffer) {
-    return LoadSectionExeFS(".code", buffer);
+    return overlay_ncch->LoadSectionExeFS(".code", buffer);
 }
 
 ResultStatus AppLoader_NCCH::ReadIcon(std::vector<u8>& buffer) {
-    return LoadSectionExeFS("icon", buffer);
+    return overlay_ncch->LoadSectionExeFS("icon", buffer);
 }
 
 ResultStatus AppLoader_NCCH::ReadBanner(std::vector<u8>& buffer) {
-    return LoadSectionExeFS("banner", buffer);
+    return overlay_ncch->LoadSectionExeFS("banner", buffer);
 }
 
 ResultStatus AppLoader_NCCH::ReadLogo(std::vector<u8>& buffer) {
-    return LoadSectionExeFS("logo", buffer);
+    return overlay_ncch->LoadSectionExeFS("logo", buffer);
 }
 
 ResultStatus AppLoader_NCCH::ReadProgramId(u64& out_program_id) {
-    if (!file.IsOpen())
-        return ResultStatus::Error;
-
-    ResultStatus result = LoadExeFS();
+    ResultStatus result = base_ncch.ReadProgramId(out_program_id);
     if (result != ResultStatus::Success)
         return result;
 
-    out_program_id = ncch_header.program_id;
     return ResultStatus::Success;
 }
 
 ResultStatus AppLoader_NCCH::ReadRomFS(std::shared_ptr<FileUtil::IOFile>& romfs_file, u64& offset,
                                        u64& size) {
-    if (!file.IsOpen())
-        return ResultStatus::Error;
-
-    // Check if the NCCH has a RomFS...
-    if (ncch_header.romfs_offset != 0 && ncch_header.romfs_size != 0) {
-        u32 romfs_offset = ncch_offset + (ncch_header.romfs_offset * kBlockSize) + 0x1000;
-        u32 romfs_size = (ncch_header.romfs_size * kBlockSize) - 0x1000;
-
-        LOG_DEBUG(Loader, "RomFS offset:           0x%08X", romfs_offset);
-        LOG_DEBUG(Loader, "RomFS size:             0x%08X", romfs_size);
-
-        if (file.GetSize() < romfs_offset + romfs_size)
-            return ResultStatus::Error;
-
-        // We reopen the file, to allow its position to be independent from file's
-        romfs_file = std::make_shared<FileUtil::IOFile>(filepath, "rb");
-        if (!romfs_file->IsOpen())
-            return ResultStatus::Error;
+    return base_ncch.ReadRomFS(romfs_file, offset, size);
+}
 
-        offset = romfs_offset;
-        size = romfs_size;
+ResultStatus AppLoader_NCCH::ReadUpdateRomFS(std::shared_ptr<FileUtil::IOFile>& romfs_file,
+                                             u64& offset, u64& size) {
+    ResultStatus result = update_ncch.ReadRomFS(romfs_file, offset, size);
 
-        return ResultStatus::Success;
-    }
-    LOG_DEBUG(Loader, "NCCH has no RomFS");
-    return ResultStatus::ErrorNotUsed;
+    if (result != ResultStatus::Success)
+        return base_ncch.ReadRomFS(romfs_file, offset, size);
 }
 
 ResultStatus AppLoader_NCCH::ReadTitle(std::string& title) {
diff --git a/src/core/loader/ncch.h b/src/core/loader/ncch.h
index e40cef764..9b56465cb 100644
--- a/src/core/loader/ncch.h
+++ b/src/core/loader/ncch.h
@@ -5,155 +5,12 @@
 #pragma once
 
 #include <memory>
-#include "common/bit_field.h"
 #include "common/common_types.h"
 #include "common/swap.h"
+#include "core/file_sys/ncch_container.h"
 #include "core/loader/loader.h"
 
 ////////////////////////////////////////////////////////////////////////////////////////////////////
-/// NCCH header (Note: "NCCH" appears to be a publicly unknown acronym)
-
-struct NCCH_Header {
-    u8 signature[0x100];
-    u32_le magic;
-    u32_le content_size;
-    u8 partition_id[8];
-    u16_le maker_code;
-    u16_le version;
-    u8 reserved_0[4];
-    u64_le program_id;
-    u8 reserved_1[0x10];
-    u8 logo_region_hash[0x20];
-    u8 product_code[0x10];
-    u8 extended_header_hash[0x20];
-    u32_le extended_header_size;
-    u8 reserved_2[4];
-    u8 flags[8];
-    u32_le plain_region_offset;
-    u32_le plain_region_size;
-    u32_le logo_region_offset;
-    u32_le logo_region_size;
-    u32_le exefs_offset;
-    u32_le exefs_size;
-    u32_le exefs_hash_region_size;
-    u8 reserved_3[4];
-    u32_le romfs_offset;
-    u32_le romfs_size;
-    u32_le romfs_hash_region_size;
-    u8 reserved_4[4];
-    u8 exefs_super_block_hash[0x20];
-    u8 romfs_super_block_hash[0x20];
-};
-
-static_assert(sizeof(NCCH_Header) == 0x200, "NCCH header structure size is wrong");
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// ExeFS (executable file system) headers
-
-struct ExeFs_SectionHeader {
-    char name[8];
-    u32 offset;
-    u32 size;
-};
-
-struct ExeFs_Header {
-    ExeFs_SectionHeader section[8];
-    u8 reserved[0x80];
-    u8 hashes[8][0x20];
-};
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
-// ExHeader (executable file system header) headers
-
-struct ExHeader_SystemInfoFlags {
-    u8 reserved[5];
-    u8 flag;
-    u8 remaster_version[2];
-};
-
-struct ExHeader_CodeSegmentInfo {
-    u32 address;
-    u32 num_max_pages;
-    u32 code_size;
-};
-
-struct ExHeader_CodeSetInfo {
-    u8 name[8];
-    ExHeader_SystemInfoFlags flags;
-    ExHeader_CodeSegmentInfo text;
-    u32 stack_size;
-    ExHeader_CodeSegmentInfo ro;
-    u8 reserved[4];
-    ExHeader_CodeSegmentInfo data;
-    u32 bss_size;
-};
-
-struct ExHeader_DependencyList {
-    u8 program_id[0x30][8];
-};
-
-struct ExHeader_SystemInfo {
-    u64 save_data_size;
-    u8 jump_id[8];
-    u8 reserved_2[0x30];
-};
-
-struct ExHeader_StorageInfo {
-    u8 ext_save_data_id[8];
-    u8 system_save_data_id[8];
-    u8 reserved[8];
-    u8 access_info[7];
-    u8 other_attributes;
-};
-
-struct ExHeader_ARM11_SystemLocalCaps {
-    u64_le program_id;
-    u32_le core_version;
-    u8 reserved_flags[2];
-    union {
-        u8 flags0;
-        BitField<0, 2, u8> ideal_processor;
-        BitField<2, 2, u8> affinity_mask;
-        BitField<4, 4, u8> system_mode;
-    };
-    u8 priority;
-    u8 resource_limit_descriptor[0x10][2];
-    ExHeader_StorageInfo storage_info;
-    u8 service_access_control[0x20][8];
-    u8 ex_service_access_control[0x2][8];
-    u8 reserved[0xf];
-    u8 resource_limit_category;
-};
-
-struct ExHeader_ARM11_KernelCaps {
-    u32_le descriptors[28];
-    u8 reserved[0x10];
-};
-
-struct ExHeader_ARM9_AccessControl {
-    u8 descriptors[15];
-    u8 descversion;
-};
-
-struct ExHeader_Header {
-    ExHeader_CodeSetInfo codeset_info;
-    ExHeader_DependencyList dependency_list;
-    ExHeader_SystemInfo system_info;
-    ExHeader_ARM11_SystemLocalCaps arm11_system_local_caps;
-    ExHeader_ARM11_KernelCaps arm11_kernel_caps;
-    ExHeader_ARM9_AccessControl arm9_access_control;
-    struct {
-        u8 signature[0x100];
-        u8 ncch_public_key_modulus[0x100];
-        ExHeader_ARM11_SystemLocalCaps arm11_system_local_caps;
-        ExHeader_ARM11_KernelCaps arm11_kernel_caps;
-        ExHeader_ARM9_AccessControl arm9_access_control;
-    } access_desc;
-};
-
-static_assert(sizeof(ExHeader_Header) == 0x800, "ExHeader structure size is wrong");
-
-////////////////////////////////////////////////////////////////////////////////////////////////////
 // Loader namespace
 
 namespace Loader {
@@ -162,7 +19,8 @@ namespace Loader {
 class AppLoader_NCCH final : public AppLoader {
 public:
     AppLoader_NCCH(FileUtil::IOFile&& file, const std::string& filepath)
-        : AppLoader(std::move(file)), filepath(filepath) {}
+        : AppLoader(std::move(file)), filepath(filepath), base_ncch(filepath),
+          overlay_ncch(&base_ncch) {}
 
     /**
      * Returns the type of the file
@@ -196,48 +54,24 @@ public:
     ResultStatus ReadRomFS(std::shared_ptr<FileUtil::IOFile>& romfs_file, u64& offset,
                            u64& size) override;
 
+    ResultStatus ReadUpdateRomFS(std::shared_ptr<FileUtil::IOFile>& romfs_file, u64& offset,
+                                 u64& size) override;
+
     ResultStatus ReadTitle(std::string& title) override;
 
 private:
     /**
-     * Reads an application ExeFS section of an NCCH file into AppLoader (e.g. .code, .logo, etc.)
-     * @param name Name of section to read out of NCCH file
-     * @param buffer Vector to read data into
-     * @return ResultStatus result of function
-     */
-    ResultStatus LoadSectionExeFS(const char* name, std::vector<u8>& buffer);
-
-    /**
      * Loads .code section into memory for booting
      * @return ResultStatus result of function
      */
     ResultStatus LoadExec();
 
-    /**
-     * Ensure ExeFS is loaded and ready for reading sections
-     * @return ResultStatus result of function
-     */
-    ResultStatus LoadExeFS();
-
     /// Reads the region lockout info in the SMDH and send it to CFG service
     void ParseRegionLockoutInfo();
 
-    bool is_exefs_loaded = false;
-    bool is_compressed = false;
-
-    u32 entry_point = 0;
-    u32 code_size = 0;
-    u32 stack_size = 0;
-    u32 bss_size = 0;
-    u32 core_version = 0;
-    u8 priority = 0;
-    u8 resource_limit_category = 0;
-    u32 ncch_offset = 0; // Offset to NCCH header, can be 0 or after NCSD header
-    u32 exefs_offset = 0;
-
-    NCCH_Header ncch_header;
-    ExeFs_Header exefs_header;
-    ExHeader_Header exheader_header;
+    FileSys::NCCHContainer base_ncch;
+    FileSys::NCCHContainer update_ncch;
+    FileSys::NCCHContainer* overlay_ncch;
 
     std::string filepath;
 };
diff --git a/src/core/memory.cpp b/src/core/memory.cpp
index 2f5cdcefe..9b394f84b 100644
--- a/src/core/memory.cpp
+++ b/src/core/memory.cpp
@@ -9,6 +9,8 @@
 #include "common/common_types.h"
 #include "common/logging/log.h"
 #include "common/swap.h"
+#include "core/arm/arm_interface.h"
+#include "core/core.h"
 #include "core/hle/kernel/memory.h"
 #include "core/hle/kernel/process.h"
 #include "core/hle/lock.h"
@@ -22,10 +24,17 @@ namespace Memory {
 static std::array<u8, Memory::VRAM_SIZE> vram;
 static std::array<u8, Memory::N3DS_EXTRA_RAM_SIZE> n3ds_extra_ram;
 
-PageTable* current_page_table = nullptr;
+static PageTable* current_page_table = nullptr;
 
-std::array<u8*, PAGE_TABLE_NUM_ENTRIES>* GetCurrentPageTablePointers() {
-    return &current_page_table->pointers;
+void SetCurrentPageTable(PageTable* page_table) {
+    current_page_table = page_table;
+    if (Core::System::GetInstance().IsPoweredOn()) {
+        Core::CPU().PageTableChanged();
+    }
+}
+
+PageTable* GetCurrentPageTable() {
+    return current_page_table;
 }
 
 static void MapPages(PageTable& page_table, u32 base, u32 size, u8* memory, PageType type) {
diff --git a/src/core/memory.h b/src/core/memory.h
index b228a48c2..1865bfea0 100644
--- a/src/core/memory.h
+++ b/src/core/memory.h
@@ -182,7 +182,8 @@ enum : VAddr {
 };
 
 /// Currently active page table
-extern PageTable* current_page_table;
+void SetCurrentPageTable(PageTable* page_table);
+PageTable* GetCurrentPageTable();
 
 bool IsValidVirtualAddress(const VAddr addr);
 bool IsValidPhysicalAddress(const PAddr addr);
@@ -259,10 +260,4 @@ enum class FlushMode {
  */
 void RasterizerFlushVirtualRegion(VAddr start, u32 size, FlushMode mode);
 
-/**
- * Dynarmic has an optimization to memory accesses when the pointer to the page exists that
- * can be used by setting up the current page table as a callback. This function is used to
- * retrieve the current page table for that purpose.
- */
-std::array<u8*, PAGE_TABLE_NUM_ENTRIES>* GetCurrentPageTablePointers();
 } // namespace Memory
diff --git a/src/tests/core/arm/arm_test_common.cpp b/src/tests/core/arm/arm_test_common.cpp
index 8384ce744..cfe0d503a 100644
--- a/src/tests/core/arm/arm_test_common.cpp
+++ b/src/tests/core/arm/arm_test_common.cpp
@@ -21,7 +21,7 @@ TestEnvironment::TestEnvironment(bool mutable_memory_)
     Memory::MapIoRegion(page_table, 0x00000000, 0x80000000, test_memory);
     Memory::MapIoRegion(page_table, 0x80000000, 0x80000000, test_memory);
 
-    Memory::current_page_table = &page_table;
+    Memory::SetCurrentPageTable(&page_table);
 }
 
 TestEnvironment::~TestEnvironment() {
diff --git a/src/video_core/command_processor.cpp b/src/video_core/command_processor.cpp
index fb65a3a0a..3ab4af374 100644
--- a/src/video_core/command_processor.cpp
+++ b/src/video_core/command_processor.cpp
@@ -243,6 +243,15 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
                     ASSERT(!g_state.geometry_pipeline.NeedIndexInput());
                     g_state.geometry_pipeline.Setup(shader_engine);
                     g_state.geometry_pipeline.SubmitVertex(output);
+
+                    // TODO: If drawing after every immediate mode triangle kills performance,
+                    // change it to flush triangles whenever a drawing config register changes
+                    // See: https://github.com/citra-emu/citra/pull/2866#issuecomment-327011550
+                    VideoCore::g_renderer->Rasterizer()->DrawTriangles();
+                    if (g_debug_context) {
+                        g_debug_context->OnEvent(DebugContext::Event::FinishedPrimitiveBatch,
+                                                 nullptr);
+                    }
                 }
             }
         }
@@ -250,16 +259,7 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
     }
 
     case PICA_REG_INDEX(pipeline.gpu_mode):
-        if (regs.pipeline.gpu_mode == PipelineRegs::GPUMode::Configuring) {
-            MICROPROFILE_SCOPE(GPU_Drawing);
-
-            // Draw immediate mode triangles when GPU Mode is set to GPUMode::Configuring
-            VideoCore::g_renderer->Rasterizer()->DrawTriangles();
-
-            if (g_debug_context) {
-                g_debug_context->OnEvent(DebugContext::Event::FinishedPrimitiveBatch, nullptr);
-            }
-        }
+        // This register likely just enables vertex processing and doesn't need any special handling
         break;
 
     case PICA_REG_INDEX_WORKAROUND(pipeline.command_buffer.trigger[0], 0x23c):
@@ -398,6 +398,11 @@ static void WritePicaReg(u32 id, u32 value, u32 mask) {
                                                       range.second, range.first);
         }
 
+        VideoCore::g_renderer->Rasterizer()->DrawTriangles();
+        if (g_debug_context) {
+            g_debug_context->OnEvent(DebugContext::Event::FinishedPrimitiveBatch, nullptr);
+        }
+
         break;
     }
 
@@ -632,6 +637,6 @@ void ProcessCommandList(const u32* list, u32 size) {
     }
 }
 
-} // namespace
+} // namespace CommandProcessor
 
-} // namespace
+} // namespace Pica
diff --git a/src/video_core/pica_types.h b/src/video_core/pica_types.h
index 5d7e10066..2eafa7e9e 100644
--- a/src/video_core/pica_types.h
+++ b/src/video_core/pica_types.h
@@ -58,11 +58,12 @@ public:
     }
 
     Float<M, E> operator*(const Float<M, E>& flt) const {
-        if ((this->value == 0.f && !std::isnan(flt.value)) ||
-            (flt.value == 0.f && !std::isnan(this->value)))
-            // PICA gives 0 instead of NaN when multiplying by inf
-            return Zero();
-        return Float<M, E>::FromFloat32(ToFloat32() * flt.ToFloat32());
+        float result = value * flt.ToFloat32();
+        // PICA gives 0 instead of NaN when multiplying by inf
+        if (!std::isnan(value) && !std::isnan(flt.ToFloat32()))
+            if (std::isnan(result))
+                result = 0.f;
+        return Float<M, E>::FromFloat32(result);
     }
 
     Float<M, E> operator/(const Float<M, E>& flt) const {
@@ -78,12 +79,7 @@ public:
     }
 
     Float<M, E>& operator*=(const Float<M, E>& flt) {
-        if ((this->value == 0.f && !std::isnan(flt.value)) ||
-            (flt.value == 0.f && !std::isnan(this->value)))
-            // PICA gives 0 instead of NaN when multiplying by inf
-            *this = Zero();
-        else
-            value *= flt.ToFloat32();
+        value = operator*(flt).value;
         return *this;
     }
 
diff --git a/src/video_core/utils.h b/src/video_core/utils.h
index 7ce83a055..d8567f314 100644
--- a/src/video_core/utils.h
+++ b/src/video_core/utils.h
@@ -8,17 +8,11 @@
 
 namespace VideoCore {
 
-/**
- * Interleave the lower 3 bits of each coordinate to get the intra-block offsets, which are
- * arranged in a Z-order curve. More details on the bit manipulation at:
- * https://fgiesen.wordpress.com/2009/12/13/decoding-morton-codes/
- */
+// 8x8 Z-Order coordinate from 2D coordinates
 static inline u32 MortonInterleave(u32 x, u32 y) {
-    u32 i = (x & 7) | ((y & 7) << 8); // ---- -210
-    i = (i ^ (i << 2)) & 0x1313;      // ---2 --10
-    i = (i ^ (i << 1)) & 0x1515;      // ---2 -1-0
-    i = (i | (i >> 7)) & 0x3F;
-    return i;
+    static const u32 xlut[] = {0x00, 0x01, 0x04, 0x05, 0x10, 0x11, 0x14, 0x15};
+    static const u32 ylut[] = {0x00, 0x02, 0x08, 0x0a, 0x20, 0x22, 0x28, 0x2a};
+    return xlut[x % 8] + ylut[y % 8];
 }
 
 /**