ci: Build the Barebone agents

Build the agent images and the kext the way frida/frida does, each
agent against the SDK its GumJS devkit was built with, and only when
src/barebone/agent/ changes. Skip the rest of CI for changes that
touch nothing else.

.github/actions/setup-llvm/action.yml

@@ -0,0 +1,19 @@
+name: Set up LLVM
+description: Install the LLVM toolchain and put its tools on the path
+inputs:
+  version:
+    required: false
+    default: "21"
+    description: The LLVM major version to install
+runs:
+  using: composite
+  steps:
+    - name: Install LLVM
+      shell: bash
+      run: |
+        curl -sSL https://apt.llvm.org/llvm.sh -o llvm.sh
+        chmod +x llvm.sh
+        sudo ./llvm.sh $VERSION
+        echo "/usr/lib/llvm-$VERSION/bin" >> $GITHUB_PATH
+      env:
+        VERSION: ${{ inputs.version }}

.github/workflows/barebone-agent.yml

@@ -0,0 +1,245 @@
+name: Barebone agent
+
+on:
+  push:
+    paths:
+      - 'src/barebone/agent/**'
+      - '.github/workflows/barebone-agent.yml'
+  pull_request:
+    branches:
+      - main
+    paths:
+      - 'src/barebone/agent/**'
+      - '.github/workflows/barebone-agent.yml'
+  workflow_dispatch:
+
+jobs:
+  gum-none:
+    strategy:
+      matrix:
+        machine:
+          - none-x86-softfloat_pic
+          - none-x86_64-softfloat_pic
+          - none-x86_64-softfloat_pic_msabi
+          - none-arm-softfloat_pic
+          - none-arm64-softfloat_pic
+      fail-fast: false
+    runs-on: ubuntu-latest
+    env:
+      FRIDA_DEPS: ${{ github.workspace }}/deps
+    steps:
+      - name: Check out repo
+        uses: actions/checkout@v6
+        with:
+          submodules: recursive
+      - name: Set up Node.js
+        uses: actions/setup-node@v6
+        with:
+          node-version: 24
+      - name: Set up C toolchain
+        uses: ./.github/actions/setup-llvm
+      - name: Grab frida-gum
+        run: |
+          python releng/meson/meson.py subprojects download frida-gum
+          git -C subprojects/frida-gum submodule update --init --recursive --depth 1
+      - name: Configure Gum
+        working-directory: subprojects/frida-gum
+        run: >-
+          ./configure
+          --host=${{ matrix.machine }}
+          --enable-gumjs
+          --with-devkits=gumjs
+          --with-devkit-symbol-scope=original
+        env:
+          CC: clang
+          CXX: clang++
+      - name: Compile
+        working-directory: subprojects/frida-gum
+        run: make
+      - name: Upload GumJS devkit
+        uses: actions/upload-artifact@v7
+        with:
+          name: frida-gumjs-devkit-${{ matrix.machine }}
+          path: subprojects/frida-gum/build/bindings/gumjs/devkit/
+      - name: Upload SDK
+        uses: actions/upload-artifact@v7
+        with:
+          name: sdk-${{ matrix.machine }}
+          path: ${{ env.FRIDA_DEPS }}/sdk-${{ matrix.machine }}/
+
+  gum-macos-kernel:
+    runs-on: macos-latest
+    env:
+      FRIDA_DEPS: ${{ github.workspace }}/deps
+    steps:
+      - name: Check out repo
+        uses: actions/checkout@v6
+        with:
+          submodules: recursive
+      - name: Grab frida-gum
+        run: |
+          python3 releng/meson/meson.py subprojects download frida-gum
+          git -C subprojects/frida-gum submodule update --init --recursive --depth 1
+      - name: Configure Gum
+        working-directory: subprojects/frida-gum
+        run: >-
+          ./configure
+          --host=macos-arm64e-kernel
+          --enable-gumjs
+          --with-devkits=gumjs
+          --with-devkit-symbol-scope=original
+      - name: Compile
+        working-directory: subprojects/frida-gum
+        run: make
+      - name: Upload GumJS devkit
+        uses: actions/upload-artifact@v7
+        with:
+          name: frida-gumjs-devkit-macos-arm64e-kernel
+          path: subprojects/frida-gum/build/bindings/gumjs/devkit/
+      - name: Upload SDK
+        uses: actions/upload-artifact@v7
+        with:
+          name: sdk-macos-arm64e-kernel
+          path: ${{ env.FRIDA_DEPS }}/sdk-macos-arm64e-kernel/
+
+  blob:
+    needs: gum-none
+    strategy:
+      matrix:
+        include:
+          - flavor: win9x
+            feature: win9x
+            arch: x86
+            machine: none-x86-softfloat_pic
+            rust_target: i686-unknown-none-pic
+          - flavor: winnt
+            feature: winnt
+            arch: x86
+            machine: none-x86-softfloat_pic
+            rust_target: i686-unknown-none-pic
+          - flavor: winnt
+            feature: winnt
+            arch: x86_64
+            machine: none-x86_64-softfloat_pic_msabi
+            rust_target: x86_64-unknown-none-pic
+          - flavor: winnt
+            feature: winnt
+            arch: arm64
+            machine: none-arm64-softfloat_pic
+            rust_target: aarch64-unknown-none-pic
+          - flavor: xnu
+            feature: xnu
+            arch: arm64
+            machine: none-arm64-softfloat_pic
+            rust_target: aarch64-unknown-none-pic
+          - flavor: linux
+            feature: linux-injected
+            arch: x86
+            machine: none-x86-softfloat_pic
+            rust_target: i686-unknown-none-pic
+          - flavor: linux
+            feature: linux-injected
+            arch: x86_64
+            machine: none-x86_64-softfloat_pic
+            rust_target: x86_64-unknown-none-pic
+          - flavor: linux
+            feature: linux-injected
+            arch: arm
+            machine: none-arm-softfloat_pic
+            rust_target: armv7a-unknown-none-pic
+          - flavor: linux
+            feature: linux-injected
+            arch: arm64
+            machine: none-arm64-softfloat_pic
+            rust_target: aarch64-unknown-none-pic
+      fail-fast: false
+    runs-on: ubuntu-latest
+    env:
+      SDK: ${{ github.workspace }}/sdk-${{ matrix.machine }}
+      GUMJS_DEVKIT_DIR: ${{ github.workspace }}/gumjs-devkit
+    steps:
+      - name: Check out repo
+        uses: actions/checkout@v6
+        with:
+          submodules: recursive
+          fetch-depth: 0
+      - name: Detect version
+        run: |
+          version=$(releng/frida_version.py)
+          test -n "$version"
+          echo "FRIDA_VERSION=$version" >> $GITHUB_ENV
+      - name: Set up C toolchain
+        uses: ./.github/actions/setup-llvm
+      - name: Set up Rust
+        run: rustup component add rust-src
+      - name: Download SDK
+        uses: actions/download-artifact@v8
+        with:
+          name: sdk-${{ matrix.machine }}
+          path: ${{ env.SDK }}
+      - name: Download GumJS devkit
+        uses: actions/download-artifact@v8
+        with:
+          name: frida-gumjs-devkit-${{ matrix.machine }}
+          path: ${{ env.GUMJS_DEVKIT_DIR }}
+      - name: Build agent
+        working-directory: src/barebone/agent
+        run: >-
+          cargo build --release
+          --features ${{ matrix.feature }}
+          --target ./${{ matrix.rust_target }}.json -Z json-target-spec
+          -Z build-std=core,alloc,compiler_builtins
+          --target-dir target-${{ matrix.flavor }}
+        env:
+          RUSTC_BOOTSTRAP: 1
+          CC: >-
+            clang
+            -I${{ env.SDK }}/include
+            -I${{ env.SDK }}/include/glib-2.0
+            -I${{ env.SDK }}/lib/glib-2.0/include
+            -I${{ env.SDK }}/include/capstone
+            -I${{ env.SDK }}/include/json-glib-1.0
+          RUSTFLAGS: -L native=${{ env.SDK }}/lib
+      - name: Upload agent
+        uses: actions/upload-artifact@v7
+        with:
+          name: frida-barebone-agent-${{ matrix.flavor }}-${{ matrix.arch }}
+          path: src/barebone/agent/target-${{ matrix.flavor }}/${{ matrix.rust_target }}/release/frida-barebone-agent
+
+  kext:
+    needs: gum-macos-kernel
+    runs-on: macos-latest
+    env:
+      FRIDA_SDK: ${{ github.workspace }}/sdk-macos-arm64e-kernel
+      GUMJS_DEVKIT_DIR: ${{ github.workspace }}/gumjs-devkit
+    steps:
+      - name: Check out repo
+        uses: actions/checkout@v6
+        with:
+          submodules: recursive
+          fetch-depth: 0
+      - name: Detect version
+        run: |
+          version=$(releng/frida_version.py)
+          test -n "$version"
+          echo "FRIDA_VERSION=$version" >> $GITHUB_ENV
+      - name: Set up Rust
+        run: rustup component add rust-src
+      - name: Download SDK
+        uses: actions/download-artifact@v8
+        with:
+          name: sdk-macos-arm64e-kernel
+          path: ${{ env.FRIDA_SDK }}
+      - name: Download GumJS devkit
+        uses: actions/download-artifact@v8
+        with:
+          name: frida-gumjs-devkit-macos-arm64e-kernel
+          path: ${{ env.GUMJS_DEVKIT_DIR }}
+      - name: Build agent
+        working-directory: src/barebone/agent/macos
+        run: make
+      - name: Upload agent
+        uses: actions/upload-artifact@v7
+        with:
+          name: frida-barebone-agent-xnu-kext-arm64e
+          path: src/barebone/agent/macos/frida-agent.kext

.github/workflows/ci.yml

@@ -2,9 +2,13 @@ name: CI
 
 on:
   push:
+    paths-ignore:
+      - 'src/barebone/agent/**'
   pull_request:
     branches:
       - main
+    paths-ignore:
+      - 'src/barebone/agent/**'
   workflow_dispatch:
     inputs:
       tag:

.gitignore

@@ -6,7 +6,7 @@
 /src/barebone/agent/linux/*.mod
 /src/barebone/agent/linux/*.mod.c
 /src/barebone/agent/linux/*.o
-/src/barebone/agent/linux/*.o.bin
+/src/barebone/agent/linux/frida-agent.elf
 /src/barebone/agent/linux/.*.cmd
 /src/barebone/agent/linux/Module.symvers
 /src/barebone/agent/linux/modules.order

compat/build.py

@@ -510,9 +510,7 @@ def compile(privdir: Path, state: State):
         if not (workdir / "build.ninja").exists():
             if options is None:
                 options = load_meson_options(top_builddir, state.role, set(subprojects.keys()))
-                version_opt = next((opt for opt in options if opt.startswith("-Dfrida_version=")), None)
-                if version_opt is None:
-                    options += [f"-Dfrida_version={state.frida_version}"]
+                options += [f"-D:frida_version={state.frida_version}"]
 
             host_machine = MachineSpec(state.host_os, group.arch, state.host_config, group.triplet)
 
@@ -525,7 +523,7 @@ def compile(privdir: Path, state: State):
                 forwarded_options = [o for o in options if not is_stripped_compat_option(o)]
                 forwarded_options += ["-Ddefault_library=static"]
             else:
-                forwarded_options = [o for o in options if not is_overridden_compat_option(o)]
+                forwarded_options = list(options)
             if state.apple_tls == "enabled":
                 forwarded_options += ["-Dapple_tls=disabled"]
 
@@ -576,17 +574,10 @@ def compile(privdir: Path, state: State):
 
 
 def is_stripped_compat_option(o: str) -> bool:
-    if is_overridden_compat_option(o):
-        return True
     name = o.split("=", 1)[0].removeprefix("-D")
     return name in STRIPPED_COMPAT_OPTIONS or name.endswith(":default_library")
 
 
-def is_overridden_compat_option(o: str) -> bool:
-    name = o.split("=", 1)[0].removeprefix("-D")
-    return name in OVERRIDDEN_COMPAT_OPTIONS
-
-
 def load_meson_options(top_builddir: Path,
                        role: Role,
                        subprojects: set[str]) -> Sequence[str]:
@@ -612,13 +603,14 @@ def option_should_be_forwarded(k: "OptionKey",
     from mesonbuild import options
     from mesonbuild.mesonlib import MachineChoice
 
+    subproject = k.subproject or ""
     our_project_id = "frida-core" if role == "subproject" else ""
-    is_for_us = k.subproject == our_project_id
-    is_for_child = k.subproject in subprojects
+    is_for_us = subproject == our_project_id
+    is_for_child = subproject in subprojects
 
     if optstore.is_project_option(k):
         if is_for_us:
-            if k.name == "compiler_backend":
+            if k.name in {"compiler_backend", "assets"}:
                 return False
             tokens = k.name.split("_")
             if tokens[0] in {"helper", "agent"} and tokens[-1] in {"modern", "legacy"}:
@@ -639,9 +631,9 @@ def option_should_be_forwarded(k: "OptionKey",
             return False
         if not str(v.value):
             return False
-        if not is_for_child and role == "subproject" and k.subproject in {"", our_project_id}:
+        if not is_for_child and role == "subproject" and subproject in {"", our_project_id}:
             if not options.BUILTIN_OPTIONS[k.evolve(subproject=None, machine=MachineChoice.HOST)].yielding:
-                return k.subproject == our_project_id
+                return subproject == our_project_id
             return True
 
     if k.get_module_prefix() == "python":
@@ -798,17 +790,6 @@ STRIPPED_COMPAT_OPTIONS = {
     "frida-gum:inspector",
 }
 
-OVERRIDDEN_COMPAT_OPTIONS = {
-    "compiler_backend",
-    "assets",
-    "helper_modern",
-    "helper_legacy",
-    "agent_modern",
-    "agent_legacy",
-    "agent_emulated_modern",
-    "agent_emulated_legacy",
-}
-
 HELPER_TARGET = "frida-helper"
 HELPER_FILE_WINDOWS = Path("src") / "frida-helper.exe"
 HELPER_FILE_UNIX = Path("src") / "frida-helper"

src/barebone/agent/Cargo.toml

@@ -9,15 +9,12 @@ name = "frida_barebone_agent"
 path = "src/lib.rs"
 crate-type = ["rlib", "staticlib"]
 
-# The XNU flavour is a freestanding executable the host injects and calls into.
-# The Linux flavour is the staticlib above, linked into a .ko by linux/Kbuild.
 [[bin]]
 name = "frida-barebone-agent"
 path = "src/main.rs"
 required-features = ["blob"]
 
 [features]
-# Flavors the host injects and calls into, as opposed to the module Linux loads.
 blob = []
 # The XNU flavour itself, and the two ways it is packaged: a blob the host injects
 # into a running kernel, and a kext that kernel loads.
@@ -26,7 +23,6 @@ xnu = ["xnu-core", "blob"]
 xnu-kext = ["xnu-core"]
 win9x = ["blob"]
 winnt = ["blob"]
-linux = []
 linux-injected = ["blob"]
 
 [profile.release]

src/barebone/agent/agent-linux-arm.lds

@@ -33,8 +33,22 @@ SECTIONS {
     _set_memory_x = .; . += 4;
     _set_memory_nx = .; . += 4;
     _modules = .; . += 4;
+    _module_mutex = .; . += 4;
+    _mutex_lock = .; . += 4;
+    _mutex_unlock = .; . += 4;
     _register_module_notifier = .; . += 4;
     _unregister_module_notifier = .; . += 4;
+    _frida_kmod_link_open = .; . += 4;
+    _frida_kmod_link_close = .; . += 4;
+    _frida_kmod_link_send = .; . += 4;
+    _frida_kmod_link_recv = .; . += 4;
+    _frida_kmod_link_pending = .; . += 4;
+    _frida_kmod_exit_worker = .; . += 4;
+    _frida_kmod_layout = .; . += 4;
+    _frida_kmod_remap_writable = .; . += 4;
+    _frida_kmod_unmap_writable = .; . += 4;
+    _execmem_alloc_rw = .; . += 4;
+    _vmalloc_noprof = .; . += 4;
     _execmem_free = .; . += 4;
     _module_alloc = .; . += 4;
     _module_memfree = .; . += 4;

src/barebone/agent/agent-linux-arm64.lds

@@ -33,8 +33,48 @@ SECTIONS {
     _set_memory_x = .; . += 8;
     _set_memory_nx = .; . += 8;
     _modules = .; . += 8;
+    _module_mutex = .; . += 8;
+    _mutex_lock = .; . += 8;
+    _mutex_unlock = .; . += 8;
     _register_module_notifier = .; . += 8;
     _unregister_module_notifier = .; . += 8;
+    _frida_kmod_link_open = .; . += 8;
+    _frida_kmod_link_close = .; . += 8;
+    _frida_kmod_link_send = .; . += 8;
+    _frida_kmod_link_recv = .; . += 8;
+    _frida_kmod_link_pending = .; . += 8;
+    _frida_kmod_exit_worker = .; . += 8;
+    _frida_kmod_layout = .; . += 8;
+    _frida_kmod_remap_writable = .; . += 8;
+    _frida_kmod_unmap_writable = .; . += 8;
+    _execmem_alloc_rw = .; . += 8;
+    _vmalloc_noprof = .; . += 8;
+    _queued_spin_lock_slowpath = .; . += 8;
+    _queued_read_lock_slowpath = .; . += 8;
+    _queued_write_lock_slowpath = .; . += 8;
+    _find_get_task_by_vpid = .; . += 8;
+    ___put_task_struct = .; . += 8;
+    _task_join_group_stop = .; . += 8;
+    ___cleanup_sighand = .; . += 8;
+    _kmem_cache_free = .; . += 8;
+    _signal_cachep = .; . += 8;
+    _put_files_struct = .; . += 8;
+    _switch_task_namespaces = .; . += 8;
+    _free_fs_struct = .; . += 8;
+    ___put_cred = .; . += 8;
+    __raw_spin_lock = .; . += 8;
+    __raw_spin_unlock = .; . += 8;
+    _capable = .; . += 8;
+    _show_map = .; . += 8;
+    _show_smap = .; . += 8;
+    _proc_fill_cache = .; . += 8;
+    _proc_pid_status = .; . += 8;
+    _do_task_stat = .; . += 8;
+    _release_task = .; . += 8;
+    _proc_task_instantiate = .; . += 8;
+    _proc_fd_instantiate = .; . += 8;
+    _proc_fdinfo_instantiate = .; . += 8;
+    ___arm64_sys_prctl = .; . += 8;
     _execmem_free = .; . += 8;
     _module_alloc = .; . += 8;
     _module_memfree = .; . += 8;

src/barebone/agent/agent-linux-x86.lds

@@ -32,8 +32,22 @@ SECTIONS {
     _set_memory_x = .; . += 4;
     _set_memory_nx = .; . += 4;
     _modules = .; . += 4;
+    _module_mutex = .; . += 4;
+    _mutex_lock = .; . += 4;
+    _mutex_unlock = .; . += 4;
     _register_module_notifier = .; . += 4;
     _unregister_module_notifier = .; . += 4;
+    _frida_kmod_link_open = .; . += 4;
+    _frida_kmod_link_close = .; . += 4;
+    _frida_kmod_link_send = .; . += 4;
+    _frida_kmod_link_recv = .; . += 4;
+    _frida_kmod_link_pending = .; . += 4;
+    _frida_kmod_exit_worker = .; . += 4;
+    _frida_kmod_layout = .; . += 4;
+    _frida_kmod_remap_writable = .; . += 4;
+    _frida_kmod_unmap_writable = .; . += 4;
+    _execmem_alloc_rw = .; . += 4;
+    _vmalloc_noprof = .; . += 4;
     _text_poke_kgdb = .; . += 4;
     _execmem_free = .; . += 4;
     _module_alloc = .; . += 4;
@@ -55,6 +69,7 @@ SECTIONS {
     _request_threaded_irq = .; . += 4;
     _vmap = .; . += 4;
     _current_task = .; . += 4;
+    _pcpu_hot = .; . += 4;
     _ioremap = .; . += 4;
     _ioremap_nocache = .; . += 4;
     ___default_kernel_pte_mask = .; . += 4;

src/barebone/agent/agent-linux-x86_64.lds

@@ -33,8 +33,45 @@ SECTIONS {
     _set_memory_x = .; . += 8;
     _set_memory_nx = .; . += 8;
     _modules = .; . += 8;
+    _module_mutex = .; . += 8;
+    _mutex_lock = .; . += 8;
+    _mutex_unlock = .; . += 8;
     _register_module_notifier = .; . += 8;
     _unregister_module_notifier = .; . += 8;
+    _frida_kmod_link_open = .; . += 8;
+    _frida_kmod_link_close = .; . += 8;
+    _frida_kmod_link_send = .; . += 8;
+    _frida_kmod_link_recv = .; . += 8;
+    _frida_kmod_link_pending = .; . += 8;
+    _frida_kmod_exit_worker = .; . += 8;
+    _frida_kmod_layout = .; . += 8;
+    _frida_kmod_remap_writable = .; . += 8;
+    _frida_kmod_unmap_writable = .; . += 8;
+    _execmem_alloc_rw = .; . += 8;
+    _vmalloc_noprof = .; . += 8;
+    _find_get_task_by_vpid = .; . += 8;
+    ___put_task_struct = .; . += 8;
+    _task_join_group_stop = .; . += 8;
+    ___cleanup_sighand = .; . += 8;
+    _kmem_cache_free = .; . += 8;
+    _signal_cachep = .; . += 8;
+    _put_files_struct = .; . += 8;
+    _switch_task_namespaces = .; . += 8;
+    _free_fs_struct = .; . += 8;
+    ___put_cred = .; . += 8;
+    __raw_spin_lock = .; . += 8;
+    __raw_spin_unlock = .; . += 8;
+    _capable = .; . += 8;
+    _show_map = .; . += 8;
+    _show_smap = .; . += 8;
+    _proc_fill_cache = .; . += 8;
+    _proc_pid_status = .; . += 8;
+    _do_task_stat = .; . += 8;
+    _release_task = .; . += 8;
+    _proc_task_instantiate = .; . += 8;
+    _proc_fd_instantiate = .; . += 8;
+    _proc_fdinfo_instantiate = .; . += 8;
+    ___x64_sys_prctl = .; . += 8;
     _text_poke_kgdb = .; . += 8;
     _execmem_free = .; . += 8;
     _module_alloc = .; . += 8;
@@ -59,6 +96,8 @@ SECTIONS {
     _page_offset_base = .; . += 8;
     _current_task = .; . += 8;
     _cpu_current_top_of_stack = .; . += 8;
+    ___preempt_count = .; . += 8;
+    _pcpu_hot = .; . += 8;
     _register_die_notifier = .; . += 8;
     _unregister_die_notifier = .; . += 8;
     _ioremap = .; . += 8;

src/barebone/agent/build.rs

@@ -47,8 +47,6 @@ fn main() {
         .write_to_file(out_dir.join("bindings.rs"))
         .expect("Couldn't write bindings");
 
-    // The Linux flavour is a staticlib: kbuild owns the final link, so it — not
-    // us — decides where the devkit and libc archives come from.
     if env::var("CARGO_FEATURE_BLOB").is_ok() {
         println!(
             "cargo:rustc-link-search=native={}",
@@ -89,6 +87,10 @@ fn main() {
         // references the entrypoint the host calls; keep the linker from
         // dropping it along with the rest of the unreferenced archive members.
         println!("cargo:rustc-link-arg=--undefined=_start");
+        if env::var("CARGO_FEATURE_LINUX_INJECTED").is_ok() {
+            println!("cargo:rustc-link-arg=--undefined=frida_agent_wake");
+            println!("cargo:rustc-link-arg=--undefined=frida_agent_stop");
+        }
 
         let note = compile_version_note(&out_dir, cc, &cc_args, arch);
         println!("cargo:rustc-link-arg={}", note.to_string_lossy());

src/barebone/agent/etc/linux-kmod-socket.json

@@ -1,5 +1,6 @@
 {
   "agent": {
+    "type": "resident",
     "transport": {
       "type": "socket",
       "path": "/tmp/frida-guest.sock"

src/barebone/agent/etc/linux-kmod.json

@@ -1,5 +1,6 @@
 {
   "agent": {
+    "type": "resident",
     "transport": {
       "type": "device",
       "path": "/dev/frida"

src/barebone/agent/linux/Kbuild

@@ -2,27 +2,15 @@ MODULE_NAME ?= frida-agent
 
 obj-m += $(MODULE_NAME).o
 
-# frida-agent-prelinked.o is not built from source here: it is the Rust agent, the
-# GumJS devkit and the C library collapsed into a single relocatable object by the prelink
-# step in Makefile. Copy it in through if_changed rather than kbuild's _shipped rule,
-# so the .cmd file modpost expects gets written.
-$(MODULE_NAME)-y := frida-kmod.o frida-agent-prelinked.o
+$(MODULE_NAME)-y := frida-kmod.o frida-agent-image.o
 
-targets += frida-agent-prelinked.o
+AFLAGS_frida-agent-image.o := -DFRIDA_AGENT_IMAGE='"$(FRIDA_AGENT_IMAGE)"'
 
-quiet_cmd_frida_agent_prelinked = COPY    $@
-      cmd_frida_agent_prelinked = cp $< $@
+frida-header-has = $(if $(shell grep -ls '$(2)' $(srctree)/$(1)),-D$(3))
+ccflags-y += $(call frida-header-has,include/linux/sched.h,list_head[[:space:]]*thread_group;,FRIDA_HAVE_TASK_THREAD_GROUP)
+ccflags-y += $(call frida-header-has,include/linux/sched/signal.h,int[[:space:]]*quick_threads;,FRIDA_HAVE_SIGNAL_QUICK_THREADS)
+ccflags-y += $(call frida-header-has,include/linux/fs_struct.h,spinlock_t[[:space:]]*lock;,FRIDA_HAVE_FS_STRUCT_LOCK)
+ccflags-y += $(call frida-header-has,arch/$(SRCARCH)/include/asm/current.h,struct[[:space:]]*pcpu_hot[[:space:]]*{,FRIDA_HAVE_PCPU_HOT)
+ccflags-y += $(call frida-header-has,arch/$(SRCARCH)/include/asm/processor.h,vl_onexec\[ARM64_VEC_MAX\],FRIDA_HAVE_VECTOR_LENGTHS)
 
-$(obj)/frida-agent-prelinked.o: $(src)/frida-agent-prelinked.o.bin FORCE
-	$(call if_changed,frida_agent_prelinked)
-
-# On x86 objtool validates every object going into a module and generates its ORC
-# unwind tables. It was not built to read a JavaScript runtime linked against a
-# freestanding libc: it gives up on the first Rust switch jump table it cannot
-# follow, and the tables would only ever describe frames no kernel stack trace walks
-# into. Both the blob and the module object it is linked into have to be excused,
-# since objtool runs over each.
-OBJECT_FILES_NON_STANDARD_frida-agent-prelinked.o := y
-OBJECT_FILES_NON_STANDARD_$(MODULE_NAME).o := y
-
-ccflags-y += $(FRIDA_KMOD_DEFS)
+$(obj)/frida-agent-image.o: $(FRIDA_AGENT_IMAGE)

src/barebone/agent/linux/Makefile

@@ -1,14 +1,3 @@
-# Builds frida-agent.ko in two steps, because kbuild insists on owning the final
-# link:
-#
-#   1. Collapse the Rust staticlib plus everything it needs — the GumJS devkit and
-#      the C library — into one relocatable object, and localize every symbol except
-#      the entry points frida-kmod.c calls. Localizing matters: Gum and the C library
-#      define names like memcpy and malloc that must not leak into the kernel's
-#      namespace.
-#   2. Hand that object to kbuild as a shipped object, so vermagic, modinfo and
-#      CRCs all come from the target kernel's own build system.
-#
 # See README.md for usage.
 
 # KDIR is the kernel to build against. When it is a source tree that was configured
@@ -17,142 +6,44 @@
 KDIR ?= /lib/modules/$(shell uname -r)/build
 PROFILE ?= release
 
-# The kernel's own name for the architecture, and the one Rust knows it by. Both
-# follow from the machine this is built for, which defaults to the one building.
 ARCH ?= $(shell uname -m \
 	| sed -e 's/^i[3-6]86$$/x86/' -e 's/^x86_64$$/x86/' -e 's/^aarch64$$/arm64/')
 BITS ?= $(shell uname -m \
 	| sed -e 's/^i[3-6]86$$/32/' -e 's/^x86_64$$/64/' -e 's/^aarch64$$/64/')
 
 ifeq ($(ARCH)-$(BITS),x86-32)
-RUST_TARGET ?= i686-unknown-none
-AGENT_TARGET_FLAGS ?= --target $(AGENT_DIR)/$(RUST_TARGET).json -Z json-target-spec
+RUST_TARGET ?= i686-unknown-none-pic
 SDK_ARCH ?= x86
-TOOLPREFIX ?=
-AGENT_LD_EMULATION ?= elf_i386
+AGENT_RUSTFLAGS ?=
+AGENT_CFLAGS ?= --target=i386-none-elf
 else ifeq ($(ARCH)-$(BITS),x86-64)
-RUST_TARGET ?= x86_64-unknown-none
+RUST_TARGET ?= x86_64-unknown-none-pic
 SDK_ARCH ?= x86_64
-TOOLPREFIX ?=
-AGENT_LD_EMULATION ?= elf_x86_64
+AGENT_RUSTFLAGS ?= -Zcf-protection=branch -Cllvm-args=--min-jump-table-entries=65536
+AGENT_CFLAGS ?= --target=x86_64-none-elf -fcf-protection=branch -fno-jump-tables
 else ifeq ($(ARCH)-$(BITS),arm64-64)
-RUST_TARGET ?= aarch64-unknown-none
+RUST_TARGET ?= aarch64-unknown-none-pic
 SDK_ARCH ?= arm64
-TOOLPREFIX ?= aarch64-none-elf-
+AGENT_RUSTFLAGS ?=
+AGENT_CFLAGS ?= --target=aarch64-none-elf -ffixed-x18
 else
-$(error Unsupported ARCH "$(ARCH)" and BITS "$(BITS)"; this flavour covers \
+$(error Unsupported ARCH "$(ARCH)" and BITS "$(BITS)"; this flavor covers \
 	x86/32, x86/64 and arm64/64)
 endif
-AGENT_TARGET_FLAGS ?= --target $(RUST_TARGET)
-
-AGENT_DIR := $(CURDIR)/..
-AGENT_TARGET_DIR := $(AGENT_DIR)/target-linux
-AGENT_LIB := $(AGENT_TARGET_DIR)/$(RUST_TARGET)/$(PROFILE)/libfrida_barebone_agent.a
 
+FRIDA_SDK ?= $(error Set FRIDA_SDK to a none-$(SDK_ARCH)-softfloat_pic SDK prefix)
 GUMJS_DEVKIT_DIR ?= $(error Set GUMJS_DEVKIT_DIR to the directory holding libfrida-gumjs.a)
 
-# Any binutils that speak the target's ELF will do; nothing here is compiled with
-# them. Building for the host's own architecture, that is the host's binutils.
-AGENT_LD ?= $(TOOLPREFIX)ld
-AGENT_AR ?= $(TOOLPREFIX)ar
-AGENT_NM ?= $(TOOLPREFIX)nm
-AGENT_OBJCOPY ?= $(TOOLPREFIX)objcopy
-
-# The ABI the kernel imposes on anything linked into it. The flags only reach crates
-# Cargo compiles, and the toolchain ships core, alloc and compiler_builtins prebuilt
-# without them, so those three have to be rebuilt from source as well. That needs the
-# rust-src component and an unstable flag, hence RUSTC_BOOTSTRAP.
-#
-# On x86_64 that is endbr64 on every address-taken function, since a kernel with
-# CONFIG_X86_KERNEL_IBT faults an indirect call landing anywhere else, and switches
-# compiled as comparisons rather than jump tables, which objtool cannot follow into.
-# Soft-float, the kernel code model and the absent red zone come with the target
-# itself. On arm64
-# it is x18, which the kernel reserves for the shadow call stack, so nothing linked in
-# may use it as a general-purpose register.
-ifeq ($(ARCH)-$(BITS),x86-32)
-AGENT_RUSTFLAGS ?=
-AGENT_CFLAGS ?= --target=i386-none-elf \
-	-Xclang -target-feature -Xclang +soft-float -mno-mmx -mno-sse \
-	-mlong-double-64 -fno-pic --sysroot=$(FRIDA_SDK)
-else ifeq ($(ARCH)-$(BITS),x86-64)
-AGENT_RUSTFLAGS ?= -Zcf-protection=branch \
-	-Cllvm-args=--min-jump-table-entries=65536 \
-	-Crelocation-model=static -Zplt=yes
-AGENT_CFLAGS ?= --target=x86_64-none-elf \
-	-Xclang -target-feature -Xclang +soft-float -mno-mmx -mno-sse -mno-red-zone \
-	-mlong-double-64 -mcmodel=kernel -fcf-protection=branch -fno-jump-tables \
-	-fno-pic --sysroot=$(FRIDA_SDK)
-else
-AGENT_RUSTFLAGS ?= -Zfixed-x18
-AGENT_CFLAGS ?= --target=aarch64-none-elf -mabi=aapcs-soft -mgeneral-regs-only \
-	-ffixed-x18 -fno-pic --sysroot=$(FRIDA_SDK)
-endif
-AGENT_BUILD_STD ?= -Z build-std=core,alloc,compiler_builtins
-
-# The devkit header reaches for GLib's and capstone's, which a devkit built against
-# an SDK leaves in that SDK rather than beside itself. The sysroot's own include
-# directory is spelled out because only clang's bare-metal driver, which it selects
-# for arm64 but not for x86_64, searches there on its own.
-AGENT_CFLAGS += -isystem $(FRIDA_SDK)/include \
+AGENT_CC ?= clang
+AGENT_CFLAGS += -I$(FRIDA_SDK)/include \
 	-I$(FRIDA_SDK)/include/glib-2.0 \
 	-I$(FRIDA_SDK)/lib/glib-2.0/include \
 	-I$(FRIDA_SDK)/include/capstone \
 	-I$(FRIDA_SDK)/include/json-glib-1.0
 
-# The soft-float SDK, whose libc is picolibc and whose compiler runtime is
-# compiler-rt. GCC's own newlib and libgcc are hard-float and cannot be used here.
-FRIDA_SDK ?= $(error Set FRIDA_SDK to a none-$(SDK_ARCH)-softfloat_nopic SDK prefix)
-LIBM = $(FRIDA_SDK)/lib/libm.a
-LIBBUILTINS = $(FRIDA_SDK)/lib/libclang_rt.builtins.a
-
-# bindgen reads the devkit header through whatever this compiler reports as its
-# include path, so it has to be one that knows where the SDK keeps its libc. It has
-# to be clang either way: -mabi=aapcs-soft landed there in 19 and GCC has it at no
-# version, and on x86_64 the soft-float ABI is reachable only as an -Xclang feature.
-AGENT_CC ?= clang
-
-PRELINKED := frida-agent-prelinked.o.bin
-
-AGENT_LIB_SOFTFLOAT := $(AGENT_TARGET_DIR)/libfrida-barebone-agent-softfloat.a
-
-# On arm64 the toolchain's own compiler_builtins is compiled for the ordinary AAPCS,
-# where a double travels in d0: its __muldf3 opens with "fmov x14, d0". Everything
-# else there is -mabi=aapcs-soft and passes that same double in x0. Whole-archiving
-# the staticlib puts those definitions in front of the SDK's compiler-rt, which has
-# the soft-float ones, and every double the C library computes then comes back
-# garbage — quietly, since the arithmetic still runs. Drop them and let compiler-rt
-# answer instead. x86_64 needs none of this: soft-float is part of the Rust target
-# there, so the copy build-std compiles already passes doubles the way the C half
-# expects.
-ifeq ($(ARCH),arm64)
-DROP_HARDFLOAT_BUILTINS = \
-	$(AGENT_AR) d $@ $$($(AGENT_NM) --print-armap $@ | sed -n 's/^__muldf3 in //p') && \
-	$(AGENT_AR) s $@
-else
-DROP_HARDFLOAT_BUILTINS = :
-endif
-
-# The agent puts malloc, the C library's locks and its process stubs on the kernel's
-# own facilities, so picolibc's versions are duplicate definitions rather than
-# fallbacks. Work out which members they live in instead of naming them: the set
-# follows from what the agent defines, and picolibc moves things between releases.
-LIBC := $(AGENT_TARGET_DIR)/libc-trimmed.a
-
-ENTRY_POINTS := frida_agent_start frida_agent_stop \
-	gum_interceptor_obtain gum_interceptor_begin_transaction \
-	gum_interceptor_end_transaction gum_interceptor_replace \
-	gum_interceptor_revert
-KEEP_GLOBAL := $(foreach s,$(ENTRY_POINTS),--keep-global-symbol=$(s))
-
-# Two structures the reparenting path reaches into lost a member upstream, and
-# neither change announces itself as anything frida-kmod.c could test for. The
-# headers it is compiled against are the authority on both, so ask them.
-FRIDA_KMOD_DEFS := \
-	$(if $(shell grep -l 'list_head[[:space:]]*thread_group;' \
-		$(KDIR)/include/linux/sched.h),-DFRIDA_HAVE_TASK_THREAD_GROUP) \
-	$(if $(shell grep -l 'spinlock_t[[:space:]]*lock;' \
-		$(KDIR)/include/linux/fs_struct.h),-DFRIDA_HAVE_FS_STRUCT_LOCK)
+AGENT_DIR := $(CURDIR)/..
+AGENT_TARGET_DIR := $(AGENT_DIR)/target-linux-$(SDK_ARCH)
+AGENT_IMAGE := frida-agent.elf
 
 all: modules
 
@@ -160,47 +51,24 @@ all: modules
 # next boot, so bring-up wants a fresh one per attempt.
 MODULE_NAME ?= frida-agent
 
-modules: $(PRELINKED)
+modules: $(AGENT_IMAGE)
 	$(MAKE) -C $(KDIR) $(if $(KOUT),O=$(KOUT)) M=$(CURDIR) ARCH=$(ARCH) \
-		FRIDA_KMOD_DEFS="$(FRIDA_KMOD_DEFS)" MODULE_NAME=$(MODULE_NAME) modules
+		MODULE_NAME=$(MODULE_NAME) FRIDA_AGENT_IMAGE=$(CURDIR)/$(AGENT_IMAGE) modules
 
-$(PRELINKED): $(AGENT_LIB_SOFTFLOAT) $(LIBC) prelink.lds
-	$(AGENT_LD) -r $(if $(AGENT_LD_EMULATION),-m $(AGENT_LD_EMULATION)) \
-		-T prelink.lds -o $@.tmp \
-		--whole-archive $(AGENT_LIB_SOFTFLOAT) --no-whole-archive \
-		--start-group \
-			$(GUMJS_DEVKIT_DIR)/libfrida-gumjs.a \
-			$(LIBC) \
-			$(LIBM) \
-			$(LIBBUILTINS) \
-		--end-group
-	$(AGENT_OBJCOPY) $(KEEP_GLOBAL) --strip-debug \
-		--remove-section=.llvmbc --remove-section=.llvmcmd $@.tmp $@
-	rm -f $@.tmp
-
-$(AGENT_LIB_SOFTFLOAT): agent
-	cp $(AGENT_LIB) $@
-	$(DROP_HARDFLOAT_BUILTINS)
-
-$(LIBC): export LC_ALL = C
-$(LIBC): $(AGENT_LIB_SOFTFLOAT)
-	cp $(FRIDA_SDK)/lib/libc.a $@
-	$(AGENT_NM) --defined-only $(AGENT_LIB_SOFTFLOAT) \
-		| awk 'NF == 3 && $$2 ~ /^[TDBR]$$/ { print $$3 }' | sort -u > $@.agent-syms
-	$(AGENT_NM) --print-armap $@ | sed -n 's/^\([^ ]*\) in \(.*\)$$/\1 \2/p' | sort -u > $@.members
-	$(AGENT_AR) d $@ $$(join $@.agent-syms $@.members | awk '{ print $$2 }' | sort -u)
-	$(AGENT_AR) s $@
-	rm -f $@.agent-syms $@.members
+$(AGENT_IMAGE): agent
+	cp $(AGENT_TARGET_DIR)/$(RUST_TARGET)/$(PROFILE)/frida-barebone-agent $@
 
 agent:
-	cd $(AGENT_DIR) && RUSTC_BOOTSTRAP=1 RUSTFLAGS="$(AGENT_RUSTFLAGS)" \
-		CC_$(subst -,_,$(RUST_TARGET))="$(AGENT_CC) $(AGENT_CFLAGS)" \
-		cargo build --$(PROFILE) \
-		$(AGENT_TARGET_FLAGS) --features linux \
-		--target-dir $(AGENT_TARGET_DIR) $(AGENT_BUILD_STD)
+	cd $(AGENT_DIR) && RUSTC_BOOTSTRAP=1 \
+		RUSTFLAGS="-L native=$(FRIDA_SDK)/lib $(AGENT_RUSTFLAGS)" \
+		CC="$(AGENT_CC) $(AGENT_CFLAGS)" \
+		GUMJS_DEVKIT_DIR=$(GUMJS_DEVKIT_DIR) \
+		cargo build --$(PROFILE) --features linux-injected \
+		--target ./$(RUST_TARGET).json -Z json-target-spec \
+		--target-dir $(AGENT_TARGET_DIR) -Z build-std=core,alloc,compiler_builtins
 
 clean:
 	$(MAKE) -C $(KDIR) $(if $(KOUT),O=$(KOUT)) M=$(CURDIR) ARCH=$(ARCH) clean
-	rm -f $(PRELINKED) $(PRELINKED).tmp $(AGENT_LIB_SOFTFLOAT) $(LIBC)
+	rm -f $(AGENT_IMAGE)
 
 .PHONY: all modules agent clean

src/barebone/agent/linux/README.md

@@ -1,291 +1,201 @@
 # Linux kernel module
 
-The same agent that the Barebone backend injects into XNU from the outside, but
-loaded from the inside as an ordinary kernel module.
+The agent the Barebone backend injects into a Linux guest from the outside,
+loaded from the inside instead. The module is a loader and nothing more: it
+carries the same `linux-injected` image a hypervisor would write into guest
+memory, maps and relocates it, and starts it. Everything past that point is the
+one implementation, whichever way the agent got in: the GumJS runtime, process
+and module enumeration, cloaking, the userspace injector.
 
-## How it differs from the XNU flavor
+## What the loader does
 
-|                   | XNU                                              | Linux                                     |
-|-------------------|--------------------------------------------------|-------------------------------------------|
-| Delivery          | remote stub injects a freestanding ELF           | `insmod frida-agent.ko`                   |
-| Kernel primitives | addresses patched into `.kernel_addrs` by the host | direct calls into `frida-kmod.c`        |
-| Bootstrap config  | GVariant blob written into guest memory          | module parameters                         |
-| Transport         | virtio hostlink or vsock                         | `/dev/frida` character device             |
-| Page permissions  | host rewrites descriptors over the hostlink      | `set_memory_*()`                          |
-| Writable alias    | shadow copy, committed via physical-memory bridge | `vmap()` of the same pages               |
-| Symbols           | table supplied by the host                       | kallsyms                                  |
+|                    | From the outside                              | From the inside (`frida-agent.ko`)               |
+|--------------------|-----------------------------------------------|--------------------------------------------------|
+| Delivery           | host writes the image into guest memory       | `insmod`; the image is embedded in the module    |
+| Relocation         | host applies `.rela.dyn`                      | loader applies `.rela.dyn` (`.rel.dyn` on x86)   |
+| `.kernel_addrs`    | host fills the slots from its symbol source   | loader fills them from kallsyms                  |
+| Struct layouts     | BTF                                           | `offsetof()` table from the module's headers, BTF behind it |
+| Symbol table       | supplied by the host                          | built by the loader from kallsyms                |
+| Transport          | virtio hostlink or vsock                      | `/dev/frida` character device                    |
+| Writable alias     | host rewrites page tables                     | `vmap()` of the same pages                       |
+| Lifetime           | ends when the host detaches                   | resident until `rmmod`                           |
 
-Because the Linux agent performs its own memory work, it never issues the
-`RemapWritablePages`, `MemoryProtect` or `PatchCode` host RPCs. The wire protocol
-is otherwise identical, so the host side sees the same `(yqv)` frames.
+The configuration the agent starts from is the same GVariant either way; the
+loader writes one naming the character device as its transport. A slot the
+loader fills that the host never does — the link functions, the layout table,
+the writable alias — is what tells the agent it was loaded from the inside.
 
 ## Building
 
-Four inputs: this repository, a GumJS devkit built for
-`none-<arch>-softfloat_nopic`, the SDK it was built against, and the kernel the module is for. `Makefile`
-covers x86_64 and arm64, and defaults to the architecture of the machine building.
-
-Building for the running kernel is a good deal shorter than cross-building for a
-device, so that walkthrough comes first; the arm64 one below it is the same steps
-with the kernel obtained from elsewhere.
+Three inputs besides this repository: a GumJS devkit for
+`none-<arch>-softfloat_pic`, the SDK it was built against, and the kernel the
+module is for. Building the image needs no kernel and runs anywhere; only the
+kbuild step needs the kernel's build tree.
 
 ### Toolchain
 
-    rustup target add x86_64-unknown-none    # or aarch64-unknown-none
-    rustup component add rust-src
-    sudo dnf install clang binutils kernel-devel-$(uname -r)
+    rustup toolchain install nightly
+    rustup component add rust-src --toolchain nightly
 
-clang must be 19 or newer, since that is where `-mabi=aapcs-soft` landed; GCC does
-not implement it at any version, nor does it reach x86_64's soft-float ABI. The
-binutils are for the prelink, which reads the archive map that only GNU `nm` prints —
-the LLVM one does not; cross-building wants the target's, natively the host's own
-will do. The host compiler is for Cargo's own build scripts, which run on the machine
-doing the building.
+The image links with `rust-lld`; nothing GNU is involved. clang has to be 19 or
+newer, which is where `-mabi=aapcs-soft` landed. The kbuild step wants whatever
+compiler the kernel was built with.
 
-The commands below are run from the top of this repository, with `releng` checked
-out — `git submodule update --init releng` if the clone did not recurse.
+### The devkit and the SDK
 
-### For the running x86_64 kernel
+    version=17.26.2
+    arch=arm64          # or x86_64
+    base=https://github.com/frida/frida/releases/download/$version
+    curl -LO $base/frida-gumjs-devkit-$version-none-$arch-softfloat_pic.tar.xz
+    mkdir -p ~/gumjs-devkit
+    tar -C ~/gumjs-devkit -xf frida-gumjs-devkit-$version-none-$arch-softfloat_pic.tar.xz
 
-No prebuilt bundles are published for `none-x86_64-softfloat_nopic`, so both inputs are
-built locally, against the same SDK:
+    releng/deps.py sync sdk none-$arch-softfloat_pic ~/sdk-none-$arch-softfloat_pic
 
-    releng/deps.py build --bundle=sdk --host=none-x86_64-softfloat_nopic
-    mkdir -p ~/sdk-none-x86_64-softfloat_nopic
-    tar -C ~/sdk-none-x86_64-softfloat_nopic -xf deps/sdk-none-x86_64-softfloat_nopic.tar.xz
+`releng/deps.toml` pins which SDK a devkit was built against, and mixing the two
+leaves symbols undefined at link time.
 
-    gum=subprojects/frida-gum
-    mkdir -p $gum/build/none-x86_64-softfloat_nopic
-    (cd $gum/build/none-x86_64-softfloat_nopic \
-        && FRIDA_DEPS=$PWD/../../../../deps ../../configure \
-            --host=none-x86_64-softfloat_nopic \
-            --enable-gumjs \
-            --with-devkits=gumjs \
-            --with-devkit-symbol-scope=original \
-        && make)
+### For the running kernel
 
     make -C src/barebone/agent/linux \
-        FRIDA_SDK=$HOME/sdk-none-x86_64-softfloat_nopic \
-        GUMJS_DEVKIT_DIR=$PWD/$gum/build/none-x86_64-softfloat_nopic/bindings/gumjs/devkit
+        FRIDA_SDK=$HOME/sdk-none-$arch-softfloat_pic \
+        GUMJS_DEVKIT_DIR=$HOME/gumjs-devkit
 
-Everything else — `KDIR`, the Rust target, the binutils, the kernel's own name for the
-architecture — follows from the running kernel.
+`KDIR`, the Rust target and the kernel's name for the architecture all follow
+from the running kernel. `make frida-agent.elf` stops after the image, which is
+also what the Barebone backend injects from the outside.
 
-### For an arm64 device
+### For an Android device
 
-Here there are published bundles to start from, and the kernel has to be assembled.
-
-#### The devkit and the SDK
-
-    version=17.17.0
-    base=https://github.com/frida/frida/releases/download/$version
-    curl -LO $base/frida-gumjs-devkit-$version-none-arm64-softfloat_nopic.tar.xz
-    mkdir -p ~/gumjs-devkit
-    tar -C ~/gumjs-devkit -xf frida-gumjs-devkit-$version-none-arm64-softfloat_nopic.tar.xz
-
-    releng/deps.py sync sdk none-arm64-softfloat_nopic ~/sdk-none-arm64-softfloat_nopic
-
-The SDK carries picolibc and the compiler-rt builtins the devkit expects to be linked
-against, and `releng/deps.toml` pins which one. Mixing a devkit with an SDK built for
-a different libc leaves undefined symbols at prelink time.
-
-#### The kernel
-
-Nothing here needs the device's own kernel to be built, but the config alone is not
-enough either. `uname -r` ends in the GKI commit and the build number:
+The image is built the same way; it is the kernel tree that takes assembling.
+`uname -r` ends in the GKI commit and the build number:
 
     adb shell uname -r
-    6.1.145-android14-11-gc1de4747ac59-ab14219743
+    6.6.102-android15-8-g6eb5b2a8c46b-ab14739656-4k
 
-which is commit `c1de4747ac59` and build `14219743`. Three pieces are keyed to them:
+which is commit `6eb5b2a8c46b` and build `14739656`. The source comes from the
+commit, the CRCs `CONFIG_MODVERSIONS` checks from what CI published for the
+build, and the config from the device itself:
 
-    commit=c1de4747ac59
-    build=14219743
+    commit=6eb5b2a8c46b
+    build=14739656
     ci=https://ci.android.com/builds/submitted/$build/kernel_aarch64/latest/raw
 
-    mkdir -p ~/kernel-prepared ~/kernel-source
-    curl -sSL $ci/modules_prepare_outdir.tar.gz | tar -xz -C ~/kernel-prepared
-    curl -sSL $ci/kernel_aarch64_Module.symvers -o ~/kernel-prepared/Module.symvers
+    mkdir -p ~/kernel
     curl -sSL https://android.googlesource.com/kernel/common/+archive/$commit.tar.gz \
-        | tar -xz -C ~/kernel-source
+        | tar -xz -C ~/kernel
+    curl -sSL $ci/kernel_aarch64_Module.symvers -o ~/kernel/Module.symvers
+    adb shell "su -c 'cat /proc/config.gz'" | gunzip > ~/kernel/.config
+    make -C ~/kernel ARCH=arm64 LLVM=1 olddefconfig modules_prepare
 
-`modules_prepare_outdir.tar.gz` is the configured build tree with `scripts/` already
-compiled as x86-64 binaries, so the kbuild half has to run on an x86-64 Linux host —
-an arm64 one gets as far as `CC [M]` and then fails to run `fixdep`. `Module.symvers`
-carries the CRCs `CONFIG_MODVERSIONS` checks. The top-level `Makefile` in that prepared tree is a stub
-that includes the source's, hence both `KDIR` and `KOUT` below.
-
-A different GKI build of the same branch does not substitute for these: vermagic is
-compared verbatim, and it names the exact build.
-
-#### Build
+That wants a Linux host, of either architecture, with clang and lld. With
+`CONFIG_MODVERSIONS` the kernel matches a module by its CRCs rather than by the
+release vermagic names, so the release this tree reports does not have to be
+the device's; the vendor's own modules do not carry it either.
 
     make -C src/barebone/agent/linux \
-        FRIDA_SDK=$HOME/sdk-none-arm64-softfloat_nopic \
+        ARCH=arm64 BITS=64 \
+        FRIDA_SDK=$HOME/sdk-none-arm64-softfloat_pic \
         GUMJS_DEVKIT_DIR=$HOME/gumjs-devkit \
-        AGENT_LD=aarch64-linux-gnu-ld \
-        AGENT_AR=aarch64-linux-gnu-ar \
-        AGENT_NM=aarch64-linux-gnu-nm \
-        AGENT_OBJCOPY=aarch64-linux-gnu-objcopy \
-        KDIR=$HOME/kernel-source \
-        KOUT=$HOME/kernel-prepared \
+        KDIR=$HOME/kernel \
         LLVM=1
 
-`make` builds the Rust staticlib itself, prelinks it against the devkit and the SDK,
-then hands the result to kbuild. The `AGENT_*` overrides are only needed because the
-defaults name an `aarch64-none-elf-` GNU toolchain; drop them if that is what is
-installed.
-
-The prelink half runs anywhere; the kbuild half needs the x86-64 Linux host above.
-Cross-building from macOS means running that step in a container holding the kernel
-tree.
-
     adb push src/barebone/agent/linux/frida-agent.ko /data/local/tmp/
 
 ## Loading
 
     insmod frida-agent.ko
 
-The module's init path returns as soon as the worker thread is running — watch
-`dmesg` for `frida: listening on /dev/frida`. `rmmod frida-agent` unwinds in the
-reverse order and does not return until the worker is off the module's text.
+Init returns once the agent's worker is running and `/dev/frida` is there.
+`rmmod frida-agent` stops the agent, waits for its worker to leave the image,
+and only then frees it.
 
 ## Talking to it
 
-The agent exposes `/dev/frida` and waits: one client at a time, `read`/`write` of the
-same length-prefixed frames the other transports carry, with `poll` for readiness.
-It never dials out, so there is no network namespace to get right and no ordering
-requirement between loading the module and attaching a client.
+The agent waits on `/dev/frida`: one client at a time, `read`/`write` of the
+same length-prefixed frames the other transports carry, with `poll` for
+readiness. It never dials out, so there is no network namespace to get right
+and no ordering between loading the module and attaching a client.
 
-On Android the node is created by devtmpfs without a policy-specific label, so an
-SELinux-enforcing system needs the client's domain granted access to it before it can
-open the device. A socket-based transport was the other option considered and is a
-worse fit: `AF_VSOCK` only reaches a peer when a hypervisor is on the other side, and
-having the kernel connect out to loopback needs something already listening before
-the module loads.
+On Android devtmpfs creates the node without a policy-specific label, so an
+SELinux-enforcing system needs the client's domain granted access to it.
 
-### Reaching it from a host
-
-Serve the Barebone device from a frida-server running on the target, so the `open()`
-happens next to the device node and the host talks to an ordinary frida-server:
+Serve the Barebone device from a frida-server running on the target, so the
+`open()` happens next to the device node:
 
     FRIDA_BAREBONE_CONFIG=/data/local/tmp/linux-kmod.json \
         frida-server --device barebone -l 127.0.0.1:27042
 
-`etc/linux-kmod.json` is that config. From the host, with `adb forward tcp:27042
-tcp:27042` in place:
+`etc/linux-kmod.json` is that config. Its `"type": "resident"` is what keeps a
+client leaving from stopping the agent. From the host, with
+`adb forward tcp:27042 tcp:27042` in place:
 
     frida -H 127.0.0.1:27042 -p 0
 
 ## Injecting into userspace
 
-Loaded on a live system the module doubles as a ptrace-free injector for ordinary
-processes — how the Linux local backend reaches its targets when the module is
-present, falling back to `ptrace` when it is not. There is no second device node;
-the channel is `prctl()` with a magic option and a token in the fifth argument,
+On x86_64 and arm64 a resident agent doubles as a ptrace-free injector for
+ordinary processes — how the Linux local backend reaches its targets when the
+agent is there, falling back to `ptrace` when it is not. The channel is
+`prctl()` with a magic option and a token in the fifth argument,
 
     prctl(0x46524944, op, arg1, arg2, 0x1d5f9e6b2c7a4038)
 
 and any call whose option or token does not match falls through to the real
-prctl, so probing for the module looks like a stock kernel. `op` 32 pings and
+prctl, so probing for the agent looks like a stock kernel. `op` 32 pings and
 returns the magic.
 
-Allocate, free, read, write and spawn — the primitives the injector drives
-against a target pid — require `CAP_SYS_ADMIN`. A spawned thread is a real
-`CLONE_THREAD` sibling that adopts the target's mm, files, credentials,
-namespaces, fs and SysV semaphore undo list, so it is a native member of the
-group rather than a kernel thread wearing its address space.
+Allocate, free, read, write and spawn — what the injector drives against a
+target pid — require `CAP_SYS_ADMIN`. A spawned thread is a real member of the
+target's thread group that adopts its mm, files, credentials, namespaces, fs
+and SysV semaphore undo list, rather than a kernel thread wearing its address
+space.
 
-Cloaking is Gum's own registry mirrored into the kernel so it holds against
-`/proc` too: the module intercepts the `maps`, `smaps`, task and status readers
+Cloaking is Gum's own registry mirrored into the kernel, so it holds against
+`/proc` too: the agent intercepts the `maps`, `smaps`, task and status readers
 and drops cloaked threads, ranges and fds while keeping the counts consistent. A
-process cloaks its own resources with the token alone — no capability — so an
-injected agent self-cloaks by forwarding Gum's updates over the same channel,
-and the loader adds the footprint the agent itself never sees. A cloaked thread
-still sees through the cloak when it reads `/proc`, so the agent is not blind to
-its own process.
+process cloaks its own resources with the token alone, so an injected agent
+self-cloaks by forwarding Gum's updates over the same channel. A cloaked thread
+still sees through the cloak when it reads `/proc`.
 
 ## What the target kernel has to provide
 
-- **x86_64 or arm64.** The register-level pieces — the page size, cache maintenance,
-  where the syscall ABI leaves prctl()'s arguments, and how a spawned thread's
-  registers are set up for its return to userspace — have a half for each. The XNU
-  flavor remains arm64-only.
-- **`CONFIG_KPROBES` and `CONFIG_KALLSYMS_ALL`.** A good deal of what the shim
-  needs is compiled in but not exported — `kallsyms_lookup_name`,
-  `kallsyms_on_each_symbol` and `set_memory_*` all are, and GKI trims its export
-  table to the KMI symbol list on top of that. The shim resolves them by name
-  through a kprobe, which is the standard way back in. Without kallsyms the module
-  still loads but cannot look up symbols or change page permissions.
+- **x86_64 or arm64** for everything above; 32-bit x86 and ARM load the agent
+  but have no injector.
+- **`CONFIG_KPROBES` and `CONFIG_KALLSYMS_ALL`.** Most of what the agent calls
+  is compiled in but not exported, and GKI trims its export table to the KMI
+  symbol list on top of that. The loader finds `kallsyms_lookup_name` through a
+  kprobe and resolves the rest by name.
 - **Unsigned modules.** `CONFIG_MODULE_SIG_FORCE` must be off. Android's
   `CONFIG_MODULE_SIG_PROTECT` is fine: it only demands signatures for modules on
   the GKI list.
+- **No BTI on arm64.** The image has no landing pads, so a CPU implementing BTI
+  under a kernel built with `CONFIG_ARM64_BTI_KERNEL` faults the first indirect
+  call into it.
 
 ## Kernel ABI constraints
 
-A hardened kernel imposes an ABI on anything linked into it, and the FP constraint
-below is why this flavor is built against a soft-float SDK
-(`none-arm64-softfloat_nopic`, `none-x86_64-softfloat_nopic`) rather than an ordinary
-bare-metal one. The remaining constraints are per-architecture.
+A hardened kernel imposes an ABI on anything running inside it, which the image
+honors whichever way it got in:
 
 - **FP/SIMD.** JavaScript numbers are doubles, so a hardfloat build would use FP
-  for as long as the runtime lives, and neither architecture lets a kernel thread
-  hold it for free — x86 wants every stretch bracketed by `kernel_fpu_begin()`, and
-  on arm64 kernels before 6.9 there is no correct way to hold it at all:
-  `kernel_neon_begin()` runs with preemption disabled so we cannot sleep under it,
-  and the scheduler does not save a kernel thread's FP registers — live values are
-  lost on preemption and the FP state of whichever user task was interrupted is
-  corrupted. Building everything soft-float removes the question rather than
-  answering it: `-mabi=aapcs-soft -mgeneral-regs-only` on arm64, and on x86_64
-  LLVM's `+soft-float`, which is also what Rust's own `x86_64-unknown-none` turns
-  on, so the two halves agree on where a double travels. `-mno-sse` alone does not
-  reach the ABI; and `long double` has to be narrowed to 64 bits with
-  `-mlong-double-64`, since soft-float has no lowering for x87's 80-bit format and
-  LLVM crashes rather than calling out to a builtin.
-- **Indirect branch tracking (x86_64).** With `CONFIG_X86_KERNEL_IBT` an indirect
-  call landing on anything but an `endbr64` faults, so everything linked in is built
-  `-fcf-protection=branch` and the Rust half `-Zcf-protection=branch`.
-- **Kernel code model and red zone (x86_64).** Modules are loaded into the top 2GB
-  of the address space and their relocations are resolved as if by `-mcmodel=kernel`,
-  and there is no red zone below the stack pointer for an interrupt to step on. Both
-  come with the Rust target; the C half is told explicitly.
-- **objtool (x86_64).** It validates every object entering a module and builds its
-  ORC unwind tables, and was not written to read a JavaScript runtime linked against
-  a freestanding libc. `Kbuild` marks the prelinked blob `OBJECT_FILES_NON_STANDARD`.
-- **Shadow call stack (arm64).** With `CONFIG_SHADOW_CALL_STACK` the kernel reserves x18, so
-  the SDK and Gum are built `-ffixed-x18` and the Rust half with `-Zfixed-x18`.
-  Anything linked in that uses x18 as a scratch register overwrites the kernel's SCS
-  pointer, and the thread then returns through whatever that address happens to hold.
-
-  The Rust flag only reaches crates Cargo compiles, so `core`, `alloc` and
-  `compiler_builtins` — which the toolchain ships prebuilt, and which between them
-  account for every x18 write that used to reach the module — are rebuilt from source
-  via `-Z build-std`. That needs `rustup component add rust-src`.
-
-## The soft-float libc
-
-The libc is picolibc and the compiler runtime is compiler-rt, both from the
-soft-float SDK that `FRIDA_SDK` points at. `Makefile` trims a copy of its
-`libc.a` before the prelink: the agent implements `malloc`, the C library's locks and
-its process stubs over the kernel's own facilities, so picolibc's are duplicate
-definitions rather than fallbacks. The set is derived from what the agent defines
-rather than listed, since picolibc moves things between releases.
-
-On arm64 those compiler-rt builtins are load-bearing in a way that is easy to miss.
-Rust's toolchain carries its own `compiler_builtins`, compiled for the ordinary AAPCS
-where a double travels in `d0`, and the prelink whole-archives the Rust staticlib — so
-without the filtering step in `Makefile` those definitions answer the C library's calls
-to `__muldf3` and friends, which pass the same double in `x0`. Nothing crashes. The
-arithmetic simply returns garbage, and the first thing to notice is GLib deciding a
-hash table need not grow. x86_64 needs none of that filtering: soft-float is part of
-the Rust target there, so the copy `-Z build-std` compiles already agrees with the C
-half.
-
-## What a script may look like
-
-The runtime compiles scripts on a kernel thread's stack, and QuickJS parses
-expressions by recursion, so the ceiling is nesting rather than length: roughly
-fourteen levels of parentheses or fifteen of object literals on arm64 Linux, whose
-kernel stack is 16 KiB — as is x86_64's. A script that exceeds it fails to compile rather than taking
-anything down, though the message is not always `stack overflow` — an over-deep
-`if`/`else` chain reports `expecting ';'`. Long scripts are fine; deep ones are not.
+  for as long as the runtime lives, and neither architecture lets a kernel
+  thread hold it for free — x86 wants every stretch bracketed by
+  `kernel_fpu_begin()`, and on arm64 kernels before 6.9 there is no correct way
+  to hold it at all. Building everything soft-float removes the question: the
+  SDK, the devkit and the Rust target all use the soft-float ABI.
+- **Indirect branch tracking (x86_64).** With `CONFIG_X86_KERNEL_IBT` an
+  indirect call landing on anything but an `endbr64` faults, so the image is
+  built `-fcf-protection=branch` and `-Zcf-protection=branch`, with switches
+  compiled as comparisons rather than `notrack` jump tables. The kernel seals
+  the `endbr64` of every function it saw no indirect call to, so the loader
+  restores it on any it hands the agent to call.
+- **Shadow call stack (arm64).** With `CONFIG_SHADOW_CALL_STACK` the kernel
+  reserves x18, so the SDK, the devkit and the Rust target all leave it alone.
+  `core`, `alloc` and `compiler_builtins` are rebuilt from source with
+  `-Z build-std` so that holds for them too.
+- **Stack.** The runtime runs on a kernel thread's stack, and QuickJS parses
+  expressions by recursion, so the ceiling is nesting rather than length: around
+  fourteen levels of parentheses or fifteen of object literals on a 16 KiB
+  kernel stack. A script that exceeds it fails to compile rather than taking
+  anything down, though the message is not always `stack overflow`.

src/barebone/agent/linux/frida-agent-image.S

@@ -0,0 +1,7 @@
+	.section .rodata, "a"
+	.balign 16
+	.globl frida_agent_image_start
+frida_agent_image_start:
+	.incbin FRIDA_AGENT_IMAGE
+	.globl frida_agent_image_end
+frida_agent_image_end:

src/barebone/agent/linux/prelink.lds

@@ -1,8 +0,0 @@
-SECTIONS
-{
-  .init_array 0 : {
-    _agent_init_start = .;
-    *(.init_array*)
-    _agent_init_end = .;
-  }
-}

src/barebone/agent/src/gthread.rs

@@ -236,7 +236,7 @@ unsafe fn rec_mutex_try_acquire(impl_: &mut RecMutexImpl) -> bool {
 
 unsafe fn rec_mutex_wait(impl_: &mut RecMutexImpl) {
     while impl_.count.load(Ordering::Relaxed) != 0 {
-        wait_on(&impl_.count, &mut || impl_.count.load(Ordering::Relaxed) == 0);
+        wait_on(&impl_.count, None, &mut || impl_.count.load(Ordering::Relaxed) == 0);
     }
 }
 
@@ -361,7 +361,7 @@ unsafe fn rw_lock_reader_try_acquire(impl_: &mut RWLockImpl) -> bool {
 
 unsafe fn rw_lock_wait_for_writers(impl_: &mut RWLockImpl) {
     while impl_.state.load(Ordering::Relaxed) & RW_WRITER_LOCK_BIT != 0 {
-        wait_on(&impl_.state, &mut || {
+        wait_on(&impl_.state, None, &mut || {
             impl_.state.load(Ordering::Relaxed) & RW_WRITER_LOCK_BIT == 0
         });
     }
@@ -369,7 +369,7 @@ unsafe fn rw_lock_wait_for_writers(impl_: &mut RWLockImpl) {
 
 unsafe fn rw_lock_wait_for_all(impl_: &mut RWLockImpl) {
     while impl_.state.load(Ordering::Relaxed) != 0 {
-        wait_on(&impl_.state, &mut || impl_.state.load(Ordering::Relaxed) == 0);
+        wait_on(&impl_.state, None, &mut || impl_.state.load(Ordering::Relaxed) == 0);
     }
 }
 
@@ -395,7 +395,7 @@ pub extern "C" fn g_cond_wait(cond: *mut GCond, mutex: *mut GMutex) {
 
         let impl_ = cond_impl(cond);
         while impl_.generation.load(Ordering::Acquire) == seen {
-            wait_on(&impl_.generation, &mut || {
+            wait_on(&impl_.generation, None, &mut || {
                 impl_.generation.load(Ordering::Acquire) != seen
             });
         }
@@ -431,11 +431,13 @@ pub extern "C" fn g_cond_wait_until(cond: *mut GCond, mutex: *mut GMutex, end_ti
 
         let impl_ = cond_impl(cond);
         let mut signalled = impl_.generation.load(Ordering::Acquire) != seen;
-        while !signalled && crate::kernel::monotonic_micros() < end_time {
-            wait_on(&impl_.generation, &mut || {
+        let mut now = crate::kernel::monotonic_micros();
+        while !signalled && now < end_time {
+            wait_on(&impl_.generation, Some((end_time - now) as u64), &mut || {
                 impl_.generation.load(Ordering::Acquire) != seen
             });
             signalled = impl_.generation.load(Ordering::Acquire) != seen;
+            now = crate::kernel::monotonic_micros();
         }
 
         g_mutex_lock(mutex);
@@ -620,7 +622,7 @@ pub extern "C" fn _g_system_thread_wait(thread: *mut GSystemThread) {
         let system_thread = thread as *mut SystemThreadImpl;
 
         while (*system_thread).thread_id.load(Ordering::Acquire) == 0 {
-            wait_on(&(*system_thread).thread_id, &mut || {
+            wait_on(&(*system_thread).thread_id, None, &mut || {
                 (*system_thread).thread_id.load(Ordering::Acquire) != 0
             });
         }
@@ -645,7 +647,7 @@ unsafe fn cleanup_wrapper_data(wrapper_data: *mut ThreadWrapperData) {
 
 unsafe fn lock_acquire(lock: &AtomicU32) {
     while !unsafe { lock_try_acquire(lock) } {
-        wait_on(lock, &mut || lock.load(Ordering::Relaxed) == 0);
+        wait_on(lock, None, &mut || lock.load(Ordering::Relaxed) == 0);
     }
 }
 
@@ -659,19 +661,11 @@ unsafe fn lock_release(lock: &AtomicU32) {
     wake_waiters(lock);
 }
 
-// Sleeps until `ready` holds or the slice expires, whichever comes first.
-//
-// These used to spin on wfe/sev. A kernel thread that spins never reaches the
-// scheduler, so a lost wakeup — or a genuine deadlock, which is how we found this —
-// starves the CPU it is on until the hardware watchdog takes the machine down. The
-// bounded sleep turns both into a stall that can be observed instead.
-//
 // The address of the word being waited on names the wait channel: XNU pairs
-// assert_wait()/thread_wakeup() on it directly, while the Linux shim has a single
-// channel and treats it as a hint. Every caller re-checks its own condition in a
-// loop, so a spurious wakeup costs nothing either way.
-fn wait_on<T>(word: &T, ready: &mut dyn FnMut() -> bool) {
-    crate::kernel::wait(token_of(word), Some(WAIT_SLICE_US), ready);
+// assert_wait()/thread_wakeup() on it directly. Every caller re-checks its own
+// condition in a loop, so a spurious wakeup costs nothing either way.
+fn wait_on<T>(word: &T, timeout_us: Option<u64>, ready: &mut dyn FnMut() -> bool) {
+    crate::kernel::wait(token_of(word), timeout_us, ready);
 
     // The wait above returns without sleeping whenever `ready` already holds, so on
     // its own it is not enough to keep a contended loop off the CPU: every caller
@@ -688,4 +682,3 @@ fn token_of<T>(word: &T) -> *const u8 {
     word as *const T as *const u8
 }
 
-const WAIT_SLICE_US: u64 = 1000;

src/barebone/agent/src/gum.rs

@@ -19,14 +19,12 @@ use alloc::vec::Vec;
 use core::ptr;
 use core::sync::atomic::{AtomicU32, Ordering};
 
-#[cfg(feature = "linux")]
-use crate::gum_linux::enumerate_exports_in_range;
-#[cfg(any(feature = "linux", feature = "linux-injected"))]
+#[cfg(feature = "linux-injected")]
 use crate::bindings::{
     GumFoundSymbolFunc, GumSymbolDetails, GumSymbolType_GUM_SYMBOL_FUNCTION,
     GumSymbolType_GUM_SYMBOL_OBJECT, GumSymbolType_GUM_SYMBOL_UNKNOWN, gboolean, gssize,
 };
-#[cfg(any(feature = "linux", feature = "linux-injected"))]
+#[cfg(feature = "linux-injected")]
 use crate::gum_modules::enumerate_symbols_in_range;
 #[cfg(any(feature = "win9x", feature = "winnt"))]
 use crate::gum_windows::{enumerate_exports_in_range, enumerate_imports_in_module};
@@ -192,7 +190,7 @@ extern "C" fn gum_native_module_iface_init(g_iface: gpointer, _iface_data: gpoin
                 Some(crate::signed_to_be_called_back(gum_native_module_enumerate_imports, 0));
         }
         (*iface).find_export_by_name = Some(crate::signed_to_be_called_back(gum_native_module_find_export_by_name, 0));
-        #[cfg(any(feature = "linux", feature = "linux-injected"))]
+        #[cfg(feature = "linux-injected")]
         {
             (*iface).enumerate_symbols =
                 Some(crate::signed_to_be_called_back(gum_native_module_enumerate_symbols, 0));
@@ -350,7 +348,7 @@ unsafe extern "C" fn gum_native_module_enumerate_imports(
     }
 }
 
-#[cfg(any(feature = "linux", feature = "linux-injected"))]
+#[cfg(feature = "linux-injected")]
 unsafe extern "C" fn gum_native_module_enumerate_symbols(
     self_: *mut GumModule,
     func: GumFoundSymbolFunc,
@@ -388,12 +386,12 @@ unsafe extern "C" fn gum_native_module_enumerate_symbols(
 /// What the backends implement for `gum_native_module_enumerate_symbols`: call
 /// `callback` for every symbol in `[start, end)`, stopping early when it returns
 /// false.
-#[cfg(any(feature = "linux", feature = "linux-injected"))]
+#[cfg(feature = "linux-injected")]
 pub(crate) type FoundSymbolCallback<'a> = dyn FnMut(*const gchar, u64, u64, u8, bool) -> bool + 'a;
 
-#[cfg(any(feature = "linux", feature = "linux-injected"))]
+#[cfg(feature = "linux-injected")]
 const SYMBOL_IS_OBJECT: u8 = 1;
-#[cfg(any(feature = "linux", feature = "linux-injected"))]
+#[cfg(feature = "linux-injected")]
 const SYMBOL_IS_FUNCTION: u8 = 2;
 
 /// What the backends implement for `gum_native_module_enumerate_exports`: call

src/barebone/agent/src/gum_injected.rs

@@ -116,6 +116,11 @@ pub extern "C" fn gum_memory_try_remap_writable_pages(
         return remap_agent_pages(first_page, n_pages);
     }
 
+    #[cfg(feature = "linux-injected")]
+    if let Some(alias) = kernel::remap_writable(first_page as u64, n_pages) {
+        return alias as gpointer;
+    }
+
     #[cfg(all(feature = "linux-injected", not(target_arch = "arm")))]
     if writable_in_place(first_page, n_pages) {
         return first_page;
@@ -271,6 +276,11 @@ pub extern "C" fn gum_memory_dispose_writable_pages(writable: gpointer, _n_pages
         return;
     }
 
+    #[cfg(feature = "linux-injected")]
+    if kernel::unmap_writable(writable as *mut core::ffi::c_void) {
+        return;
+    }
+
     if let Some(executable) = aliases().remove(&(writable as u64)) {
         let size = _n_pages as usize * unsafe { gum_query_page_size() } as usize;
         libc::__clear_cache(executable as *const u8, (executable + size as u64) as *const u8);
@@ -591,7 +601,7 @@ pub extern "C" fn gum_memory_allocate(
     #[cfg(all(not(feature = "xnu-core"), not(feature = "linux-injected")))]
     let ptr = kernel::alloc_code(size as usize);
     #[cfg(feature = "linux-injected")]
-    let ptr = kernel::alloc_heap(size as usize);
+    let ptr = kernel::alloc_code(size as usize);
     #[cfg(not(feature = "xnu-core"))]
     let _ = address;
     #[cfg(not(feature = "xnu-kext"))]

src/barebone/agent/src/gum_linux.rs

@@ -1,216 +0,0 @@
-// Linux half of Gum's platform backend. Everything is done locally: we are the
-// kernel, so page permissions go through set_memory_*(), writable aliases of
-// write-protected text come from vmap(), and symbols come from kallsyms.
-
-use crate::{
-    bindings::{
-        _GumPageProtection_GUM_PAGE_EXECUTE, _GumPageProtection_GUM_PAGE_READ,
-        _GumPageProtection_GUM_PAGE_WRITE, _GumRwxSupport_GUM_RWX_NONE, GArray,
-        GumDebugSymbolDetails, GumModuleRegistry, GumPageProtection, GumRwxSupport,
-        g_array_append_vals, g_array_new, g_pattern_spec_free, g_pattern_spec_match_string,
-        g_pattern_spec_new, g_strdup, gboolean, gchar, gconstpointer, gpointer, gsize, guint,
-        gum_mprotect,
-    },
-    gum::{self, FoundExportCallback},
-    kernel,
-};
-use core::ffi::CStr;
-use core::ptr;
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_barebone_query_platform() -> *const crate::bindings::gchar {
-    c"linux".as_ptr() as *const crate::bindings::gchar
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_barebone_query_stack_size() -> gsize {
-    crate::linux::stack_headroom() as gsize
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_query_rwx_support() -> GumRwxSupport {
-    _GumRwxSupport_GUM_RWX_NONE
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_memory_can_remap_writable() -> gboolean {
-    1
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_memory_try_remap_writable_pages(
-    first_page: gpointer,
-    n_pages: guint,
-) -> gpointer {
-    kernel::remap_writable(first_page as u64, n_pages) as gpointer
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_memory_dispose_writable_pages(writable: gpointer, _n_pages: guint) {
-    kernel::unmap_writable(writable as *mut core::ffi::c_void);
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_try_mprotect(
-    address: gpointer,
-    size: gsize,
-    prot: GumPageProtection,
-) -> gboolean {
-    if kernel::protect(address as u64, size as usize, prot as u32) {
-        1
-    } else {
-        0
-    }
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_memory_allocate(
-    _address: gpointer,
-    size: gsize,
-    _alignment: gsize,
-    prot: GumPageProtection,
-) -> gpointer {
-    let ptr = kernel::alloc_code(size as usize);
-    unsafe {
-        core::ptr::write_bytes(ptr, 0, size as usize);
-        if (prot & _GumPageProtection_GUM_PAGE_EXECUTE) != 0 {
-            gum::register_slab(ptr as u64, size as usize);
-            gum_mprotect(ptr as gpointer, size, prot);
-        }
-    }
-
-    ptr as gpointer
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_memory_free(address: gpointer, size: gsize) -> gboolean {
-    // vfree() wants the range writable again: the pages go back to the vmalloc
-    // allocator, and leaving them RX would fault the next consumer.
-    if gum::is_agent_slab(address as u64) {
-        unsafe {
-            gum_mprotect(
-                address,
-                size,
-                (_GumPageProtection_GUM_PAGE_READ | _GumPageProtection_GUM_PAGE_WRITE)
-                    as GumPageProtection,
-            );
-        }
-        gum::unregister_slab(address as u64);
-    }
-    kernel::free_code(address as *mut u8, size as usize);
-    1
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_barebone_on_registry_activating(registry: *mut GumModuleRegistry) {
-    crate::gum_modules::publish(registry);
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_barebone_on_registry_deactivating(_registry: *mut GumModuleRegistry) {
-    crate::gum_modules::unpublish();
-}
-
-pub(crate) unsafe fn enumerate_exports_in_range(
-    start_address: u64,
-    end_address: u64,
-    callback: &mut FoundExportCallback<'_>,
-) {
-    if crate::gum_modules::enumerate_exports_in_module(start_address, callback) {
-        return;
-    }
-
-    kernel::enumerate_symbols(&mut |name, address| {
-        if address < start_address || address >= end_address {
-            return true;
-        }
-        callback(name.as_ptr() as *const gchar, address)
-    });
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_symbol_details_from_address(
-    address: gpointer,
-    details: *mut GumDebugSymbolDetails,
-) -> gboolean {
-    let Some((name, symbol_address)) = kernel::symbol_from_address(address as u64) else {
-        return 0;
-    };
-
-    unsafe {
-        let name_bytes = name.as_bytes();
-        let copy_len = core::cmp::min(name_bytes.len(), (*details).symbol_name.len() - 1);
-        core::ptr::copy_nonoverlapping(
-            name_bytes.as_ptr(),
-            (*details).symbol_name.as_mut_ptr() as *mut u8,
-            copy_len,
-        );
-        (*details).symbol_name[copy_len] = 0;
-
-        (*details).address = symbol_address;
-
-        (*details).module_name[0] = 0;
-        (*details).file_name[0] = 0;
-        (*details).line_number = 0;
-        (*details).column = 0;
-    }
-
-    1
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_symbol_name_from_address(address: gpointer) -> *mut gchar {
-    let Some((name, _)) = kernel::symbol_from_address(address as u64) else {
-        return ptr::null_mut();
-    };
-
-    let mut name = name;
-    name.push('\0');
-    unsafe { g_strdup(name.as_ptr() as *const gchar) }
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_find_function(name: *const gchar) -> gpointer {
-    let address = kernel::find_function(unsafe { CStr::from_ptr(name) });
-    address as gpointer
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_find_functions_named(name: *const gchar) -> *mut GArray {
-    let wanted = unsafe { CStr::from_ptr(name) };
-
-    collect_addresses(&mut |candidate, _address| candidate == wanted)
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_find_functions_matching(pattern: *const gchar) -> *mut GArray {
-    unsafe {
-        let spec = g_pattern_spec_new(pattern);
-        let array = collect_addresses(&mut |candidate, _address| {
-            g_pattern_spec_match_string(spec, candidate.as_ptr() as *const gchar) != 0
-        });
-        g_pattern_spec_free(spec);
-        array
-    }
-}
-
-fn collect_addresses(predicate: &mut dyn FnMut(&CStr, u64) -> bool) -> *mut GArray {
-    unsafe {
-        let array = g_array_new(0, 0, core::mem::size_of::<gpointer>() as guint);
-
-        kernel::enumerate_symbols(&mut |name, address| {
-            if predicate(name, address) {
-                let addr = address as gpointer;
-                g_array_append_vals(array, &addr as *const gpointer as gconstpointer, 1);
-            }
-            true
-        });
-
-        array
-    }
-}
-
-#[unsafe(no_mangle)]
-pub extern "C" fn gum_load_symbols(_path: *const gchar) -> gboolean {
-    0
-}

src/barebone/agent/src/hostlink_chardev.rs

@@ -6,6 +6,7 @@
 // stream; whoever opens the device decides how it reaches the host.
 
 use core::cell::UnsafeCell;
+use core::sync::atomic::{AtomicBool, Ordering};
 use core::ffi::{c_int, c_void};
 
 use alloc::vec::Vec;
@@ -30,6 +31,7 @@ impl Hostlink {
         if unsafe { frida_kmod_link_open() } != 0 {
             return Err(());
         }
+        LINK_OPEN.store(true, Ordering::Release);
 
         Ok(Hostlink {
             state: UnsafeCell::new(Inner {
@@ -97,25 +99,55 @@ impl Hostlink {
 
 impl Drop for Hostlink {
     fn drop(&mut self) {
+        LINK_OPEN.store(false, Ordering::Release);
         unsafe { frida_kmod_link_close() };
     }
 }
 
-#[cfg(feature = "xnu-kext")]
 pub fn a_turn_is_wanted() -> bool {
-    unsafe { frida_kmod_link_pending() }
+    LINK_OPEN.load(Ordering::Acquire) && unsafe { frida_kmod_link_pending() }
 }
 
+static LINK_OPEN: AtomicBool = AtomicBool::new(false);
+
 fn recv_nonblocking(dst: &mut [u8]) -> usize {
     let n = unsafe { frida_kmod_link_recv(dst.as_mut_ptr() as *mut c_void, dst.len()) };
     if n <= 0 { 0 } else { n as usize }
 }
 
+#[cfg(feature = "xnu-kext")]
 unsafe extern "C" {
     fn frida_kmod_link_open() -> c_int;
     fn frida_kmod_link_close();
     fn frida_kmod_link_send(data: *const c_void, size: usize) -> c_int;
     fn frida_kmod_link_recv(data: *mut c_void, size: usize) -> isize;
-    #[cfg(feature = "xnu-kext")]
+    fn frida_kmod_link_pending() -> bool;
+}
+
+#[cfg(all(feature = "linux-injected", not(target_arch = "x86")))]
+unsafe extern "C" {
+    #[link_name = "_frida_kmod_link_open"]
+    static frida_kmod_link_open: unsafe extern "C" fn() -> c_int;
+    #[link_name = "_frida_kmod_link_close"]
+    static frida_kmod_link_close: unsafe extern "C" fn();
+    #[link_name = "_frida_kmod_link_send"]
+    static frida_kmod_link_send: unsafe extern "C" fn(*const c_void, usize) -> c_int;
+    #[link_name = "_frida_kmod_link_recv"]
+    static frida_kmod_link_recv: unsafe extern "C" fn(*mut c_void, usize) -> isize;
+    #[link_name = "_frida_kmod_link_pending"]
+    static frida_kmod_link_pending: unsafe extern "C" fn() -> bool;
+}
+
+#[cfg(all(feature = "linux-injected", target_arch = "x86"))]
+unsafe extern "C" {
+    #[link_name = "frida_k_frida_kmod_link_open"]
+    fn frida_kmod_link_open() -> c_int;
+    #[link_name = "frida_k_frida_kmod_link_close"]
+    fn frida_kmod_link_close();
+    #[link_name = "frida_k_frida_kmod_link_send"]
+    fn frida_kmod_link_send(data: *const c_void, size: usize) -> c_int;
+    #[link_name = "frida_k_frida_kmod_link_recv"]
+    fn frida_kmod_link_recv(data: *mut c_void, size: usize) -> isize;
+    #[link_name = "frida_k_frida_kmod_link_pending"]
     fn frida_kmod_link_pending() -> bool;
 }

src/barebone/agent/src/kernel.rs

@@ -1,14 +1,8 @@
 // Facade over the host kernel the agent is running inside of. Both backends
 // expose the same primitives, so the rest of the agent never names a specific
 // kernel.
-//
-// A kernel is reached from the outside -- a remote stub injects the agent and
-// patches the addresses it needs into .kernel_addrs -- or from the inside,
-// where the agent is a kernel module whose glue is an ordinary C translation
-// unit compiled against the target kernel's headers. Linux is reached either
-// way; the others only from the outside.
 
-#[cfg(any(feature = "linux", feature = "linux-injected"))]
+#[cfg(feature = "linux-injected")]
 pub use crate::linux::*;
 #[cfg(feature = "win9x")]
 pub use crate::win9x::*;

src/barebone/agent/src/lib.rs

@@ -4,16 +4,12 @@
     feature = "xnu-core",
     feature = "win9x",
     feature = "winnt",
-    feature = "linux",
     feature = "linux-injected"
 )))]
 compile_error!(
-    "pick a flavor: --features xnu, --features win9x, --features winnt, --features linux or --features linux-injected"
+    "pick a flavor: --features xnu, --features win9x, --features winnt or --features linux-injected"
 );
 
-#[cfg(all(feature = "blob", feature = "linux"))]
-compile_error!("pick one flavour, not both");
-
 use alloc::boxed::Box;
 use alloc::collections::BTreeMap;
 use alloc::format;
@@ -58,17 +54,15 @@ mod ring;
 
 pub mod kernel;
 
-#[cfg(feature = "linux")]
-mod gum_linux;
 #[cfg(feature = "linux-injected")]
 mod gum_btf;
-#[cfg(any(feature = "linux", feature = "linux-injected"))]
+#[cfg(feature = "linux-injected")]
 mod gum_modules;
-#[cfg(any(feature = "linux", feature = "xnu-kext"))]
+#[cfg(any(feature = "linux-injected", feature = "xnu-kext"))]
 mod hostlink_chardev;
 #[cfg(any(feature = "linux-injected", feature = "xnu-core"))]
 mod heap;
-#[cfg(any(feature = "linux", feature = "linux-injected"))]
+#[cfg(feature = "linux-injected")]
 mod linux;
 
 #[cfg(any(feature = "win9x", feature = "winnt"))]
@@ -269,6 +263,19 @@ mod entrypoint_blob {
         unsafe { enter(config_data, config_size) }
     }
 
+    #[cfg(feature = "linux-injected")]
+    #[unsafe(no_mangle)]
+    pub extern "C" fn frida_agent_wake() {
+        nudge_the_loop(ptr::addr_of!(glib::WAKEUP_TOKEN) as *const u8);
+    }
+
+    #[cfg(feature = "linux-injected")]
+    #[unsafe(no_mangle)]
+    pub extern "C" fn frida_agent_stop() {
+        STOP_REQUESTED.store(true, Ordering::Release);
+        frida_agent_wake();
+    }
+
     unsafe fn enter(config_data: *const u8, config_size: usize) {
         unsafe {
             #[cfg(any(feature = "win9x", feature = "winnt", feature = "linux-injected", feature = "xnu-core"))]
@@ -312,6 +319,12 @@ mod entrypoint_blob {
             #[cfg(feature = "linux-injected")]
             crate::gum_btf::publish();
 
+            #[cfg(all(feature = "linux-injected", any(target_arch = "x86_64", target_arch = "aarch64")))]
+            {
+                kernel::learn_when_to_sleep();
+                kernel::open_control_channel();
+            }
+
             let wake_token = ptr::addr_of_mut!(glib::WAKEUP_TOKEN) as *const u8;
             // Install this before the loop waits the first time, thus a copy placed in a process later
             // can receive a handle to the same event.
@@ -342,6 +355,10 @@ mod entrypoint_blob {
                     linux::hostlink_pipe_goldfish::Hostlink::init(&path, Some(on_frame_from_host), wake_token)
                         .unwrap(),
                 ),
+                #[cfg(feature = "linux-injected")]
+                TransportConfig::CharDevice => Transport::CharDevice(
+                    hostlink_chardev::Hostlink::init(Some(on_frame_from_host)).unwrap(),
+                ),
                 #[cfg(feature = "xnu-core")]
                 TransportConfig::Vsock { host_port } => Transport::Vsock(
                     hostlink_vsock::Hostlink::init(host_port, Some(on_frame_from_host), wake_token)
@@ -361,6 +378,9 @@ mod entrypoint_blob {
 
             destroy_all_scripts(context);
 
+            #[cfg(all(feature = "linux-injected", any(target_arch = "x86_64", target_arch = "aarch64")))]
+            kernel::close_control_channel();
+
             #[cfg(any(feature = "win9x", feature = "winnt", feature = "linux-injected",
                 feature = "xnu-core"))]
             {
@@ -384,6 +404,8 @@ mod entrypoint_blob {
 
             (&raw mut frida_agent_left).write_volatile(1);
 
+            #[cfg(feature = "linux-injected")]
+            kernel::hand_back_to_the_loader();
         }
     }
 
@@ -475,6 +497,12 @@ mod entrypoint_blob {
                     #[cfg(not(feature = "linux-injected"))]
                     { panic!("pipe-goldfish is linux-injected's") }
                 }
+                6 => {
+                    #[cfg(feature = "linux-injected")]
+                    { TransportConfig::CharDevice }
+                    #[cfg(not(feature = "linux-injected"))]
+                    { panic!("the character device is linux-injected's") }
+                }
                 _ => panic!("Unsupported transport kind: {}", transport_kind),
             };
             g_variant_unref(transport_cfg_inner);
@@ -628,7 +656,7 @@ mod entrypoint_xnu_kext {
             kernel::wake(ptr::addr_of!(glib::WAKEUP_TOKEN) as *const u8);
             kernel::wait(
                 ptr::addr_of!(glib::WAKEUP_TOKEN) as *const u8,
-                Some(10_000),
+                None,
                 &mut || WORKER_EXITED.load(Ordering::Acquire),
             );
         }
@@ -678,94 +706,6 @@ mod entrypoint_xnu_kext {
     }
 }
 
-#[cfg(feature = "linux")]
-mod entrypoint_linux {
-    use super::*;
-    use core::ffi::c_int;
-    use core::sync::atomic::AtomicBool;
-
-    static STOP_REQUESTED: AtomicBool = AtomicBool::new(false);
-    static WORKER_EXITED: AtomicBool = AtomicBool::new(false);
-
-    /// Called from the module's init path. Everything interesting happens on our
-    /// own kernel thread: `insmod` runs with the module mutex held, and bringing
-    /// up Gum plus connecting to the host both block.
-    #[unsafe(no_mangle)]
-    pub extern "C" fn frida_agent_start(own_base: u64, own_size: u64) -> c_int {
-        kernel::log("frida: agent starting\n\0");
-
-        unsafe {
-            OWN_RANGE = GumMemoryRange {
-                base_address: own_base,
-                size: own_size as gsize,
-            };
-        }
-
-        if kernel::spawn_thread(worker, ptr::null_mut()) != 0 {
-            kernel::log("frida: failed to spawn worker\n\0");
-            return -1;
-        }
-
-        0
-    }
-
-    /// Called from the module's exit path, which must not return until the worker
-    /// is off our text — the module's pages go away right after.
-    #[unsafe(no_mangle)]
-    pub extern "C" fn frida_agent_stop() {
-        kernel::log("frida: agent stopping\n\0");
-
-        STOP_REQUESTED.store(true, Ordering::Release);
-
-        while !WORKER_EXITED.load(Ordering::Acquire) {
-            kernel::wake(ptr::addr_of!(glib::WAKEUP_TOKEN) as *const u8);
-            kernel::wait(
-                ptr::addr_of!(glib::WAKEUP_TOKEN) as *const u8,
-                Some(10_000),
-                &mut || WORKER_EXITED.load(Ordering::Acquire),
-            );
-        }
-
-        kernel::log("frida: agent stopped\n\0");
-    }
-
-    pub fn stop_requested() -> bool {
-        STOP_REQUESTED.load(Ordering::Acquire)
-    }
-
-    unsafe extern "C" fn worker(_parameter: *mut c_void, _wait_result: i32) {
-        unsafe {
-            kernel::log("frida: worker entry\n\0");
-
-            crate::run_constructors();
-            init_gum();
-
-            kernel::install_hooks();
-
-            match hostlink_chardev::Hostlink::init(Some(on_frame_from_host)) {
-                Ok(hostlink) => transport_set(Transport::CharDevice(hostlink)),
-                Err(_) => {
-                    kernel::log("frida: failed to connect to peer\n\0");
-                    WORKER_EXITED.store(true, Ordering::Release);
-                    return;
-                }
-            }
-
-            let main_context = adopt_js_context();
-
-            run_main_loop(main_context);
-
-            destroy_all_scripts(main_context);
-            transport_teardown();
-
-            WORKER_EXITED.store(true, Ordering::Release);
-            kernel::wake(ptr::addr_of!(glib::WAKEUP_TOKEN) as *const u8);
-        }
-    }
-}
-
-#[cfg(feature = "linux")]
-pub use entrypoint_linux::{frida_agent_start, frida_agent_stop};
 #[cfg(feature = "xnu-kext")]
 pub use entrypoint_xnu_kext::{frida_agent_start, frida_agent_stop, frida_agent_wake};
 #[cfg(feature = "blob")]
@@ -782,7 +722,7 @@ pub enum Transport {
     PipeVsock(linux::hostlink_pipe_vsock::Hostlink),
     #[cfg(feature = "linux-injected")]
     PipeGoldfish(linux::hostlink_pipe_goldfish::Hostlink),
-    #[cfg(any(feature = "linux", feature = "xnu-kext"))]
+    #[cfg(any(feature = "linux-injected", feature = "xnu-kext"))]
     CharDevice(hostlink_chardev::Hostlink),
 }
 
@@ -799,7 +739,7 @@ impl Transport {
             Transport::PipeVsock(h) => h.send(payload),
             #[cfg(feature = "linux-injected")]
             Transport::PipeGoldfish(h) => h.send(payload),
-            #[cfg(any(feature = "linux", feature = "xnu-kext"))]
+            #[cfg(any(feature = "linux-injected", feature = "xnu-kext"))]
             Transport::CharDevice(h) => h.send(payload),
         }
     }
@@ -816,7 +756,7 @@ impl Transport {
             Transport::PipeVsock(h) => h.shutdown(),
             #[cfg(feature = "linux-injected")]
             Transport::PipeGoldfish(h) => h.shutdown(),
-            #[cfg(any(feature = "linux", feature = "xnu-kext"))]
+            #[cfg(any(feature = "linux-injected", feature = "xnu-kext"))]
             Transport::CharDevice(_) => {}
         }
     }
@@ -833,7 +773,7 @@ impl Transport {
             Transport::PipeVsock(h) => h.process(),
             #[cfg(feature = "linux-injected")]
             Transport::PipeGoldfish(h) => h.process(),
-            #[cfg(any(feature = "linux", feature = "xnu-kext"))]
+            #[cfg(any(feature = "linux-injected", feature = "xnu-kext"))]
             Transport::CharDevice(h) => h.process(),
         }
     }
@@ -872,6 +812,8 @@ pub enum TransportConfig {
     PipeVsock { path: alloc::string::String },
     #[cfg(feature = "linux-injected")]
     PipeGoldfish { path: alloc::string::String },
+    #[cfg(feature = "linux-injected")]
+    CharDevice,
 }
 
 static mut TRANSPORT_DRIVER: *mut Transport = core::ptr::null_mut();
@@ -1262,11 +1204,6 @@ fn run_main_loop(main_context: *mut GMainContext) {
             #[cfg(not(any(feature = "blob", feature = "xnu-core")))]
             transport_get_unchecked().process();
 
-            #[cfg(feature = "linux")]
-            if entrypoint_linux::stop_requested() {
-                return;
-            }
-
             if stop_requested() {
                 return;
             }
@@ -1298,7 +1235,7 @@ fn kernel_half_has_work() -> bool {
         return true;
     }
 
-    #[cfg(feature = "xnu-kext")]
+    #[cfg(any(feature = "linux-injected", feature = "xnu-kext"))]
     if hostlink_chardev::a_turn_is_wanted() {
         return true;
     }
@@ -1337,6 +1274,11 @@ fn kernel_half_has_work() -> bool {
         return true;
     }
 
+    #[cfg(all(feature = "linux-injected", any(target_arch = "x86_64", target_arch = "aarch64")))]
+    if kernel::an_operation_waits() {
+        return true;
+    }
+
     kernel::injected_arenas()
         .iter()
         .any(|arena| kernel::holds_a_frame_from_target(*arena))
@@ -1370,6 +1312,9 @@ fn serve_the_kernel_half() {
     #[cfg(feature = "linux-injected")]
     kernel::report_what_the_copies_hit();
 
+    #[cfg(all(feature = "linux-injected", any(target_arch = "x86_64", target_arch = "aarch64")))]
+    kernel::carry_out_operations();
+
     relay_frames_from_targets();
 }
 
@@ -1528,7 +1473,7 @@ unsafe extern "C" fn on_script_unloaded(
 
 static UNLOADS_IN_FLIGHT: AtomicU32 = AtomicU32::new(0);
 
-#[cfg(any(feature = "linux", feature = "xnu-kext"))]
+#[cfg(feature = "xnu-kext")]
 fn transport_teardown() {
     unsafe {
         let driver = TRANSPORT_DRIVER;
@@ -2247,7 +2192,6 @@ fn text_or_empty(text: *const u8) -> *const core::ffi::c_char {
 }
 
 fn send_command_reply(request_id: u16, response: HandlerResponse) {
-
     unsafe {
         let message = g_variant_new(
             c"(yquv)".as_ptr(),
@@ -2433,9 +2377,14 @@ fn handle_stop() -> HandlerResponse {
         kernel::show_our_threads_again();
     }
 
-    #[cfg(feature = "blob")]
+    #[cfg(all(feature = "blob", not(feature = "linux-injected")))]
     STOP_REQUESTED.store(true, Ordering::Release);
 
+    #[cfg(feature = "linux-injected")]
+    if !kernel::loaded_from_the_inside() {
+        STOP_REQUESTED.store(true, Ordering::Release);
+    }
+
     HandlerResponse::success(unsafe { g_variant_new_uint32(0) })
 }
 

src/barebone/agent/src/libc.rs

@@ -50,7 +50,7 @@ pub extern "C" fn sysconf(_name: i32) -> isize {
 #[unsafe(no_mangle)]
 pub extern "C" fn __clear_cache(_start: *const u8, _end: *const u8) {}
 
-#[cfg(all(target_arch = "arm", not(any(feature = "linux", feature = "linux-injected"))))]
+#[cfg(all(target_arch = "arm", not(feature = "linux-injected")))]
 #[unsafe(no_mangle)]
 pub extern "C" fn __clear_cache(start: *const u8, end: *const u8) {
     unsafe {
@@ -68,7 +68,7 @@ pub extern "C" fn __clear_cache(start: *const u8, end: *const u8) {
     }
 }
 
-#[cfg(all(target_arch = "arm", any(feature = "linux", feature = "linux-injected")))]
+#[cfg(all(target_arch = "arm", feature = "linux-injected"))]
 #[unsafe(no_mangle)]
 pub extern "C" fn __clear_cache(start: *const u8, end: *const u8) {
     unsafe {

src/barebone/agent/src/linux/cloak.rs

@@ -0,0 +1,402 @@
+use alloc::collections::BTreeMap;
+use alloc::vec::Vec;
+use core::ffi::{c_int, c_void};
+use core::sync::atomic::{AtomicBool, Ordering};
+
+use super::layout::field_offset;
+use super::native;
+
+pub fn add_thread(tgid: i32, tid: i32) {
+    let fresh = Cloak::of(tgid);
+    registry().with(|cloaks| {
+        let threads = &mut cloaks.entry(tgid).or_insert(fresh).threads;
+        if !threads.contains(&tid) {
+            threads.push(tid);
+        }
+    });
+}
+
+pub fn remove_thread(tgid: i32, tid: i32) {
+    registry().with(|cloaks| {
+        if let Some(cloak) = cloaks.get_mut(&tgid) {
+            cloak.threads.retain(|t| *t != tid);
+        }
+    });
+}
+
+pub fn add_range(tgid: i32, base: usize, size: usize) {
+    let fresh = Cloak::of(tgid);
+    registry().with(|cloaks| {
+        cloaks.entry(tgid).or_insert(fresh).ranges.push((base, base + size));
+    });
+}
+
+pub fn remove_range(tgid: i32, base: usize, size: usize) {
+    registry().with(|cloaks| {
+        if let Some(cloak) = cloaks.get_mut(&tgid) {
+            if let Some(index) = cloak.ranges.iter().position(|r| *r == (base, base + size)) {
+                cloak.ranges.swap_remove(index);
+            }
+        }
+    });
+}
+
+pub fn add_fd(tgid: i32, fd: i32) {
+    let fresh = Cloak::of(tgid);
+    registry().with(|cloaks| {
+        let fds = &mut cloaks.entry(tgid).or_insert(fresh).fds;
+        if !fds.contains(&fd) {
+            fds.push(fd);
+        }
+    });
+}
+
+pub fn remove_fd(tgid: i32, fd: i32) {
+    registry().with(|cloaks| {
+        if let Some(cloak) = cloaks.get_mut(&tgid) {
+            cloak.fds.retain(|f| *f != fd);
+        }
+    });
+}
+
+pub fn install_hooks() {
+    let Some(layout) = describe_readers() else {
+        return;
+    };
+    unsafe { LAYOUT = Some(layout) };
+
+    let interceptor = unsafe { crate::bindings::gum_interceptor_obtain() };
+    unsafe {
+        crate::bindings::gum_interceptor_begin_transaction(interceptor);
+        replace(interceptor, _show_map, frida_cb_show_map as *mut c_void, &raw mut SHOW_MAP);
+        replace(interceptor, _show_smap, frida_cb_show_smap as *mut c_void, &raw mut SHOW_SMAP);
+        replace(interceptor, _proc_fill_cache, frida_cb_proc_fill_cache as *mut c_void,
+            &raw mut PROC_FILL_CACHE);
+        replace(interceptor, _proc_pid_status, frida_cb_proc_pid_status as *mut c_void,
+            &raw mut PROC_PID_STATUS);
+        replace(interceptor, _do_task_stat, frida_cb_do_task_stat as *mut c_void, &raw mut DO_TASK_STAT);
+        replace(interceptor, _release_task, frida_cb_release_task as *mut c_void, &raw mut RELEASE_TASK);
+        crate::bindings::gum_interceptor_end_transaction(interceptor);
+    }
+}
+
+pub fn remove_hooks() {
+    if unsafe { (*(&raw const LAYOUT)).is_none() } {
+        return;
+    }
+
+    let interceptor = unsafe { crate::bindings::gum_interceptor_obtain() };
+    unsafe {
+        crate::bindings::gum_interceptor_begin_transaction(interceptor);
+        for target in [_show_map, _show_smap, _proc_fill_cache, _proc_pid_status, _do_task_stat, _release_task] {
+            if target != 0 {
+                crate::bindings::gum_interceptor_revert(interceptor, target as *mut c_void);
+            }
+        }
+        crate::bindings::gum_interceptor_end_transaction(interceptor);
+    }
+
+    registry().with(|cloaks| cloaks.clear());
+}
+
+unsafe fn replace(interceptor: *mut crate::bindings::GumInterceptor, target: usize, replacement: *mut c_void,
+        original: *mut usize) {
+    if target == 0 {
+        return;
+    }
+
+    unsafe {
+        crate::bindings::gum_interceptor_replace(interceptor, target as *mut c_void, replacement,
+            original as *mut *mut c_void, core::ptr::null_mut());
+    }
+}
+
+unsafe extern "C" fn frida_cb_show_map(m: *mut c_void, vma: *mut c_void) -> c_int {
+    if hides_the_area(vma) {
+        return 0;
+    }
+
+    unsafe { call::<ShowFn>(SHOW_MAP)(m, vma) }
+}
+
+unsafe extern "C" fn frida_cb_show_smap(m: *mut c_void, vma: *mut c_void) -> c_int {
+    if hides_the_area(vma) {
+        return 0;
+    }
+
+    unsafe { call::<ShowFn>(SHOW_SMAP)(m, vma) }
+}
+
+unsafe extern "C" fn frida_cb_proc_fill_cache(file: *mut c_void, ctx: *mut c_void, name: *const u8, len: u32,
+        instantiate: usize, task: *mut c_void, ptr: *const c_void) -> bool {
+    if hides_the_entry(instantiate, task as usize, ptr as usize) {
+        return true;
+    }
+
+    unsafe { call::<FillCacheFn>(PROC_FILL_CACHE)(file, ctx, name, len, instantiate, task, ptr) }
+}
+
+unsafe extern "C" fn frida_cb_proc_pid_status(m: *mut c_void, ns: *mut c_void, pid: *mut c_void,
+        task: *mut c_void) -> c_int {
+    let layout = layout();
+    let start = read_word(m as usize + layout.seq_count);
+
+    let result = unsafe { call::<StatusFn>(PROC_PID_STATUS)(m, ns, pid, task) };
+
+    if let Some(hidden) = threads_hidden_from_the_reader(task as usize) {
+        let said = said_since(m as usize, start);
+        if let Some(at) = find(said, b"Threads:\t") {
+            take_from_the_number_at(m as usize, start + at + b"Threads:\t".len(), hidden);
+        }
+    }
+
+    result
+}
+
+unsafe extern "C" fn frida_cb_do_task_stat(m: *mut c_void, ns: *mut c_void, pid: *mut c_void,
+        task: *mut c_void, whole: c_int) -> c_int {
+    let layout = layout();
+    let start = read_word(m as usize + layout.seq_count);
+
+    let result = unsafe { call::<StatFn>(DO_TASK_STAT)(m, ns, pid, task, whole) };
+
+    if let Some(hidden) = threads_hidden_from_the_reader(task as usize) {
+        if let Some(at) = where_the_thread_count_is(said_since(m as usize, start)) {
+            take_from_the_number_at(m as usize, start + at, hidden);
+        }
+    }
+
+    result
+}
+
+unsafe extern "C" fn frida_cb_release_task(task: *mut c_void) {
+    let layout = layout();
+    let task = task as usize;
+    let tgid = read_i32(task + layout.task_tgid);
+
+    remove_thread(tgid, read_i32(task + layout.task_pid));
+    if read_word(task + layout.task_group_leader) == task {
+        registry().with(|cloaks| {
+            cloaks.remove(&tgid);
+        });
+    }
+
+    unsafe { call::<ReleaseFn>(RELEASE_TASK)(task as *mut c_void) };
+}
+
+fn hides_the_area(vma: *mut c_void) -> bool {
+    let layout = layout();
+    let vma = vma as usize;
+    let mm = read_word(vma + layout.vma_mm);
+    let start = read_word(vma + layout.vma_start);
+    let end = read_word(vma + layout.vma_end);
+
+    !the_reader_is_cloaked() && registry().with(|cloaks| {
+        cloaks.values().any(|cloak| {
+            cloak.mm == mm && cloak.ranges.iter().any(|(s, e)| start < *e && *s < end)
+        })
+    })
+}
+
+fn hides_the_entry(instantiate: usize, task: usize, ptr: usize) -> bool {
+    if the_reader_is_cloaked() {
+        return false;
+    }
+
+    let layout = layout();
+    if instantiate == unsafe { _proc_task_instantiate } {
+        let tid = read_i32(task + layout.task_pid);
+        return registry().with(|cloaks| cloaks.values().any(|cloak| cloak.threads.contains(&tid)));
+    }
+
+    if instantiate == unsafe { _proc_fd_instantiate } || instantiate == unsafe { _proc_fdinfo_instantiate } {
+        let tgid = read_i32(task + layout.task_tgid);
+        let fd = read_i32(ptr + layout.fd_data_fd);
+        return registry().with(|cloaks| cloaks.get(&tgid).is_some_and(|cloak| cloak.fds.contains(&fd)));
+    }
+
+    false
+}
+
+fn threads_hidden_from_the_reader(task: usize) -> Option<i32> {
+    if the_reader_is_cloaked() {
+        return None;
+    }
+
+    let tgid = read_i32(task + layout().task_tgid);
+    let hidden = registry().with(|cloaks| cloaks.get(&tgid).map_or(0, |cloak| cloak.threads.len()));
+
+    (hidden != 0).then_some(hidden as i32)
+}
+
+// A cloaked thread is one of ours, so let it see through the cloak; RASP threads in the same
+// process are not cloaked and keep getting the censored view.
+fn the_reader_is_cloaked() -> bool {
+    let tid = read_i32(native::current_task() as usize + layout().task_pid);
+
+    registry().with(|cloaks| cloaks.values().any(|cloak| cloak.threads.contains(&tid)))
+}
+
+fn said_since(m: usize, start: usize) -> &'static [u8] {
+    let layout = layout();
+    let count = read_word(m + layout.seq_count);
+    if start >= count || count >= read_word(m + layout.seq_size) {
+        return &[];
+    }
+
+    unsafe { core::slice::from_raw_parts((read_word(m + layout.seq_buf) + start) as *const u8, count - start) }
+}
+
+fn find(haystack: &[u8], needle: &[u8]) -> Option<usize> {
+    haystack.windows(needle.len()).position(|window| window == needle)
+}
+
+fn where_the_thread_count_is(said: &[u8]) -> Option<usize> {
+    let name_end = said.iter().rposition(|b| *b == b')')?;
+
+    said[name_end..]
+        .iter()
+        .enumerate()
+        .filter(|(_, b)| **b == b' ')
+        .nth(17)
+        .map(|(at, _)| name_end + at + 1)
+}
+
+fn take_from_the_number_at(m: usize, at: usize, hidden: i32) {
+    let layout = layout();
+    let buf = read_word(m + layout.seq_buf);
+    let count = read_word(m + layout.seq_count);
+    let said = unsafe { core::slice::from_raw_parts_mut(buf as *mut u8, count) };
+
+    let digits = said[at..].iter().take_while(|b| b.is_ascii_digit()).count();
+    let value = said[at..at + digits].iter().fold(0i32, |n, b| n * 10 + i32::from(b - b'0'));
+    let shown = if value > hidden { value - hidden } else { 1 };
+    let rendered = alloc::format!("{shown}");
+
+    said.copy_within(at + digits..count, at + rendered.len());
+    said[at..at + rendered.len()].copy_from_slice(rendered.as_bytes());
+    write_word(m + layout.seq_count, count - digits + rendered.len());
+}
+
+struct Cloak {
+    mm: usize,
+    threads: Vec<i32>,
+    ranges: Vec<(usize, usize)>,
+    fds: Vec<i32>,
+}
+
+impl Cloak {
+    fn of(tgid: i32) -> Self {
+        let mm = super::processes::HeldTask::with_id(tgid as u32)
+            .map_or(0, |task| read_word(task.address() + layout().task_mm));
+
+        Self { mm, threads: Vec::new(), ranges: Vec::new(), fds: Vec::new() }
+    }
+}
+
+struct Registry {
+    busy: AtomicBool,
+    cloaks: core::cell::UnsafeCell<BTreeMap<i32, Cloak>>,
+}
+
+unsafe impl Sync for Registry {}
+
+impl Registry {
+    fn with<R>(&self, action: impl FnOnce(&mut BTreeMap<i32, Cloak>) -> R) -> R {
+        native::disable_preemption();
+        while self.busy.swap(true, Ordering::Acquire) {
+            core::hint::spin_loop();
+        }
+        let result = action(unsafe { &mut *self.cloaks.get() });
+        self.busy.store(false, Ordering::Release);
+        native::enable_preemption();
+        result
+    }
+}
+
+fn registry() -> &'static Registry {
+    &REGISTRY
+}
+
+static REGISTRY: Registry = Registry {
+    busy: AtomicBool::new(false),
+    cloaks: core::cell::UnsafeCell::new(BTreeMap::new()),
+};
+
+struct ReaderLayout {
+    task_pid: usize,
+    task_tgid: usize,
+    task_mm: usize,
+    task_group_leader: usize,
+    vma_mm: usize,
+    vma_start: usize,
+    vma_end: usize,
+    seq_buf: usize,
+    seq_size: usize,
+    seq_count: usize,
+    fd_data_fd: usize,
+}
+
+fn describe_readers() -> Option<ReaderLayout> {
+    Some(ReaderLayout {
+        task_pid: field_offset("task_struct", "pid")?,
+        task_tgid: field_offset("task_struct", "tgid")?,
+        task_mm: field_offset("task_struct", "mm")?,
+        task_group_leader: field_offset("task_struct", "group_leader")?,
+        vma_mm: field_offset("vm_area_struct", "vm_mm")?,
+        vma_start: field_offset("vm_area_struct", "vm_start")?,
+        vma_end: field_offset("vm_area_struct", "vm_end")?,
+        seq_buf: field_offset("seq_file", "buf")?,
+        seq_size: field_offset("seq_file", "size")?,
+        seq_count: field_offset("seq_file", "count")?,
+        fd_data_fd: field_offset("fd_data", "fd")?,
+    })
+}
+
+fn layout() -> &'static ReaderLayout {
+    unsafe { (*(&raw const LAYOUT)).as_ref().unwrap() }
+}
+
+static mut LAYOUT: Option<ReaderLayout> = None;
+
+unsafe fn call<F: Copy>(original: usize) -> F {
+    unsafe { core::mem::transmute_copy(&original) }
+}
+
+fn read_word(address: usize) -> usize {
+    unsafe { (address as *const usize).read() }
+}
+
+fn write_word(address: usize, value: usize) {
+    unsafe { (address as *mut usize).write(value) };
+}
+
+fn read_i32(address: usize) -> i32 {
+    unsafe { (address as *const i32).read() }
+}
+
+type ShowFn = unsafe extern "C" fn(*mut c_void, *mut c_void) -> c_int;
+type FillCacheFn =
+    unsafe extern "C" fn(*mut c_void, *mut c_void, *const u8, u32, usize, *mut c_void, *const c_void) -> bool;
+type StatusFn = unsafe extern "C" fn(*mut c_void, *mut c_void, *mut c_void, *mut c_void) -> c_int;
+type StatFn = unsafe extern "C" fn(*mut c_void, *mut c_void, *mut c_void, *mut c_void, c_int) -> c_int;
+type ReleaseFn = unsafe extern "C" fn(*mut c_void);
+
+static mut SHOW_MAP: usize = 0;
+static mut SHOW_SMAP: usize = 0;
+static mut PROC_FILL_CACHE: usize = 0;
+static mut PROC_PID_STATUS: usize = 0;
+static mut DO_TASK_STAT: usize = 0;
+static mut RELEASE_TASK: usize = 0;
+
+unsafe extern "C" {
+    static _show_map: usize;
+    static _show_smap: usize;
+    static _proc_fill_cache: usize;
+    static _proc_pid_status: usize;
+    static _do_task_stat: usize;
+    static _release_task: usize;
+    static _proc_task_instantiate: usize;
+    static _proc_fd_instantiate: usize;
+    static _proc_fdinfo_instantiate: usize;
+}

src/barebone/agent/src/linux/control.rs

@@ -0,0 +1,362 @@
+use alloc::vec;
+use core::ffi::{c_int, c_long, c_void};
+use core::mem::size_of;
+use core::ptr;
+use core::sync::atomic::{AtomicBool, AtomicPtr, Ordering};
+
+use super::layout::field_offset;
+use super::{cloak, native, process_memory, user_threads};
+
+pub fn open() {
+    cloak::install_hooks();
+
+    let Some(arguments) = describe_arguments() else {
+        return;
+    };
+    let target = unsafe { SYS_PRCTL_TARGET };
+    if target == 0 {
+        return;
+    }
+    unsafe { ARGUMENTS = Some(arguments) };
+
+    let interceptor = unsafe { crate::bindings::gum_interceptor_obtain() };
+    unsafe {
+        crate::bindings::gum_interceptor_begin_transaction(interceptor);
+        crate::bindings::gum_interceptor_replace(interceptor, target as *mut c_void,
+            frida_cb_sys_prctl as *mut c_void, (&raw mut SYS_PRCTL) as *mut *mut c_void, ptr::null_mut());
+        crate::bindings::gum_interceptor_end_transaction(interceptor);
+    }
+}
+
+pub fn close() {
+    if unsafe { (*(&raw const ARGUMENTS)).is_some() } {
+        let interceptor = unsafe { crate::bindings::gum_interceptor_obtain() };
+        unsafe {
+            crate::bindings::gum_interceptor_begin_transaction(interceptor);
+            crate::bindings::gum_interceptor_revert(interceptor, SYS_PRCTL_TARGET as *mut c_void);
+            crate::bindings::gum_interceptor_end_transaction(interceptor);
+        }
+    }
+
+    cloak::remove_hooks();
+}
+
+pub fn an_operation_waits() -> bool {
+    !PENDING.load(Ordering::Acquire).is_null()
+}
+
+pub fn carry_out_operations() {
+    let operation = PENDING.swap(ptr::null_mut(), Ordering::AcqRel);
+    if operation.is_null() {
+        return;
+    }
+
+    unsafe {
+        (*operation).result = match (*operation).request {
+            Request::Allocate { id, size, protection } => process_memory::allocate(id, size, protection) as c_long,
+            Request::Release { id, address, size } => process_memory::release(id, address, size) as c_long,
+            Request::StartThread { id, entry, stack, argument, tls } => {
+                user_threads::start_in_process(id, entry, stack, argument, tls) as c_long
+            }
+        };
+        (*operation).done.store(true, Ordering::Release);
+    }
+
+    native::wake(wake_token());
+}
+
+unsafe extern "C" fn frida_cb_sys_prctl(regs: *const c_void) -> c_long {
+    let argument = |index: usize| read_word(regs as usize + unsafe { arguments().at[index] });
+
+    if argument(0) as c_int != CONTROL_MAGIC || argument(4) as u64 != CONTROL_TOKEN {
+        let original: unsafe extern "C" fn(*const c_void) -> c_long = unsafe { core::mem::transmute(SYS_PRCTL) };
+        return unsafe { original(regs) };
+    }
+
+    let op = argument(1) as u32;
+    let (first, second) = (argument(2), argument(3));
+
+    if op == OP_PING {
+        return CONTROL_MAGIC as c_long;
+    }
+
+    if op >= OP_ALLOC && !unsafe { _capable(CAP_SYS_ADMIN) } {
+        return -(EPERM as c_long);
+    }
+
+    let tgid = read_i32(native::current_task() as usize + unsafe { arguments().task_tgid });
+    match op {
+        OP_ADD_THREAD => cloak::add_thread(tgid, first as i32),
+        OP_REMOVE_THREAD => cloak::remove_thread(tgid, first as i32),
+        OP_ADD_RANGE => cloak::add_range(tgid, first, second),
+        OP_REMOVE_RANGE => cloak::remove_range(tgid, first, second),
+        OP_ADD_FD => cloak::add_fd(tgid, first as i32),
+        OP_REMOVE_FD => cloak::remove_fd(tgid, first as i32),
+        OP_ALLOC => return allocate(first),
+        OP_FREE => return release(first),
+        OP_WRITE => return write(first),
+        OP_READ => return read(first),
+        OP_SPAWN => return start_thread(first),
+        OP_CLOAK_THREAD => return cloak_thread(first),
+        OP_CLOAK_RANGE => return cloak_range(first),
+        _ => return -(EINVAL as c_long),
+    }
+
+    0
+}
+
+fn allocate(arguments: usize) -> c_long {
+    let Some(args) = arguments_at::<AllocArgs>(arguments) else {
+        return -(EFAULT as c_long);
+    };
+
+    hand_to_the_worker(Request::Allocate { id: args.pid, size: args.size as usize, protection: args.prot as usize })
+}
+
+fn release(arguments: usize) -> c_long {
+    let Some(args) = arguments_at::<FreeArgs>(arguments) else {
+        return -(EFAULT as c_long);
+    };
+
+    hand_to_the_worker(Request::Release { id: args.pid, address: args.address as usize, size: args.size as usize })
+}
+
+fn write(arguments: usize) -> c_long {
+    let Some(args) = arguments_at::<RwArgs>(arguments) else {
+        return -(EFAULT as c_long);
+    };
+
+    let mut bounce = vec![0u8; args.size as usize];
+    if !process_memory::read_own(args.buffer as usize, &mut bounce) {
+        return -(EFAULT as c_long);
+    }
+
+    process_memory::write(args.pid, args.address as usize, &bounce) as c_long
+}
+
+fn read(arguments: usize) -> c_long {
+    let Some(args) = arguments_at::<RwArgs>(arguments) else {
+        return -(EFAULT as c_long);
+    };
+
+    let mut bounce = vec![0u8; args.size as usize];
+    let transferred = process_memory::read(args.pid, args.address as usize, &mut bounce);
+
+    if !process_memory::write_own(args.buffer as usize, &bounce) {
+        return -(EFAULT as c_long);
+    }
+
+    transferred as c_long
+}
+
+fn start_thread(arguments: usize) -> c_long {
+    let Some(args) = arguments_at::<SpawnArgs>(arguments) else {
+        return -(EFAULT as c_long);
+    };
+
+    hand_to_the_worker(Request::StartThread {
+        id: args.pid,
+        entry: args.entry as usize,
+        stack: args.stack as usize,
+        argument: args.arg as usize,
+        tls: args.tls as usize,
+    })
+}
+
+fn cloak_thread(arguments: usize) -> c_long {
+    let Some(args) = arguments_at::<CloakThreadArgs>(arguments) else {
+        return -(EFAULT as c_long);
+    };
+
+    cloak::add_thread(args.tgid as i32, args.tid as i32);
+
+    0
+}
+
+fn cloak_range(arguments: usize) -> c_long {
+    let Some(args) = arguments_at::<CloakRangeArgs>(arguments) else {
+        return -(EFAULT as c_long);
+    };
+
+    cloak::add_range(args.tgid as i32, args.base as usize, args.size as usize);
+
+    0
+}
+
+fn arguments_at<T: Default>(address: usize) -> Option<T> {
+    let mut value = T::default();
+    let raw = unsafe { core::slice::from_raw_parts_mut(&mut value as *mut T as *mut u8, size_of::<T>()) };
+
+    process_memory::read_own(address, raw).then_some(value)
+}
+
+fn hand_to_the_worker(request: Request) -> c_long {
+    let token = wake_token();
+
+    while BUSY.swap(true, Ordering::Acquire) {
+        native::wait(token, None, &mut || !BUSY.load(Ordering::Acquire));
+    }
+
+    let mut operation = Operation { request, result: 0, done: AtomicBool::new(false) };
+    PENDING.store(&mut operation, Ordering::Release);
+    crate::nudge_the_loop(token);
+
+    while !operation.done.load(Ordering::Acquire) {
+        native::wait(token, None, &mut || operation.done.load(Ordering::Acquire));
+    }
+
+    BUSY.store(false, Ordering::Release);
+    native::wake(token);
+
+    operation.result
+}
+
+fn wake_token() -> *const u8 {
+    (&raw const crate::glib::WAKEUP_TOKEN) as *const u8
+}
+
+struct Operation {
+    request: Request,
+    result: c_long,
+    done: AtomicBool,
+}
+
+enum Request {
+    Allocate { id: u32, size: usize, protection: usize },
+    Release { id: u32, address: usize, size: usize },
+    StartThread { id: u32, entry: usize, stack: usize, argument: usize, tls: usize },
+}
+
+static PENDING: AtomicPtr<Operation> = AtomicPtr::new(ptr::null_mut());
+static BUSY: AtomicBool = AtomicBool::new(false);
+
+#[repr(C)]
+#[derive(Default)]
+struct AllocArgs {
+    size: u64,
+    pid: u32,
+    prot: u32,
+}
+
+#[repr(C)]
+#[derive(Default)]
+struct FreeArgs {
+    address: u64,
+    size: u64,
+    pid: u32,
+}
+
+#[repr(C)]
+#[derive(Default)]
+struct RwArgs {
+    address: u64,
+    buffer: u64,
+    size: u64,
+    pid: u32,
+}
+
+#[repr(C)]
+#[derive(Default)]
+struct SpawnArgs {
+    entry: u64,
+    stack: u64,
+    arg: u64,
+    tls: u64,
+    pid: u32,
+}
+
+#[repr(C)]
+#[derive(Default)]
+struct CloakThreadArgs {
+    tgid: u32,
+    tid: u32,
+}
+
+#[repr(C)]
+#[derive(Default)]
+struct CloakRangeArgs {
+    base: u64,
+    size: u64,
+    tgid: u32,
+}
+
+struct ArgumentLayout {
+    at: [usize; 5],
+    task_tgid: usize,
+}
+
+#[cfg(target_arch = "aarch64")]
+fn describe_arguments() -> Option<ArgumentLayout> {
+    let regs = field_offset("pt_regs", "regs")?;
+
+    Some(ArgumentLayout {
+        at: core::array::from_fn(|index| regs + index * size_of::<usize>()),
+        task_tgid: field_offset("task_struct", "tgid")?,
+    })
+}
+
+#[cfg(target_arch = "x86_64")]
+fn describe_arguments() -> Option<ArgumentLayout> {
+    Some(ArgumentLayout {
+        at: [
+            field_offset("pt_regs", "di")?,
+            field_offset("pt_regs", "si")?,
+            field_offset("pt_regs", "dx")?,
+            field_offset("pt_regs", "r10")?,
+            field_offset("pt_regs", "r8")?,
+        ],
+        task_tgid: field_offset("task_struct", "tgid")?,
+    })
+}
+
+unsafe fn arguments() -> &'static ArgumentLayout {
+    unsafe { (*(&raw const ARGUMENTS)).as_ref().unwrap() }
+}
+
+static mut ARGUMENTS: Option<ArgumentLayout> = None;
+static mut SYS_PRCTL: usize = 0;
+
+fn read_word(address: usize) -> usize {
+    unsafe { (address as *const usize).read() }
+}
+
+fn read_i32(address: usize) -> i32 {
+    unsafe { (address as *const i32).read() }
+}
+
+const CONTROL_MAGIC: c_int = 0x4652_4944;
+const CONTROL_TOKEN: u64 = 0x1d5f_9e6b_2c7a_4038;
+
+const OP_ADD_THREAD: u32 = 1;
+const OP_REMOVE_THREAD: u32 = 2;
+const OP_ADD_RANGE: u32 = 3;
+const OP_REMOVE_RANGE: u32 = 4;
+const OP_ADD_FD: u32 = 5;
+const OP_REMOVE_FD: u32 = 6;
+const OP_PING: u32 = 32;
+const OP_ALLOC: u32 = 64;
+const OP_FREE: u32 = 65;
+const OP_WRITE: u32 = 66;
+const OP_READ: u32 = 67;
+const OP_SPAWN: u32 = 68;
+const OP_CLOAK_THREAD: u32 = 69;
+const OP_CLOAK_RANGE: u32 = 70;
+
+const CAP_SYS_ADMIN: c_int = 21;
+
+const EPERM: c_int = 1;
+const EFAULT: c_int = 14;
+const EINVAL: c_int = 22;
+
+#[cfg(target_arch = "aarch64")]
+use self::___arm64_sys_prctl as SYS_PRCTL_TARGET;
+#[cfg(target_arch = "x86_64")]
+use self::___x64_sys_prctl as SYS_PRCTL_TARGET;
+
+unsafe extern "C" {
+    #[cfg(target_arch = "aarch64")]
+    static ___arm64_sys_prctl: usize;
+    #[cfg(target_arch = "x86_64")]
+    static ___x64_sys_prctl: usize;
+    static _capable: unsafe extern "C" fn(c_int) -> bool;
+}

src/barebone/agent/src/linux/injection.rs

@@ -1,6 +1,7 @@
 use core::ffi::{c_int, c_void};
 use core::ptr;
 
+use alloc::boxed::Box;
 use alloc::collections::BTreeMap;
 
 #[cfg(not(target_arch = "x86"))]
@@ -63,17 +64,28 @@ fn take_what_the_copy_opened(placed: &mut Placement) {
     let copy = COPY_TASK.load(core::sync::atomic::Ordering::Acquire) as *mut c_void;
     placed.says = native::take_the_file_of(copy, says_fd);
     placed.hears = native::take_the_file_of(copy, hears_fd);
-    listen_to_the_copy(placed.says);
+    listen_to_the_copy(placed);
 }
 
-fn listen_to_the_copy(says: *mut c_void) {
-    native::spawn_thread(carry_what_the_copy_says, says);
+fn listen_to_the_copy(placed: &Placement) {
+    let listener = Box::into_raw(Box::new(Listener { says: placed.says, arena: placed.arena }));
+    native::spawn_thread(carry_what_the_copy_says, listener as *mut c_void);
 }
 
 unsafe extern "C" fn carry_what_the_copy_says(argument: *mut c_void, _reason: i32) {
-    while native::wait_for_a_word(argument) {
+    let listener = unsafe { Box::from_raw(argument as *mut Listener) };
+
+    while native::wait_for_a_word(listener.says) {
         crate::nudge_the_loop(crate::glib::wakeup_token());
     }
+
+    Arena::at(listener.arena).gone();
+    crate::nudge_the_loop(crate::glib::wakeup_token());
+}
+
+struct Listener {
+    says: *mut c_void,
+    arena: usize,
 }
 
 struct Placement {
@@ -362,23 +374,25 @@ pub unsafe extern "C" fn frida_cb_user(argument: *mut c_void) -> c_int {
 
     COPY_TASK.store(native::current_task() as usize, core::sync::atomic::Ordering::Release);
 
+    hand_userland_to(bootstrap, hand_over_argument(stack, arena), arena);
+
+    0
+}
+
+pub(crate) fn hand_userland_to(pc: usize, stack: usize, first_argument: usize) {
     let Some(places) = describe_registers() else {
-        return 0;
+        return;
     };
     let registers = registers_of_this_task(&places);
 
-    let stack = hand_over_argument(stack, arena);
-
     unsafe {
-        ((registers + places.pc) as *mut usize).write(bootstrap & !THUMB_BIT);
+        ((registers + places.pc) as *mut usize).write(pc & !THUMB_BIT);
         ((registers + places.stack) as *mut usize).write(stack);
         ((registers + places.flags) as *mut usize).write(
-            USER_EXECUTION | (bootstrap & THUMB_BIT) * THUMB_STATE);
-        ((registers + places.first) as *mut usize).write(arena);
+            USER_EXECUTION | (pc & THUMB_BIT) * THUMB_STATE);
+        ((registers + places.first) as *mut usize).write(first_argument);
         write_selectors(registers, &places);
     }
-
-    0
 }
 
 // What a task starts with in userland is at the top of the kernel stack it was given.
@@ -390,7 +404,7 @@ fn registers_of_this_task(places: &Places) -> usize {
 fn registers_of_task(task: usize, places: &Places) -> usize {
     let stack = unsafe { ((task + places.stack_of_task) as *const usize).read() };
 
-    stack + STACK_SPAN - TOP_OF_STACK_PADDING - places.size
+    stack + places.stack_span - TOP_OF_STACK_PADDING - places.size
 }
 
 #[cfg(target_arch = "x86_64")]
@@ -684,6 +698,7 @@ unsafe fn write_selectors(registers: usize, _places: &Places) {
     unsafe {
         ((registers + 136) as *mut usize).write(0x33);
         ((registers + 160) as *mut usize).write(0x2b);
+        ((registers + 120) as *mut isize).write(-1);
     }
 }
 
@@ -733,6 +748,7 @@ struct Places {
     pc: usize,
     flags: usize,
     stack_of_task: usize,
+    stack_span: usize,
 }
 
 #[cfg(target_arch = "aarch64")]
@@ -744,6 +760,7 @@ fn describe_registers() -> Option<Places> {
         pc: field_offset("user_pt_regs", "pc")?,
         flags: field_offset("user_pt_regs", "pstate")?,
         stack_of_task: field_offset("task_struct", "stack")?,
+        stack_span: struct_size("thread_union").unwrap_or(STACK_SPAN),
     })
 }
 
@@ -756,6 +773,7 @@ fn describe_registers() -> Option<Places> {
         pc: 15 * 4,
         flags: 16 * 4,
         stack_of_task: 0,
+        stack_span: STACK_SPAN,
     })
 }
 
@@ -768,6 +786,7 @@ fn describe_registers() -> Option<Places> {
         pc: 48,
         flags: 56,
         stack_of_task: 0,
+        stack_span: STACK_SPAN,
     })
 }
 
@@ -780,6 +799,7 @@ fn describe_registers() -> Option<Places> {
         pc: 128,
         flags: 144,
         stack_of_task: 0,
+        stack_span: STACK_SPAN,
     })
 }
 
@@ -932,7 +952,7 @@ pub unsafe extern "C" fn frida_cb_exec(_data: *mut c_void, task: *mut c_void, _w
         ((program as usize + at) as *const *const u8).read()
     });
 
-    native::wake(task as *const u8);
+    native::wake(crate::glib::wakeup_token());
 }
 
 #[cfg(not(target_arch = "x86"))]
@@ -969,7 +989,7 @@ unsafe extern "C" fn frida_replaced_set_task_comm(task: *mut c_void, buf: *const
         if let Some(id) = id_of(task as usize) {
             native::send_signal(STOP, task as usize);
             super::spawn::note_pending_spawn(id, name.to_bytes());
-            native::wake(ptr::null());
+            native::wake(crate::glib::wakeup_token());
         }
     }
 
@@ -1062,9 +1082,8 @@ fn ask_it_to_leave(placed: &Placement) {
     arena.tell_it_to_go();
     wake_the_copy(placed);
 
-    let waited_on = &placed.arena as *const usize as *const u8;
     while !arena.has_gone() {
-        native::wait(waited_on, Some(LEAVING_GRACE_US), &mut || arena.has_gone());
+        native::wait(crate::glib::wakeup_token(), None, &mut || arena.has_gone());
     }
 }
 
@@ -1091,7 +1110,6 @@ fn take_back_what_it_was_given(placed: &Placement, memory: *mut c_void) {
     }
 }
 
-const LEAVING_GRACE_US: u64 = 100_000;
 
 pub fn tell_the_copy(id: u32) {
     let Some(placed) = (unsafe { placements() }).get(&id) else {
@@ -1290,7 +1308,6 @@ const EXEC_PROBE: unsafe extern "C" fn(*mut c_void, *mut c_void, c_int, *mut c_v
 const EXEC_PROBE: unsafe extern "C" fn(*mut c_void, *mut c_void, c_int, *mut c_void) =
     frida_cb_exec;
 
-
 #[cfg(target_arch = "x86")]
 const THREAD_APPEARED_PROBE: unsafe extern "C" fn(*mut c_void, *mut c_void, *mut c_void) =
     frida_kcb_thread_appeared;

src/barebone/agent/src/linux/kfault.rs

@@ -3,6 +3,8 @@ use core::ffi::{c_int, c_void};
 use core::ffi::c_long;
 #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
 use core::ptr;
+#[cfg(any(target_arch = "arm", target_arch = "aarch64"))]
+use core::sync::atomic::{AtomicBool, Ordering};
 
 use crate::bindings::{
     gum_barebone_handle_exception, GumCpuContext,
@@ -322,10 +324,15 @@ pub fn install() {
         _patch_text(addr.add(1) as *mut c_void, frida_arm_fixup as *const () as usize as u32);
         _patch_text(addr as *mut c_void, 0xe51ff004);
     }
+    ACTIVATED.store(true, Ordering::Release);
 }
 
 #[cfg(target_arch = "arm")]
 pub fn uninstall() {
+    if !ACTIVATED.swap(false, Ordering::AcqRel) {
+        return;
+    }
+
     let target = unsafe { _fixup_exception } as *mut u32;
     unsafe {
         _patch_text(target.add(1) as *mut c_void, PROLOGUE[1]);
@@ -437,6 +444,9 @@ unsafe extern "C" {
 #[cfg(target_arch = "aarch64")]
 const PROLOGUE: [u32; 4] = [0xd503233f, 0xa9be7bfd, 0x910003fd, 0xa90153f3];
 
+#[cfg(any(target_arch = "arm", target_arch = "aarch64"))]
+static ACTIVATED: AtomicBool = AtomicBool::new(false);
+
 #[cfg(target_arch = "aarch64")]
 #[unsafe(no_mangle)]
 static mut FRIDA_ARM64_ORIGINAL_CONT: u64 = 0;
@@ -452,7 +462,7 @@ pub fn install() {
 
     unsafe { FRIDA_ARM64_ORIGINAL_CONT = target.add(4) as u64 };
 
-    let jump = detour_to(frida_arm64_fixup as *const () as usize);
+    let jump = redirect_to(frida_arm64_fixup as *const () as usize);
     unsafe {
         let addr = target as *mut u32;
         _aarch64_insn_patch_text_nosync(addr.add(2) as *mut c_void, jump[2]);
@@ -460,10 +470,15 @@ pub fn install() {
         _aarch64_insn_patch_text_nosync(addr.add(1) as *mut c_void, jump[1]);
         _aarch64_insn_patch_text_nosync(addr as *mut c_void, jump[0]);
     }
+    ACTIVATED.store(true, Ordering::Release);
 }
 
 #[cfg(target_arch = "aarch64")]
 pub fn uninstall() {
+    if !ACTIVATED.swap(false, Ordering::AcqRel) {
+        return;
+    }
+
     let target = unsafe { _fixup_exception } as *mut u32;
     unsafe {
         for slot in 0..4 {
@@ -473,7 +488,7 @@ pub fn uninstall() {
 }
 
 #[cfg(target_arch = "aarch64")]
-fn detour_to(destination: usize) -> [u32; 4] {
+fn redirect_to(destination: usize) -> [u32; 4] {
     [0x58000050, 0xd61f0200, destination as u32, (destination >> 32) as u32]
 }
 
@@ -542,7 +557,7 @@ unsafe fn recovered(regs: *mut c_void) -> bool {
 extern "C" fn frida_arm64_park() -> ! {
     static mut PARK_TOKEN: u8 = 0;
     loop {
-        crate::kernel::wait(core::ptr::addr_of!(PARK_TOKEN), Some(1_000_000), &mut || false);
+        crate::kernel::wait(core::ptr::addr_of!(PARK_TOKEN), None, &mut || false);
     }
 }
 

src/barebone/agent/src/linux/kmod.rs

@@ -1,357 +0,0 @@
-// The agent as a kernel module: every primitive is an ordinary call into
-// linux/frida-kmod.c, compiled by kbuild against the target kernel's headers,
-// which is what keeps us out of the business of guessing struct layouts,
-// config-dependent macros and per-version symbol names.
-
-use core::ffi::{CStr, c_char, c_int, c_void};
-
-use alloc::string::String;
-use alloc::vec::Vec;
-
-use super::{LoadedModule, ModuleEvent};
-use crate::kernel::ThreadEntry;
-
-pub fn log(msg: &str) {
-    unsafe { frida_kmod_log(msg.as_ptr() as *const c_char) };
-}
-
-pub fn panic(msg: &str) -> ! {
-    unsafe { frida_kmod_panic(msg.as_ptr() as *const c_char) }
-}
-
-pub fn run_when_ready(action: fn()) {
-    action();
-}
-
-pub fn spawn_thread(entry: ThreadEntry, parameter: *mut c_void) -> isize {
-    unsafe { frida_kmod_spawn_thread(entry, parameter) as isize }
-}
-
-pub fn install_hooks() {
-    unsafe { frida_kmod_install_hooks() };
-}
-
-pub fn alloc(size: usize) -> *mut u8 {
-    unsafe { frida_kmod_alloc(size) }
-}
-
-pub fn free(ptr: *mut u8, size: usize) {
-    unsafe { frida_kmod_free(ptr, size) };
-}
-
-pub fn alloc_code(size: usize) -> *mut u8 {
-    unsafe { frida_kmod_alloc_code(size) }
-}
-
-pub fn free_code(ptr: *mut u8, size: usize) {
-    unsafe { frida_kmod_free_code(ptr, size) };
-}
-
-pub fn page_size() -> usize {
-    crate::gum::page_size_the_kernel_runs_with()
-}
-
-pub fn alloc_heap(size: usize) -> *mut u8 {
-    alloc(size)
-}
-
-pub fn alloc_dma(size: usize) -> *mut u8 {
-    alloc_code(size)
-}
-
-pub fn free_dma(ptr: *mut u8, size: usize) {
-    free_code(ptr, size);
-}
-
-pub fn own_range() -> (u64, u64) {
-    let mut base: u64 = 0;
-    let mut size: u64 = 0;
-    unsafe { frida_kmod_own_range(&mut base, &mut size) };
-    (base, size)
-}
-
-// The sequence number is taken before `check` runs, so a wakeup racing with the
-// condition we are about to observe still makes the commit return immediately.
-pub fn wait(_token: *const u8, timeout_us: Option<u64>, check: &mut dyn FnMut() -> bool) {
-    let seq = unsafe { frida_kmod_wait_prepare() };
-
-    if check() {
-        return;
-    }
-
-    unsafe { frida_kmod_wait_commit(seq, timeout_us.map_or(-1, |us| us as i64)) };
-}
-
-pub fn poke_loop_wakeup() {}
-
-pub fn wake(_token: *const u8) {
-    unsafe { frida_kmod_wake() };
-}
-
-/// Gives the scheduler a chance to run something else. A kernel thread that loops
-/// without doing this is a hard lockup waiting to happen.
-pub fn yield_now() {
-    unsafe { frida_kmod_yield() };
-}
-
-pub fn monotonic_micros() -> i64 {
-    unsafe { frida_kmod_monotonic_micros() }
-}
-
-pub fn wall_clock_micros() -> (u32, u32) {
-    let mut secs: u32 = 0;
-    let mut micros: u32 = 0;
-    unsafe { frida_kmod_wall_clock_micros(&mut secs, &mut micros) };
-    (secs, micros)
-}
-
-pub fn current_process_id() -> u32 {
-    KERNEL_PROCESS
-}
-
-pub fn current_thread_id() -> u64 {
-    unsafe { frida_kmod_current_thread_id() }
-}
-
-pub fn protect(address: u64, size: usize, gum_prot: u32) -> bool {
-    unsafe { frida_kmod_protect(address, size, gum_prot) != 0 }
-}
-
-pub fn remap_writable(first_page: u64, n_pages: u32) -> *mut c_void {
-    unsafe { frida_kmod_remap_writable(first_page, n_pages) }
-}
-
-/// Tears down an alias handed out by [`remap_writable`] and flushes the icache
-/// over the range it aliased; the shim remembers which range that was.
-pub fn unmap_writable(mapping: *mut c_void) {
-    unsafe { frida_kmod_unmap_writable(mapping) };
-}
-
-pub fn get_kernel_base() -> u64 {
-    unsafe { frida_kmod_kernel_base() }
-}
-
-pub fn get_kernel_size() -> u64 {
-    unsafe { frida_kmod_kernel_size() }
-}
-
-pub fn enumerate_modules() -> Vec<LoadedModule> {
-    unsafe extern "C" fn on_module(
-        name: *const c_char,
-        version: *const c_char,
-        base: u64,
-        size: u64,
-        user_data: *mut c_void,
-    ) -> c_int {
-        unsafe {
-            let modules = &mut *(user_data as *mut Vec<LoadedModule>);
-            modules.push(LoadedModule {
-                name: String::from(CStr::from_ptr(name).to_str().unwrap_or("?")),
-                version: String::from(CStr::from_ptr(version).to_str().unwrap_or("")),
-                base,
-                size,
-            });
-            1
-        }
-    }
-
-    let mut modules: Vec<LoadedModule> = Vec::new();
-    unsafe {
-        frida_kmod_enumerate_modules(on_module, &mut modules as *mut _ as *mut c_void);
-    }
-    modules
-}
-
-pub fn watch_modules(on_event: fn(ModuleEvent, LoadedModule)) {
-    unsafe extern "C" fn on_module_event(
-        loaded: c_int,
-        name: *const c_char,
-        version: *const c_char,
-        base: u64,
-        size: u64,
-        user_data: *mut c_void,
-    ) {
-        unsafe {
-            let on_event: fn(ModuleEvent, LoadedModule) = core::mem::transmute(user_data);
-            on_event(
-                if loaded != 0 {
-                    ModuleEvent::Loaded
-                } else {
-                    ModuleEvent::Unloaded
-                },
-                LoadedModule {
-                    name: String::from(CStr::from_ptr(name).to_str().unwrap_or("?")),
-                    version: String::from(CStr::from_ptr(version).to_str().unwrap_or("")),
-                    base,
-                    size,
-                },
-            );
-        }
-    }
-
-    unsafe {
-        frida_kmod_watch_modules(on_module_event, on_event as *mut c_void);
-    }
-}
-
-pub fn unwatch_modules() {
-    unsafe { frida_kmod_unwatch_modules() };
-}
-
-pub fn enumerate_module_symbols(
-    base: u64,
-    on_symbol: &mut dyn FnMut(*const c_char, u64, u64, u8, bool) -> bool,
-) -> bool {
-    unsafe extern "C" fn on_module_symbol(
-        name: *const c_char,
-        address: u64,
-        size: u64,
-        kind: u8,
-        is_global: c_int,
-        user_data: *mut c_void,
-    ) -> c_int {
-        unsafe {
-            let callback =
-                &mut *(user_data as *mut &mut dyn FnMut(*const c_char, u64, u64, u8, bool) -> bool);
-            if callback(name, address, size, kind, is_global != 0) { 1 } else { 0 }
-        }
-    }
-
-    let mut boxed: &mut dyn FnMut(*const c_char, u64, u64, u8, bool) -> bool = on_symbol;
-    unsafe {
-        frida_kmod_enumerate_module_symbols(base, on_module_symbol, &mut boxed as *mut _ as *mut c_void) != 0
-    }
-}
-
-pub fn enumerate_module_exports(
-    base: u64,
-    on_export: &mut dyn FnMut(*const c_char, u64) -> bool,
-) -> bool {
-    unsafe extern "C" fn on_module_export(
-        name: *const c_char,
-        address: u64,
-        user_data: *mut c_void,
-    ) -> c_int {
-        unsafe {
-            let callback = &mut *(user_data as *mut &mut dyn FnMut(*const c_char, u64) -> bool);
-            if callback(name, address) { 1 } else { 0 }
-        }
-    }
-
-    let mut boxed: &mut dyn FnMut(*const c_char, u64) -> bool = on_export;
-    unsafe {
-        frida_kmod_enumerate_module_exports(base, on_module_export, &mut boxed as *mut _ as *mut c_void) != 0
-    }
-}
-
-pub fn find_symbol(name: &CStr) -> u64 {
-    unsafe { frida_kmod_find_symbol(name.as_ptr()) }
-}
-
-pub fn find_function(name: &CStr) -> u64 {
-    unsafe { frida_kmod_find_function(name.as_ptr()) }
-}
-
-/// Resolves `address` to `(name, symbol_address)`, or `None` when it falls
-/// outside every known symbol.
-pub fn symbol_from_address(address: u64) -> Option<(String, u64)> {
-    const NAME_CAPACITY: usize = 256;
-
-    let mut buffer = [0u8; NAME_CAPACITY];
-    let mut symbol_address: u64 = 0;
-    let found = unsafe {
-        frida_kmod_symbol_name_from_address(
-            address,
-            buffer.as_mut_ptr() as *mut c_char,
-            NAME_CAPACITY,
-            &mut symbol_address,
-        )
-    };
-    if found == 0 {
-        return None;
-    }
-
-    let name = CStr::from_bytes_until_nul(&buffer).unwrap();
-    Some((String::from(name.to_str().unwrap()), symbol_address))
-}
-
-/// Visits every symbol the running kernel knows about. Returns false when
-/// kallsyms is unavailable, which is the case on kernels built without
-/// CONFIG_KALLSYMS or when the shim could not resolve its entry points.
-pub fn enumerate_symbols(callback: &mut dyn FnMut(&CStr, u64) -> bool) -> bool {
-    unsafe extern "C" fn on_symbol(
-        name: *const c_char,
-        address: u64,
-        user_data: *mut c_void,
-    ) -> c_int {
-        unsafe {
-            let callback = &mut *(user_data as *mut &mut dyn FnMut(&CStr, u64) -> bool);
-            if callback(CStr::from_ptr(name), address) { 1 } else { 0 }
-        }
-    }
-
-    let mut boxed: &mut dyn FnMut(&CStr, u64) -> bool = callback;
-    unsafe { frida_kmod_enumerate_symbols(on_symbol, &mut boxed as *mut _ as *mut c_void) != 0 }
-}
-
-const KERNEL_PROCESS: u32 = 0;
-
-type FoundSymbolFunc =
-    unsafe extern "C" fn(name: *const c_char, address: u64, user_data: *mut c_void) -> c_int;
-type ModuleEventFunc = unsafe extern "C" fn(c_int, *const c_char, *const c_char, u64, u64, *mut c_void);
-type FoundModuleSymbolFunc =
-    unsafe extern "C" fn(*const c_char, u64, u64, u8, c_int, *mut c_void) -> c_int;
-type FoundModuleExportFunc = unsafe extern "C" fn(*const c_char, u64, *mut c_void) -> c_int;
-type FoundModuleFunc = unsafe extern "C" fn(
-    name: *const c_char,
-    version: *const c_char,
-    base: u64,
-    size: u64,
-    user_data: *mut c_void,
-) -> c_int;
-
-unsafe extern "C" {
-    fn frida_kmod_log(message: *const c_char);
-    fn frida_kmod_panic(message: *const c_char) -> !;
-    fn frida_kmod_spawn_thread(entry: ThreadEntry, parameter: *mut c_void) -> c_int;
-    fn frida_kmod_install_hooks();
-    fn frida_kmod_alloc(size: usize) -> *mut u8;
-    fn frida_kmod_free(ptr: *mut u8, size: usize);
-    fn frida_kmod_alloc_code(size: usize) -> *mut u8;
-    fn frida_kmod_free_code(ptr: *mut u8, size: usize);
-    fn frida_kmod_own_range(base: *mut u64, size: *mut u64);
-    fn frida_kmod_wait_prepare() -> u32;
-    fn frida_kmod_wait_commit(seq: u32, timeout_us: i64);
-    fn frida_kmod_wake();
-    fn frida_kmod_yield();
-    fn frida_kmod_monotonic_micros() -> i64;
-    fn frida_kmod_wall_clock_micros(secs: *mut u32, micros: *mut u32);
-    fn frida_kmod_current_thread_id() -> u64;
-    fn frida_kmod_protect(address: u64, size: usize, gum_prot: u32) -> c_int;
-    fn frida_kmod_remap_writable(first_page: u64, n_pages: u32) -> *mut c_void;
-    fn frida_kmod_unmap_writable(mapping: *mut c_void);
-    fn frida_kmod_kernel_base() -> u64;
-    fn frida_kmod_kernel_size() -> u64;
-    fn frida_kmod_enumerate_modules(func: FoundModuleFunc, user_data: *mut c_void);
-    fn frida_kmod_watch_modules(func: ModuleEventFunc, user_data: *mut c_void);
-    fn frida_kmod_unwatch_modules();
-    fn frida_kmod_enumerate_module_symbols(
-        base: u64,
-        func: FoundModuleSymbolFunc,
-        user_data: *mut c_void,
-    ) -> c_int;
-    fn frida_kmod_enumerate_module_exports(
-        base: u64,
-        func: FoundModuleExportFunc,
-        user_data: *mut c_void,
-    ) -> c_int;
-    fn frida_kmod_find_symbol(name: *const c_char) -> u64;
-    fn frida_kmod_find_function(name: *const c_char) -> u64;
-    fn frida_kmod_symbol_name_from_address(
-        address: u64,
-        buffer: *mut c_char,
-        size: usize,
-        symbol_address: *mut u64,
-    ) -> c_int;
-    fn frida_kmod_enumerate_symbols(func: FoundSymbolFunc, user_data: *mut c_void) -> c_int;
-}

src/barebone/agent/src/linux/layout.rs

@@ -1,13 +1,17 @@
 use alloc::collections::BTreeMap;
 use alloc::string::String;
 use alloc::vec::Vec;
-use core::ffi::c_void;
+use core::ffi::{CStr, c_char, c_void};
 
 pub fn field_offset(container: &str, field: &str) -> Option<usize> {
+    if let Some(told) = told_by_the_loader(container, field) {
+        return Some(told);
+    }
+
     let types = described_types()?;
     let described = types.named(container)?;
 
-    types.offset_of(&described, field).map(|bits| bits / 8)
+    types.offset_of_path(&described, field).map(|bits| bits / 8)
 }
 
 pub fn resolve_struct(container: &str) -> Option<usize> {
@@ -96,11 +100,43 @@ pub fn types_are_described() -> bool {
 }
 
 pub fn struct_size(container: &str) -> Option<usize> {
+    if let Some(told) = told_by_the_loader(container, "") {
+        return Some(told);
+    }
+
     let types = described_types()?;
 
     Some(types.named(container)?.size as usize)
 }
 
+fn told_by_the_loader(container: &str, field: &str) -> Option<usize> {
+    let mut entry = unsafe { _frida_kmod_layout } as *const ToldLayout;
+    if entry.is_null() {
+        return None;
+    }
+
+    unsafe {
+        while !(*entry).container.is_null() {
+            let told = &*entry;
+            if CStr::from_ptr(told.container).to_bytes() == container.as_bytes()
+                && CStr::from_ptr(told.field).to_bytes() == field.as_bytes()
+            {
+                return Some(told.value);
+            }
+            entry = entry.add(1);
+        }
+    }
+
+    None
+}
+
+#[repr(C)]
+struct ToldLayout {
+    container: *const c_char,
+    field: *const c_char,
+    value: usize,
+}
+
 struct DescribedTypes {
     types: &'static [u8],
     names: &'static [u8],
@@ -437,6 +473,25 @@ impl DescribedTypes {
         described.kind == KIND_ENUM || described.kind == KIND_WIDE_ENUM
     }
 
+    fn offset_of_path(&self, described: &Described, path: &str) -> Option<usize> {
+        let Some((member, rest)) = path.split_once('.') else {
+            return self.offset_of(described, path);
+        };
+
+        let (placement, type_id) = self.member_named(described, member)?;
+        let within = self.with_id(type_id)?;
+
+        Some(placement + self.offset_of_path(&within, rest)?)
+    }
+
+    fn member_named(&self, described: &Described, wanted: &str) -> Option<(usize, u32)> {
+        (0..described.members).find_map(|index| {
+            let member = described.body + (index * MEMBER_SIZE);
+            (self.name_of(word_at(self.types, member)) == wanted)
+                .then(|| (self.placement_of(described, member), word_at(self.types, member + 4)))
+        })
+    }
+
     // A structure the kernel lays out itself keeps its fields in a member with no name of its
     // own, so a field is looked for through those as well, and answers where it sits in the
     // structure the question was asked about.
@@ -619,4 +674,5 @@ const MAX_DEPTH: usize = 8;
 unsafe extern "C" {
     static ___start_BTF: *const c_void;
     static ___stop_BTF: *const c_void;
+    static _frida_kmod_layout: usize;
 }

src/barebone/agent/src/linux/mod.rs

@@ -1,10 +1,3 @@
-// Linux backend. The agent reaches the kernel in one of two ways, and the rest
-// of the agent cannot tell which: loaded from the inside as a kernel module,
-// where every primitive is a call into a shim kbuild compiled against the
-// target's headers; or injected from the outside like XNU and Windows, where
-// the host patches the addresses it needs into .kernel_addrs and the memory it
-// cannot touch is reached over the hostlink.
-
 use alloc::string::String;
 
 pub fn stack_headroom() -> usize {
@@ -20,8 +13,6 @@ pub(crate) const STACK_SPAN: usize = 16 * 1024;
 #[cfg(any(target_arch = "arm", target_arch = "x86"))]
 pub(crate) const STACK_SPAN: usize = 8 * 1024;
 
-#[cfg(feature = "linux")]
-mod kmod;
 #[cfg(feature = "linux-injected")]
 mod arena;
 #[cfg(feature = "linux-injected")]
@@ -52,9 +43,15 @@ mod threads;
 pub(crate) mod user;
 #[cfg(feature = "linux-injected")]
 mod processes;
+#[cfg(all(feature = "linux-injected", any(target_arch = "x86_64", target_arch = "aarch64")))]
+mod cloak;
+#[cfg(all(feature = "linux-injected", any(target_arch = "x86_64", target_arch = "aarch64")))]
+mod control;
+#[cfg(all(feature = "linux-injected", any(target_arch = "x86_64", target_arch = "aarch64")))]
+mod process_memory;
+#[cfg(all(feature = "linux-injected", any(target_arch = "x86_64", target_arch = "aarch64")))]
+mod user_threads;
 
-#[cfg(feature = "linux")]
-pub use self::kmod::*;
 #[cfg(feature = "linux-injected")]
 pub use self::injection::*;
 #[cfg(feature = "linux-injected")]
@@ -63,17 +60,23 @@ pub use self::relay::*;
 pub use self::facade::*;
 #[cfg(feature = "linux-injected")]
 pub use self::native::{
-    alloc_dma, free_dma, get_kernel_base, get_kernel_size, install_interrupt_handler, map_io,
-    map_pages,
-    mmio_interrupt, pci_interrupt, release_fault_reporter, release_interrupt, run_when_ready,
-    patch_text, set_kernel_base, set_protection, virt_to_phys,
+    alloc_dma, free_dma, get_kernel_base, get_kernel_size, hand_back_to_the_loader,
+    install_interrupt_handler, loaded_from_the_inside, map_io, map_pages, mmio_interrupt, pci_interrupt,
+    release_fault_reporter, release_interrupt, remap_writable, run_when_ready, patch_text,
+    set_kernel_base, set_protection, unmap_writable, virt_to_phys,
 };
+#[cfg(all(feature = "linux-injected", any(target_arch = "x86_64", target_arch = "aarch64")))]
+pub use self::native::learn_when_to_sleep;
 #[cfg(feature = "linux-injected")]
 pub use self::mapped::*;
 #[cfg(feature = "linux-injected")]
 pub use self::modules::*;
 #[cfg(feature = "linux-injected")]
 pub use self::processes::*;
+#[cfg(all(feature = "linux-injected", any(target_arch = "x86_64", target_arch = "aarch64")))]
+pub use self::control::{
+    an_operation_waits, carry_out_operations, close as close_control_channel, open as open_control_channel,
+};
 #[cfg(feature = "linux-injected")]
 pub use self::spawn::*;
 #[cfg(feature = "linux-injected")]

src/barebone/agent/src/linux/modules.rs

@@ -9,6 +9,7 @@ use crate::bindings::{g_free, gsize, gum_memory_read};
 use super::{LoadedModule, ModuleEvent};
 
 pub fn enumerate_modules() -> Vec<LoadedModule> {
+    let _held = ModuleListLock::take();
     let Some(layout) = module_layout() else {
         return Vec::new();
     };
@@ -38,6 +39,7 @@ pub fn enumerate_module_symbols(
     base: u64,
     on_symbol: &mut dyn FnMut(*const c_char, u64, u64, u8, bool) -> bool,
 ) -> bool {
+    let _held = ModuleListLock::take();
     let Some(layout) = module_layout() else {
         return false;
     };
@@ -87,6 +89,7 @@ pub fn enumerate_module_exports(
     base: u64,
     on_export: &mut dyn FnMut(*const c_char, u64) -> bool,
 ) -> bool {
+    let _held = ModuleListLock::take();
     let Some(layout) = module_layout() else {
         return false;
     };
@@ -152,6 +155,52 @@ fn relative_target(at: usize) -> usize {
     at.wrapping_add_signed(offset as isize)
 }
 
+struct ModuleListLock {
+    mutex: *mut c_void,
+}
+
+impl ModuleListLock {
+    fn take() -> Self {
+        let mutex = unsafe { _module_mutex } as *mut c_void;
+        let lockable = !mutex.is_null() && unsafe { _mutex_lock.is_some() && _mutex_unlock.is_some() };
+        if !lockable {
+            return Self { mutex: ptr::null_mut() };
+        }
+
+        unsafe { lock_mutex(mutex) };
+
+        Self { mutex }
+    }
+}
+
+impl Drop for ModuleListLock {
+    fn drop(&mut self) {
+        if !self.mutex.is_null() {
+            unsafe { unlock_mutex(self.mutex) };
+        }
+    }
+}
+
+#[cfg(not(target_arch = "x86"))]
+unsafe fn lock_mutex(mutex: *mut c_void) {
+    unsafe { _mutex_lock.unwrap()(mutex) };
+}
+
+#[cfg(not(target_arch = "x86"))]
+unsafe fn unlock_mutex(mutex: *mut c_void) {
+    unsafe { _mutex_unlock.unwrap()(mutex) };
+}
+
+#[cfg(target_arch = "x86")]
+unsafe fn lock_mutex(mutex: *mut c_void) {
+    unsafe { frida_k_mutex_lock(mutex) };
+}
+
+#[cfg(target_arch = "x86")]
+unsafe fn unlock_mutex(mutex: *mut c_void) {
+    unsafe { frida_k_mutex_unlock(mutex) };
+}
+
 pub fn watch_modules(on_event: fn(ModuleEvent, LoadedModule)) {
     if module_layout().is_none() {
         return;
@@ -778,10 +827,15 @@ unsafe extern "C" {
     ) -> c_int;
     fn frida_k_register_module_notifier(block: *mut c_void) -> c_int;
     fn frida_k_unregister_module_notifier(block: *mut c_void) -> c_int;
+    fn frida_k_mutex_lock(mutex: *mut c_void);
+    fn frida_k_mutex_unlock(mutex: *mut c_void);
 }
 
 unsafe extern "C" {
     static _modules: *const c_void;
+    static _module_mutex: *const c_void;
+    static _mutex_lock: Option<unsafe extern "C" fn(*mut c_void)>;
+    static _mutex_unlock: Option<unsafe extern "C" fn(*mut c_void)>;
     static _register_module_notifier: Option<unsafe extern "C" fn(*mut c_void) -> c_int>;
     #[allow(dead_code)]
     static _unregister_module_notifier: Option<unsafe extern "C" fn(*mut c_void) -> c_int>;

src/barebone/agent/src/linux/native.rs

@@ -1,14 +1,9 @@
-// The agent injected from the outside: the host resolves the kernel functions
-// below and patches their addresses into .kernel_addrs, the same way it does
-// for XNU and Windows. Everything the kernel would otherwise have to lend us --
-// its symbol table, its module list, the right to write to its pages -- comes
-// from the host instead, over the hostlink.
-
 use core::ffi::{c_char, c_int, c_uint, c_void};
-use core::ptr;
 use core::ptr::read_volatile;
 use core::sync::atomic::{AtomicBool, AtomicU64, Ordering};
-#[cfg(target_arch = "aarch64")]
+#[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+use core::sync::atomic::{AtomicI32, fence};
+#[cfg(not(target_arch = "arm"))]
 use core::sync::atomic::AtomicUsize;
 
 use crate::kernel::ThreadEntry;
@@ -47,6 +42,16 @@ pub fn release_fault_reporter() {
     super::kfault::uninstall();
 }
 
+pub fn loaded_from_the_inside() -> bool {
+    unsafe { _frida_kmod_exit_worker != 0 }
+}
+
+pub fn hand_back_to_the_loader() {
+    if loaded_from_the_inside() {
+        unsafe { frida_kmod_exit_worker() };
+    }
+}
+
 pub fn send_signal(signal: c_int, task: usize) {
     unsafe { _send_sig(signal, task as *mut c_void, 1) };
 }
@@ -152,6 +157,17 @@ pub(crate) fn kcfi_thunk(callback: usize, _type_ref: Option<usize>) -> usize {
     callback
 }
 
+#[cfg(target_arch = "aarch64")]
+pub(crate) fn kernel_thread_function(cache: &AtomicUsize, callback: KthreadFn) -> KthreadFn {
+    let ref_fn = unsafe { _kthread_worker_fn.map(|f| f as usize) };
+    unsafe { core::mem::transmute(kcfi_thunk_cached(cache, callback as usize, ref_fn)) }
+}
+
+#[cfg(target_arch = "x86_64")]
+pub(crate) fn kernel_thread_function(_cache: &AtomicUsize, callback: KthreadFn) -> KthreadFn {
+    callback
+}
+
 #[cfg(target_arch = "aarch64")]
 fn kcfi_thunk_cached(cache: &AtomicUsize, callback: usize, type_ref: Option<usize>) -> usize {
     let existing = cache.load(Ordering::Acquire);
@@ -212,28 +228,38 @@ pub fn spawn_thread(entry: ThreadEntry, parameter: *mut c_void) -> isize {
 
 #[cfg(target_arch = "x86")]
 pub fn alloc(size: usize) -> *mut u8 {
-    unsafe { frida_k_alloc(size, GFP_KERNEL as usize) }
+    unsafe { frida_k_alloc(size, allocation_flags() as usize) }
 }
 
 #[cfg(not(target_arch = "x86"))]
 pub fn alloc(size: usize) -> *mut u8 {
+    let flags = allocation_flags();
+
     unsafe {
         if let Some(f) = ___kmalloc_noprof {
-            f(size, GFP_KERNEL)
+            f(size, flags)
         } else if let Some(f) = _kmalloc {
-            f(size, GFP_KERNEL)
+            f(size, flags)
         } else {
-            ___kmalloc.unwrap()(size, GFP_KERNEL)
+            ___kmalloc.unwrap()(size, flags)
         }
     }
 }
 
+fn allocation_flags() -> u32 {
+    if can_sleep() { GFP_KERNEL } else { GFP_ATOMIC }
+}
+
 pub fn free(ptr: *mut u8, _size: usize) {
     unsafe { _kfree(ptr) };
 }
 
 pub fn alloc_heap(size: usize) -> *mut u8 {
     unsafe {
+        #[cfg(not(target_arch = "x86"))]
+        if let Some(f) = _vmalloc_noprof {
+            return f(size);
+        }
         if let Some(f) = _vmalloc {
             return f(size);
         }
@@ -242,7 +268,6 @@ pub fn alloc_heap(size: usize) -> *mut u8 {
     alloc(size)
 }
 
-
 #[cfg(target_arch = "x86")]
 pub fn alloc_code(size: usize) -> *mut u8 {
     unsafe {
@@ -253,7 +278,9 @@ pub fn alloc_code(size: usize) -> *mut u8 {
 #[cfg(not(target_arch = "x86"))]
 pub fn alloc_code(size: usize) -> *mut u8 {
     unsafe {
-        let code = if let Some(f) = _execmem_alloc {
+        let code = if let Some(f) = _execmem_alloc_rw {
+            f(EXECMEM_MODULE_TEXT, size)
+        } else if let Some(f) = _execmem_alloc {
             f(EXECMEM_MODULE_TEXT, size)
         } else {
             _module_alloc.unwrap()(size)
@@ -283,6 +310,29 @@ pub fn free_code(ptr: *mut u8, _size: usize) {
     }
 }
 
+#[cfg(not(target_arch = "x86"))]
+pub fn remap_writable(first_page: u64, n_pages: u32) -> Option<*mut c_void> {
+    let remap = unsafe { _frida_kmod_remap_writable }?;
+    let alias = unsafe { remap(first_page, n_pages) };
+
+    (!alias.is_null()).then_some(alias)
+}
+
+#[cfg(target_arch = "x86")]
+pub fn remap_writable(_first_page: u64, _n_pages: u32) -> Option<*mut c_void> {
+    None
+}
+
+#[cfg(not(target_arch = "x86"))]
+pub fn unmap_writable(alias: *mut c_void) -> bool {
+    unsafe { _frida_kmod_unmap_writable }.is_some_and(|unmap| unsafe { unmap(alias) })
+}
+
+#[cfg(target_arch = "x86")]
+pub fn unmap_writable(_alias: *mut c_void) -> bool {
+    false
+}
+
 pub fn page_size() -> usize {
     crate::gum::page_size_the_kernel_runs_with()
 }
@@ -392,13 +442,18 @@ pub fn free_dma(ptr: *mut u8, size: usize) {
     free(ptr, size);
 }
 
-pub fn wait(_token: *const u8, timeout_us: Option<u64>, check: &mut dyn FnMut() -> bool) {
+pub fn wait(token: *const u8, timeout_us: Option<u64>, check: &mut dyn FnMut() -> bool) {
+    if !can_sleep() {
+        spin_until(check, timeout_us);
+        return;
+    }
+
     let waiters = waiters();
 
     let mut waiter = [0usize; WAIT_ENTRY_WORDS];
-    let waiter = queue_this_thread(&mut waiter);
+    let waiter = queue_this_thread(&mut waiter, token);
 
-    unsafe { _prepare_to_wait_event(waiters, waiter, TASK_INTERRUPTIBLE) };
+    unsafe { _prepare_to_wait_event(waiters, waiter, TASK_IDLE) };
 
     if !check() {
         match timeout_us {
@@ -413,18 +468,19 @@ pub fn wait(_token: *const u8, timeout_us: Option<u64>, check: &mut dyn FnMut()
     unsafe { _finish_wait(waiters, waiter) };
 }
 
-pub fn wake(_token: *const u8) {
-    unsafe { ___wake_up(waiters(), TASK_NORMAL, WAKE_EVERY_WAITER, ptr::null_mut()) };
+pub fn wake(token: *const u8) {
+    unsafe { ___wake_up(waiters(), TASK_NORMAL, WAKE_EVERY_WAITER, token as *mut c_void) };
 }
 
 pub fn poke_loop_wakeup() {}
 
-fn queue_this_thread(entry: &mut [usize; WAIT_ENTRY_WORDS]) -> *mut c_void {
+fn queue_this_thread(entry: &mut [usize; WAIT_ENTRY_WORDS], token: *const u8) -> *mut c_void {
     let entry = entry.as_mut_ptr();
 
     unsafe {
         entry.add(WAIT_ENTRY_THREAD).write(current_task() as usize);
-        entry.add(WAIT_ENTRY_WAKE).write(_autoremove_wake_function as usize);
+        entry.add(WAIT_ENTRY_WAKE).write(waking_function());
+        entry.add(WAIT_ENTRY_TOKEN).write(token as usize);
 
         let queued = entry.add(WAIT_ENTRY_QUEUED);
         queued.write(queued as usize);
@@ -434,6 +490,37 @@ fn queue_this_thread(entry: &mut [usize; WAIT_ENTRY_WORDS]) -> *mut c_void {
     entry as *mut c_void
 }
 
+#[cfg(target_arch = "aarch64")]
+fn waking_function() -> usize {
+    let ref_fn = unsafe { _autoremove_wake_function } as usize;
+    kcfi_thunk_cached(&WAKE_THUNK, frida_cb_wake as *const () as usize, Some(ref_fn))
+}
+
+#[cfg(not(any(target_arch = "aarch64", target_arch = "x86")))]
+fn waking_function() -> usize {
+    frida_cb_wake as *const () as usize
+}
+
+#[cfg(target_arch = "x86")]
+fn waking_function() -> usize {
+    unsafe { _autoremove_wake_function as usize }
+}
+
+#[cfg(not(target_arch = "x86"))]
+unsafe extern "C" fn frida_cb_wake(entry: *mut c_void, mode: c_uint, flags: c_int, key: *mut c_void) -> c_int {
+    let awaited = unsafe { (entry as *const usize).add(WAIT_ENTRY_TOKEN).read() };
+    if key as usize != awaited {
+        return 0;
+    }
+
+    let wake_this_one: unsafe extern "C" fn(*mut c_void, c_uint, c_int, *mut c_void) -> c_int =
+        unsafe { core::mem::transmute(_autoremove_wake_function) };
+    unsafe { wake_this_one(entry, mode, flags, key) }
+}
+
+#[cfg(target_arch = "aarch64")]
+static WAKE_THUNK: AtomicUsize = AtomicUsize::new(0);
+
 fn waiters() -> *mut c_void {
     let queue = (&raw mut WAITERS) as *mut c_void;
 
@@ -492,7 +579,7 @@ fn pci_irq_vector() -> Option<unsafe extern "C" fn(*mut c_void, c_uint) -> c_int
 #[cfg(target_arch = "arm")]
 pub fn mmio_interrupt(mmio_base: u64) -> Option<u32> {
     let path = alloc::format!("/virtio_mmio@{:x}\0", mmio_base);
-    let node = unsafe { _of_find_node_opts_by_path(path.as_ptr() as *const c_char, ptr::null_mut()) };
+    let node = unsafe { _of_find_node_opts_by_path(path.as_ptr() as *const c_char, core::ptr::null_mut()) };
     if node.is_null() {
         return None;
     }
@@ -513,7 +600,240 @@ pub fn mmio_interrupt(_mmio_base: u64) -> Option<u32> {
 }
 
 pub fn yield_now() {
-    unsafe { _schedule() };
+    if can_sleep() {
+        unsafe { _schedule() };
+    } else {
+        core::hint::spin_loop();
+    }
+}
+
+#[cfg(any(target_arch = "aarch64", target_arch = "x86_64"))]
+pub fn learn_when_to_sleep() {
+    let above = |found: Option<usize>| found.map_or(usize::MAX, |at| at + 1);
+    RCU_DEPTH_AT.store(above(super::layout::field_offset("task_struct", "rcu_read_lock_nesting")), Ordering::Relaxed);
+    PREEMPT_COUNT_AT.store(above(where_the_preempt_count_is()), Ordering::Release);
+}
+
+#[cfg(any(target_arch = "aarch64", target_arch = "x86_64"))]
+fn can_sleep() -> bool {
+    let Some(places) = sleeping_places() else {
+        return true;
+    };
+    let rcu_depth = places.rcu_read_lock_nesting
+        .map_or(0, |at| unsafe { ((current_task() as usize + at) as *const i32).read_volatile() });
+
+    current_preempt_count(places.preempt_count) == 0 && !interrupts_masked() && rcu_depth == 0
+}
+
+#[cfg(not(any(target_arch = "aarch64", target_arch = "x86_64")))]
+fn can_sleep() -> bool {
+    true
+}
+
+#[cfg(any(target_arch = "aarch64", target_arch = "x86_64"))]
+fn sleeping_places() -> Option<SleepingPlaces> {
+    let found = |at: usize| (at != 0 && at != usize::MAX).then(|| at - 1);
+    Some(SleepingPlaces {
+        preempt_count: found(PREEMPT_COUNT_AT.load(Ordering::Acquire))?,
+        rcu_read_lock_nesting: found(RCU_DEPTH_AT.load(Ordering::Relaxed)),
+    })
+}
+
+#[cfg(target_arch = "aarch64")]
+fn where_the_preempt_count_is() -> Option<usize> {
+    super::layout::field_offset("task_struct", "thread_info.preempt_count")
+}
+
+#[cfg(target_arch = "x86_64")]
+fn where_the_preempt_count_is() -> Option<usize> {
+    let standalone = unsafe { ___preempt_count };
+    if standalone != 0 {
+        return Some(standalone);
+    }
+
+    hot_per_cpu_place("preempt_count")
+}
+
+#[cfg(target_arch = "aarch64")]
+fn current_preempt_count(at: usize) -> u32 {
+    unsafe { ((current_task() as usize + at) as *const u32).read_volatile() }
+}
+
+#[cfg(target_arch = "x86_64")]
+fn current_preempt_count(at: usize) -> u32 {
+    let count: u32;
+    unsafe { core::arch::asm!("mov {:e}, gs:[{}]", out(reg) count, in(reg) at, options(nostack, readonly)) };
+
+    count & !PREEMPT_NEED_RESCHED
+}
+
+#[cfg(target_arch = "aarch64")]
+fn interrupts_masked() -> bool {
+    let daif: usize;
+    unsafe { core::arch::asm!("mrs {}, daif", out(reg) daif, options(nomem, nostack)) };
+
+    daif & DAIF_IRQ_MASKED != 0
+}
+
+#[cfg(target_arch = "x86_64")]
+fn interrupts_masked() -> bool {
+    let flags: usize;
+    unsafe { core::arch::asm!("pushfq", "pop {}", out(reg) flags, options(nomem, preserves_flags)) };
+
+    flags & RFLAGS_IF == 0
+}
+
+#[cfg(any(target_arch = "aarch64", target_arch = "x86_64"))]
+struct SleepingPlaces {
+    preempt_count: usize,
+    rcu_read_lock_nesting: Option<usize>,
+}
+
+#[cfg(any(target_arch = "aarch64", target_arch = "x86_64"))]
+static PREEMPT_COUNT_AT: AtomicUsize = AtomicUsize::new(0);
+#[cfg(any(target_arch = "aarch64", target_arch = "x86_64"))]
+static RCU_DEPTH_AT: AtomicUsize = AtomicUsize::new(0);
+
+fn spin_until(check: &mut dyn FnMut() -> bool, timeout_us: Option<u64>) {
+    let deadline = timeout_us.map(|us| monotonic_micros() + us as i64);
+
+    while !check() && deadline.is_none_or(|deadline| monotonic_micros() < deadline) {
+        core::hint::spin_loop();
+    }
+}
+
+#[cfg(target_arch = "x86_64")]
+pub(crate) fn lock_spinlock(lock: usize) {
+    unsafe { __raw_spin_lock.unwrap()(lock as *mut c_void) };
+}
+
+#[cfg(target_arch = "x86_64")]
+pub(crate) fn unlock_spinlock(lock: usize) {
+    unsafe { __raw_spin_unlock.unwrap()(lock as *mut c_void) };
+}
+
+#[cfg(target_arch = "aarch64")]
+pub(crate) fn lock_spinlock(lock: usize) {
+    if let Some(lock_it) = unsafe { __raw_spin_lock } {
+        return unsafe { lock_it(lock as *mut c_void) };
+    }
+
+    disable_preemption();
+    if let Err(seen) = counter_at(lock).compare_exchange(0, SPIN_LOCKED, Ordering::Acquire, Ordering::Relaxed) {
+        unsafe { _queued_spin_lock_slowpath(lock as *mut c_void, seen as u32) };
+    }
+}
+
+#[cfg(target_arch = "aarch64")]
+pub(crate) fn unlock_spinlock(lock: usize) {
+    if let Some(unlock_it) = unsafe { __raw_spin_unlock } {
+        return unsafe { unlock_it(lock as *mut c_void) };
+    }
+
+    unsafe { core::sync::atomic::AtomicU8::from_ptr(lock as *mut u8) }.store(0, Ordering::Release);
+    enable_preemption();
+}
+
+#[cfg(target_arch = "aarch64")]
+pub(crate) fn read_lock_irqsave(lock: usize) -> usize {
+    let flags = mask_interrupts();
+    disable_preemption();
+    if counter_at(lock).fetch_add(READER_BIAS, Ordering::Acquire) & WRITER_MASK != 0 {
+        unsafe { _queued_read_lock_slowpath(lock as *mut c_void) };
+    }
+    flags
+}
+
+#[cfg(target_arch = "aarch64")]
+pub(crate) fn read_unlock_irqrestore(lock: usize, flags: usize) {
+    counter_at(lock).fetch_sub(READER_BIAS, Ordering::Release);
+    enable_preemption();
+    restore_interrupts(flags);
+}
+
+#[cfg(target_arch = "aarch64")]
+pub(crate) fn write_lock_irqsave(lock: usize) -> usize {
+    let flags = mask_interrupts();
+    disable_preemption();
+    if counter_at(lock).compare_exchange(0, WRITER_LOCKED, Ordering::Acquire, Ordering::Relaxed).is_err() {
+        unsafe { _queued_write_lock_slowpath(lock as *mut c_void) };
+    }
+    flags
+}
+
+#[cfg(target_arch = "aarch64")]
+pub(crate) fn write_unlock_irqrestore(lock: usize, flags: usize) {
+    unsafe { core::sync::atomic::AtomicU8::from_ptr(lock as *mut u8) }.store(0, Ordering::Release);
+    enable_preemption();
+    restore_interrupts(flags);
+}
+
+#[cfg(target_arch = "aarch64")]
+pub(crate) fn disable_preemption() {
+    let count = preempt_count();
+    unsafe { count.write_volatile(count.read_volatile() + 1) };
+    core::sync::atomic::compiler_fence(Ordering::SeqCst);
+}
+
+#[cfg(target_arch = "aarch64")]
+pub(crate) fn enable_preemption() {
+    core::sync::atomic::compiler_fence(Ordering::SeqCst);
+    let count = preempt_count();
+    unsafe { count.write_volatile(count.read_volatile() - 1) };
+}
+
+#[cfg(target_arch = "x86_64")]
+pub(crate) fn disable_preemption() {
+    let at = sleeping_places().unwrap().preempt_count;
+    unsafe { core::arch::asm!("add dword ptr gs:[{}], 1", in(reg) at, options(nostack)) };
+    core::sync::atomic::compiler_fence(Ordering::SeqCst);
+}
+
+#[cfg(target_arch = "x86_64")]
+pub(crate) fn enable_preemption() {
+    core::sync::atomic::compiler_fence(Ordering::SeqCst);
+    let at = sleeping_places().unwrap().preempt_count;
+    unsafe { core::arch::asm!("sub dword ptr gs:[{}], 1", in(reg) at, options(nostack)) };
+}
+
+#[cfg(target_arch = "aarch64")]
+fn preempt_count() -> *mut u32 {
+    let at = sleeping_places().unwrap().preempt_count;
+    (current_task() as usize + at) as *mut u32
+}
+
+#[cfg(target_arch = "aarch64")]
+fn mask_interrupts() -> usize {
+    let flags: usize;
+    unsafe {
+        core::arch::asm!("mrs {}, daif", "msr daifset, #3", out(reg) flags, options(nostack));
+    }
+    flags
+}
+
+#[cfg(target_arch = "aarch64")]
+fn restore_interrupts(flags: usize) {
+    unsafe { core::arch::asm!("msr daif, {}", in(reg) flags, options(nostack)) };
+}
+
+#[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+pub(crate) fn hold(counter: usize) {
+    counter_at(counter).fetch_add(1, Ordering::Relaxed);
+}
+
+#[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+pub(crate) fn let_go(counter: usize) -> bool {
+    if counter_at(counter).fetch_sub(1, Ordering::Release) != 1 {
+        return false;
+    }
+
+    fence(Ordering::Acquire);
+    true
+}
+
+#[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+fn counter_at(address: usize) -> &'static AtomicI32 {
+    unsafe { AtomicI32::from_ptr(address as *mut i32) }
 }
 
 pub fn monotonic_micros() -> i64 {
@@ -567,33 +887,64 @@ pub fn current_task() -> u64 {
     task
 }
 
-#[cfg(target_arch = "x86_64")]
+#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
 pub fn current_task() -> u64 {
-    let task: u64;
-    unsafe {
-        core::arch::asm!("mov {}, gs:[{}]", out(reg) task, in(reg) _current_task, options(nostack));
-    }
-    task
+    let place = per_cpu_place(unsafe { _current_task }, "current_task", &CURRENT_TASK_AT);
+
+    read_per_cpu_word(place) as u64
 }
 
 #[cfg(target_arch = "x86_64")]
 pub fn top_of_stack() -> usize {
-    let top: usize;
-    unsafe {
-        core::arch::asm!("mov {}, gs:[{}]", out(reg) top, in(reg) _cpu_current_top_of_stack, options(nostack));
+    let place = per_cpu_place(unsafe { _cpu_current_top_of_stack }, "top_of_stack", &TOP_OF_STACK_AT);
+
+    read_per_cpu_word(place)
+}
+
+#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+fn per_cpu_place(standalone: usize, hot_member: &str, cache: &AtomicUsize) -> usize {
+    let known = cache.load(Ordering::Relaxed);
+    if known != 0 {
+        return known;
     }
-    top
+
+    let place = if standalone != 0 { standalone } else { hot_per_cpu_place(hot_member).unwrap() };
+    cache.store(place, Ordering::Relaxed);
+
+    place
+}
+
+#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+fn hot_per_cpu_place(member: &str) -> Option<usize> {
+    let hot = unsafe { _pcpu_hot };
+    if hot == 0 {
+        return None;
+    }
+
+    Some(hot + super::layout::field_offset("pcpu_hot", member)?)
+}
+
+#[cfg(target_arch = "x86_64")]
+fn read_per_cpu_word(place: usize) -> usize {
+    let value: usize;
+    unsafe { core::arch::asm!("mov {}, gs:[{}]", out(reg) value, in(reg) place, options(nostack, readonly)) };
+
+    value
 }
 
 #[cfg(target_arch = "x86")]
-pub fn current_task() -> u64 {
-    let task: u32;
-    unsafe {
-        core::arch::asm!("mov {}, fs:[{}]", out(reg) task, in(reg) _current_task as u32, options(nostack));
-    }
-    task as u64
+fn read_per_cpu_word(place: usize) -> usize {
+    let value: usize;
+    unsafe { core::arch::asm!("mov {}, fs:[{}]", out(reg) value, in(reg) place, options(nostack, readonly)) };
+
+    value
 }
 
+#[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+static CURRENT_TASK_AT: AtomicUsize = AtomicUsize::new(0);
+#[cfg(target_arch = "x86_64")]
+static TOP_OF_STACK_AT: AtomicUsize = AtomicUsize::new(0);
+
 pub fn get_kernel_base() -> u64 {
     KERNEL_BASE.load(Ordering::Relaxed)
 }
@@ -814,22 +1165,38 @@ const KERNEL_PROCESS: u32 = 0;
 const JS_SAFE_THREAD_ID_MASK: u64 = (1 << 48) - 1;
 
 const GFP_KERNEL: u32 = 0xcc0;
+const GFP_ATOMIC: u32 = 0x820;
 const NUMA_NO_NODE: c_int = -1;
-const EXECMEM_MODULE_TEXT: u32 = 1;
+const EXECMEM_MODULE_TEXT: u32 = 0;
 const PAGE_SIZE: usize = 4096;
 #[cfg(target_arch = "aarch64")]
 const TCR_T1SZ_SHIFT: u64 = 16;
 #[cfg(target_arch = "aarch64")]
 const TCR_T1SZ_MASK: u64 = 0x3f;
-const TASK_INTERRUPTIBLE: c_uint = 1;
+const TASK_IDLE: c_uint = 0x402;
 const TASK_NORMAL: c_uint = 3;
 const WAKE_EVERY_WAITER: c_int = 0;
 const WAKE_SLACK_NS: u64 = 50_000;
+#[cfg(target_arch = "aarch64")]
+const SPIN_LOCKED: i32 = 1;
+#[cfg(target_arch = "aarch64")]
+const READER_BIAS: i32 = 1 << 9;
+#[cfg(target_arch = "aarch64")]
+const WRITER_LOCKED: i32 = 0xff;
+#[cfg(target_arch = "aarch64")]
+const WRITER_MASK: i32 = 0x1ff;
+#[cfg(target_arch = "aarch64")]
+const DAIF_IRQ_MASKED: usize = 1 << 7;
+#[cfg(target_arch = "x86_64")]
+const PREEMPT_NEED_RESCHED: u32 = 1 << 31;
+#[cfg(target_arch = "x86_64")]
+const RFLAGS_IF: usize = 1 << 9;
 const HRTIMER_MODE_REL: c_int = 1;
 const WAIT_ENTRY_WORDS: usize = 8;
 const WAIT_ENTRY_THREAD: usize = 1;
 const WAIT_ENTRY_WAKE: usize = 2;
 const WAIT_ENTRY_QUEUED: usize = 3;
+const WAIT_ENTRY_TOKEN: usize = WAIT_ENTRY_WORDS - 1;
 const WAIT_QUEUE_WORDS: usize = 16;
 const LOCK_KEY_WORDS: usize = 8;
 #[cfg(target_arch = "aarch64")]
@@ -864,7 +1231,33 @@ struct Timespec64 {
     tv_nsec: i64,
 }
 
-type KthreadFn = unsafe extern "C" fn(data: *mut c_void) -> c_int;
+pub(crate) type KthreadFn = unsafe extern "C" fn(data: *mut c_void) -> c_int;
+
+#[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+unsafe extern "C" {
+    static __raw_spin_lock: Option<unsafe extern "C" fn(*mut c_void)>;
+    static __raw_spin_unlock: Option<unsafe extern "C" fn(*mut c_void)>;
+}
+
+#[cfg(target_arch = "aarch64")]
+unsafe extern "C" {
+    static _queued_spin_lock_slowpath: unsafe extern "C" fn(*mut c_void, u32);
+    static _queued_read_lock_slowpath: unsafe extern "C" fn(*mut c_void);
+    static _queued_write_lock_slowpath: unsafe extern "C" fn(*mut c_void);
+}
+
+unsafe extern "C" {
+    static _frida_kmod_exit_worker: usize;
+    #[cfg(target_arch = "x86")]
+    #[link_name = "frida_k_frida_kmod_exit_worker"]
+    fn frida_kmod_exit_worker() -> !;
+}
+
+#[cfg(not(target_arch = "x86"))]
+unsafe fn frida_kmod_exit_worker() -> ! {
+    let exit: unsafe extern "C" fn() -> ! = unsafe { core::mem::transmute(_frida_kmod_exit_worker) };
+    unsafe { exit() }
+}
 type IrqHandlerFn = unsafe extern "C" fn(irq: c_int, cookie: *mut c_void) -> c_int;
 
 unsafe extern "C" {
@@ -900,6 +1293,14 @@ unsafe extern "C" {
     static _vmalloc: Option<unsafe extern "C" fn(usize) -> *mut u8>;
     // Executable memory moved out of the module loader in 6.12.
     static _execmem_alloc: Option<unsafe extern "C" fn(u32, usize) -> *mut u8>;
+    #[cfg(not(target_arch = "x86"))]
+    static _execmem_alloc_rw: Option<unsafe extern "C" fn(u32, usize) -> *mut u8>;
+    #[cfg(not(target_arch = "x86"))]
+    static _vmalloc_noprof: Option<unsafe extern "C" fn(usize) -> *mut u8>;
+    #[cfg(not(target_arch = "x86"))]
+    static _frida_kmod_remap_writable: Option<unsafe extern "C" fn(u64, u32) -> *mut c_void>;
+    #[cfg(not(target_arch = "x86"))]
+    static _frida_kmod_unmap_writable: Option<unsafe extern "C" fn(*mut c_void) -> bool>;
     static _set_memory_rw: Option<unsafe extern "C" fn(usize, c_int) -> c_int>;
     static _set_memory_ro: Option<unsafe extern "C" fn(usize, c_int) -> c_int>;
     static _set_memory_x: Option<unsafe extern "C" fn(usize, c_int) -> c_int>;
@@ -971,6 +1372,10 @@ unsafe extern "C" {
     static _current_task: usize;
     #[cfg(target_arch = "x86_64")]
     static _cpu_current_top_of_stack: usize;
+    #[cfg(target_arch = "x86_64")]
+    static ___preempt_count: usize;
+    #[cfg(any(target_arch = "x86", target_arch = "x86_64"))]
+    static _pcpu_hot: usize;
     #[cfg(target_arch = "arm")]
     static ___pv_offset: *const u64;
     #[cfg(target_arch = "arm")]

src/barebone/agent/src/linux/process_memory.rs

@@ -0,0 +1,99 @@
+use core::ffi::{c_int, c_void};
+
+use super::processes::HeldTask;
+
+pub fn allocate(id: u32, size: usize, protection: usize) -> usize {
+    let Some(mm) = BorrowedMemory::of(id) else {
+        return 0;
+    };
+
+    let address = unsafe { _vm_mmap(core::ptr::null_mut(), 0, size, protection, MAP_PRIVATE | MAP_ANONYMOUS, 0) };
+    drop(mm);
+
+    if address >= FIRST_ERROR_ADDRESS { 0 } else { address }
+}
+
+pub fn release(id: u32, address: usize, size: usize) -> c_int {
+    let Some(mm) = BorrowedMemory::of(id) else {
+        return -ESRCH;
+    };
+
+    let result = unsafe { _vm_munmap(address, size) };
+    drop(mm);
+
+    result
+}
+
+pub fn write(id: u32, address: usize, data: &[u8]) -> usize {
+    transfer(id, address, data.as_ptr() as *mut c_void, data.len(), FOLL_WRITE)
+}
+
+pub fn read(id: u32, address: usize, buffer: &mut [u8]) -> usize {
+    transfer(id, address, buffer.as_mut_ptr() as *mut c_void, buffer.len(), 0)
+}
+
+pub fn read_own(address: usize, buffer: &mut [u8]) -> bool {
+    let len = buffer.len();
+    transfer_with(super::native::current_task() as usize, address, buffer.as_mut_ptr() as *mut c_void,
+        len, 0) == len
+}
+
+pub fn write_own(address: usize, data: &[u8]) -> bool {
+    transfer_with(super::native::current_task() as usize, address, data.as_ptr() as *mut c_void,
+        data.len(), FOLL_WRITE) == data.len()
+}
+
+fn transfer(id: u32, address: usize, buffer: *mut c_void, size: usize, flags: u32) -> usize {
+    let Some(task) = HeldTask::with_id(id) else {
+        return 0;
+    };
+
+    transfer_with(task.address(), address, buffer, size, flags)
+}
+
+fn transfer_with(task: usize, address: usize, buffer: *mut c_void, size: usize, flags: u32) -> usize {
+    let transferred = unsafe { _access_process_vm(task as *mut c_void, address, buffer, size as c_int, flags) };
+
+    transferred.max(0) as usize
+}
+
+struct BorrowedMemory(*mut c_void);
+
+impl BorrowedMemory {
+    fn of(id: u32) -> Option<Self> {
+        let task = HeldTask::with_id(id)?;
+        let mm = unsafe { _get_task_mm(task.address() as *mut c_void) };
+        if mm.is_null() {
+            return None;
+        }
+
+        unsafe { _kthread_use_mm(mm) };
+
+        Some(Self(mm))
+    }
+}
+
+impl Drop for BorrowedMemory {
+    fn drop(&mut self) {
+        unsafe {
+            _kthread_unuse_mm(self.0);
+            _mmput(self.0);
+        }
+    }
+}
+
+const MAP_PRIVATE: usize = 0x02;
+const MAP_ANONYMOUS: usize = 0x20;
+const FOLL_WRITE: u32 = 0x01;
+const FIRST_ERROR_ADDRESS: usize = usize::MAX - 4095;
+const ESRCH: c_int = 3;
+
+unsafe extern "C" {
+    static _get_task_mm: unsafe extern "C" fn(*mut c_void) -> *mut c_void;
+    static _mmput: unsafe extern "C" fn(*mut c_void);
+    static _kthread_use_mm: unsafe extern "C" fn(*mut c_void);
+    static _kthread_unuse_mm: unsafe extern "C" fn(*mut c_void);
+    static _vm_mmap: unsafe extern "C" fn(*mut c_void, usize, usize, usize, usize, usize) -> usize;
+    static _vm_munmap: unsafe extern "C" fn(usize, usize) -> c_int;
+    static _access_process_vm: unsafe extern "C" fn(*mut c_void, usize, *mut c_void, c_int, u32) -> c_int;
+}

src/barebone/agent/src/linux/processes.rs

@@ -72,6 +72,53 @@ pub unsafe fn let_go_of(file: *mut c_void) {
     unsafe { _fput(file) };
 }
 
+#[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+pub(crate) struct HeldTask(usize);
+
+#[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+impl HeldTask {
+    pub(crate) fn with_id(id: u32) -> Option<Self> {
+        let task = unsafe { _find_get_task_by_vpid(id as c_int) } as usize;
+        (task != 0).then(|| Self(task))
+    }
+
+    pub(crate) fn leader(&self) -> Option<Self> {
+        let group_leader = super::layout::field_offset("task_struct", "group_leader")?;
+        let usage = super::layout::field_offset("task_struct", "usage")?;
+
+        let flags = lock_tasklist();
+        let leader = unsafe { ((self.0 + group_leader) as *const usize).read() };
+        super::native::hold(leader + usage);
+        unlock_tasklist(flags);
+
+        Some(Self(leader))
+    }
+
+    pub(crate) fn address(&self) -> usize {
+        self.0
+    }
+
+    pub(crate) fn into_address(self) -> usize {
+        let task = self.0;
+        core::mem::forget(self);
+        task
+    }
+
+    pub(crate) unsafe fn from_address(task: usize) -> Self {
+        Self(task)
+    }
+}
+
+#[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+impl Drop for HeldTask {
+    fn drop(&mut self) {
+        let usage = super::layout::field_offset("task_struct", "usage").unwrap();
+        if super::native::let_go(self.0 + usage) {
+            unsafe { ___put_task_struct(self.0 as *mut c_void) };
+        }
+    }
+}
+
 pub fn task_with_id(id: u32) -> Option<usize> {
     let layout = task_layout()?;
 
@@ -573,12 +620,16 @@ fn lock_tasklist() -> usize {
 #[cfg(not(target_arch = "x86"))]
 fn lock_tasklist() -> usize {
     unsafe {
-        match (__raw_read_lock_irqsave, __raw_read_unlock_irqrestore) {
-            (Some(lock), Some(_)) => lock(_tasklist_lock),
-            _ => {
-                __raw_read_lock.unwrap()(_tasklist_lock);
+        match (__raw_read_lock_irqsave, __raw_read_unlock_irqrestore, __raw_read_lock) {
+            (Some(lock), Some(_), _) => lock(_tasklist_lock),
+            (_, _, Some(lock)) => {
+                lock(_tasklist_lock);
                 0
             }
+            #[cfg(target_arch = "aarch64")]
+            _ => super::native::read_lock_irqsave(_tasklist_lock as usize),
+            #[cfg(not(target_arch = "aarch64"))]
+            _ => panic!("no way to lock the task list"),
         }
     }
 }
@@ -591,9 +642,13 @@ fn unlock_tasklist(flags: usize) {
 #[cfg(not(target_arch = "x86"))]
 fn unlock_tasklist(flags: usize) {
     unsafe {
-        match (__raw_read_lock_irqsave, __raw_read_unlock_irqrestore) {
-            (Some(_), Some(unlock)) => unlock(_tasklist_lock, flags),
-            _ => __raw_read_unlock.unwrap()(_tasklist_lock),
+        match (__raw_read_lock_irqsave, __raw_read_unlock_irqrestore, __raw_read_unlock) {
+            (Some(_), Some(unlock), _) => unlock(_tasklist_lock, flags),
+            (_, _, Some(unlock)) => unlock(_tasklist_lock),
+            #[cfg(target_arch = "aarch64")]
+            _ => super::native::read_unlock_irqrestore(_tasklist_lock as usize, flags),
+            #[cfg(not(target_arch = "aarch64"))]
+            _ => panic!("no way to unlock the task list"),
         }
     }
 }
@@ -602,10 +657,6 @@ const PIDTYPE_TGID: c_int = 1;
 const PIDTYPE_MAX: usize = 4;
 
 pub fn cloak_task(task: *mut c_void) {
-    // detach_pid's arity differs across kernels (the classic detach_pid(task, type)
-    // versus the batched detach_pid(pids, task, type) paired with free_pids); calling
-    // the wrong one drives a UBSAN/CFI trap on hardened kernels. Leave the thread
-    // visible rather than risk it.
     if true {
         let _ = task;
         return;
@@ -638,8 +689,17 @@ unsafe fn detach_pid(pids: *mut *mut c_void, task: *mut c_void, kind: c_int) {
 }
 
 #[cfg(not(target_arch = "x86"))]
-unsafe fn detach_pid(pids: *mut *mut c_void, task: *mut c_void, kind: c_int) {
-    unsafe { _detach_pid(pids, task, kind) };
+pub(crate) unsafe fn detach_pid(pids: *mut *mut c_void, task: *mut c_void, kind: c_int) {
+    type ReleasingDetachPid = unsafe extern "C" fn(*mut c_void, c_int);
+
+    unsafe {
+        if _free_pids.is_some() {
+            _detach_pid(pids, task, kind);
+        } else {
+            let releasing: ReleasingDetachPid = core::mem::transmute(_detach_pid);
+            releasing(task, kind);
+        }
+    }
 }
 
 #[cfg(target_arch = "x86")]
@@ -648,11 +708,13 @@ unsafe fn free_pids(pids: *mut *mut c_void) {
 }
 
 #[cfg(not(target_arch = "x86"))]
-unsafe fn free_pids(pids: *mut *mut c_void) {
-    unsafe { _free_pids(pids) };
+pub(crate) unsafe fn free_pids(pids: *mut *mut c_void) {
+    if let Some(free) = unsafe { _free_pids } {
+        unsafe { free(pids) };
+    }
 }
 
-unsafe fn unlink(node: usize) {
+pub(crate) unsafe fn unlink(node: usize) {
     let next = unsafe { (node as *const usize).read() };
     let prev = unsafe { ((node + WORD_SIZE) as *const usize).read() };
     unsafe {
@@ -664,34 +726,64 @@ unsafe fn unlink(node: usize) {
 }
 
 #[cfg(target_arch = "x86")]
-fn write_lock_tasklist() -> usize {
+pub(crate) fn write_lock_tasklist() -> usize {
     unsafe { frida_k_write_lock_tasklist() }
 }
 
-#[cfg(not(target_arch = "x86"))]
-fn write_lock_tasklist() -> usize {
+#[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+pub(crate) unsafe fn unlink_for_readers(node: usize) {
+    let next = unsafe { (node as *const usize).read() };
+    let prev = unsafe { ((node + WORD_SIZE) as *const usize).read() };
     unsafe {
-        match (__raw_write_lock_irqsave, __raw_write_unlock_irqrestore) {
-            (Some(lock), Some(_)) => lock(_tasklist_lock),
-            _ => {
-                __raw_write_lock.unwrap()(_tasklist_lock);
+        ((next + WORD_SIZE) as *mut usize).write(prev);
+        (prev as *mut usize).write(next);
+    }
+}
+
+#[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+pub(crate) unsafe fn append_for_readers(node: usize, head: usize) {
+    let prev = unsafe { ((head + WORD_SIZE) as *const usize).read() };
+    unsafe {
+        (node as *mut usize).write(head);
+        ((node + WORD_SIZE) as *mut usize).write(prev);
+        core::sync::atomic::AtomicUsize::from_ptr(prev as *mut usize)
+            .store(node, core::sync::atomic::Ordering::Release);
+        ((head + WORD_SIZE) as *mut usize).write(node);
+    }
+}
+
+#[cfg(not(target_arch = "x86"))]
+pub(crate) fn write_lock_tasklist() -> usize {
+    unsafe {
+        match (__raw_write_lock_irqsave, __raw_write_unlock_irqrestore, __raw_write_lock) {
+            (Some(lock), Some(_), _) => lock(_tasklist_lock),
+            (_, _, Some(lock)) => {
+                lock(_tasklist_lock);
                 0
             }
+            #[cfg(target_arch = "aarch64")]
+            _ => super::native::write_lock_irqsave(_tasklist_lock as usize),
+            #[cfg(not(target_arch = "aarch64"))]
+            _ => panic!("no way to lock the task list"),
         }
     }
 }
 
 #[cfg(target_arch = "x86")]
-fn write_unlock_tasklist(flags: usize) {
+pub(crate) fn write_unlock_tasklist(flags: usize) {
     unsafe { frida_k_write_unlock_tasklist(flags) };
 }
 
 #[cfg(not(target_arch = "x86"))]
-fn write_unlock_tasklist(flags: usize) {
+pub(crate) fn write_unlock_tasklist(flags: usize) {
     unsafe {
-        match (__raw_write_lock_irqsave, __raw_write_unlock_irqrestore) {
-            (Some(_), Some(unlock)) => unlock(_tasklist_lock, flags),
-            _ => __raw_write_unlock.unwrap()(_tasklist_lock),
+        match (__raw_write_lock_irqsave, __raw_write_unlock_irqrestore, __raw_write_unlock) {
+            (Some(_), Some(unlock), _) => unlock(_tasklist_lock, flags),
+            (_, _, Some(unlock)) => unlock(_tasklist_lock),
+            #[cfg(target_arch = "aarch64")]
+            _ => super::native::write_unlock_irqrestore(_tasklist_lock as usize, flags),
+            #[cfg(not(target_arch = "aarch64"))]
+            _ => panic!("no way to unlock the task list"),
         }
     }
 }
@@ -728,7 +820,11 @@ unsafe extern "C" {
     #[cfg(not(target_arch = "x86"))]
     static _detach_pid: unsafe extern "C" fn(*mut *mut c_void, *mut c_void, c_int);
     #[cfg(not(target_arch = "x86"))]
-    static _free_pids: unsafe extern "C" fn(*mut *mut c_void);
+    static _free_pids: Option<unsafe extern "C" fn(*mut *mut c_void)>;
+    #[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+    static _find_get_task_by_vpid: unsafe extern "C" fn(c_int) -> *mut c_void;
+    #[cfg(any(target_arch = "x86_64", target_arch = "aarch64"))]
+    static ___put_task_struct: unsafe extern "C" fn(*mut c_void);
     #[cfg(not(target_arch = "x86"))]
     static __raw_write_lock: Option<unsafe extern "C" fn(*mut c_void)>;
     #[cfg(not(target_arch = "x86"))]

src/barebone/agent/src/linux/regparm.c

@@ -625,6 +625,24 @@ int frida_k_unregister_module_notifier (void * a0)
   return ((fn_t) _unregister_module_notifier) (a0);
 }
 
+extern void * _mutex_lock;
+
+void frida_k_mutex_lock (void * a0)
+{
+  typedef void (__attribute__ ((regparm (3))) * fn_t) (void *);
+
+  ((fn_t) _mutex_lock) (a0);
+}
+
+extern void * _mutex_unlock;
+
+void frida_k_mutex_unlock (void * a0)
+{
+  typedef void (__attribute__ ((regparm (3))) * fn_t) (void *);
+
+  ((fn_t) _mutex_unlock) (a0);
+}
+
 extern int frida_cb_module_state (void * nb, unsigned long action, void * data);
 
 __attribute__ ((regparm (3))) int
@@ -658,3 +676,57 @@ frida_kcb_die (void * nb, unsigned long action, void * data)
 {
   return frida_cb_die (nb, action, data);
 }
+
+extern void * _frida_kmod_link_open;
+
+int frida_k_frida_kmod_link_open (void)
+{
+  typedef int (__attribute__ ((regparm (3))) * fn_t) (void);
+
+  return ((fn_t) _frida_kmod_link_open) ();
+}
+
+extern void * _frida_kmod_link_close;
+
+void frida_k_frida_kmod_link_close (void)
+{
+  typedef void (__attribute__ ((regparm (3))) * fn_t) (void);
+
+  ((fn_t) _frida_kmod_link_close) ();
+}
+
+extern void * _frida_kmod_link_send;
+
+int frida_k_frida_kmod_link_send (const void * data, unsigned long size)
+{
+  typedef int (__attribute__ ((regparm (3))) * fn_t) (const void *, unsigned long);
+
+  return ((fn_t) _frida_kmod_link_send) (data, size);
+}
+
+extern void * _frida_kmod_link_recv;
+
+long frida_k_frida_kmod_link_recv (void * data, unsigned long size)
+{
+  typedef long (__attribute__ ((regparm (3))) * fn_t) (void *, unsigned long);
+
+  return ((fn_t) _frida_kmod_link_recv) (data, size);
+}
+
+extern void * _frida_kmod_link_pending;
+
+_Bool frida_k_frida_kmod_link_pending (void)
+{
+  typedef _Bool (__attribute__ ((regparm (3))) * fn_t) (void);
+
+  return ((fn_t) _frida_kmod_link_pending) ();
+}
+
+extern void * _frida_kmod_exit_worker;
+
+void frida_k_frida_kmod_exit_worker (void)
+{
+  typedef void (__attribute__ ((regparm (3))) * fn_t) (void);
+
+  ((fn_t) _frida_kmod_exit_worker) ();
+}

src/barebone/agent/src/linux/user.rs

@@ -125,7 +125,6 @@ fn leave() -> ! {
         syscall(CLOSE, descriptor as usize, 0, 0, 0, 0, 0);
     }
 
-    Arena::at(unsafe { ARENA }).gone();
     syscall(EXIT_GROUP, 0, 0, 0, 0, 0, 0);
 
     loop {}

src/barebone/agent/src/linux/user_threads.rs

@@ -0,0 +1,419 @@
+use core::ffi::{c_int, c_void};
+use core::mem::size_of;
+use core::ptr;
+use core::sync::atomic::{AtomicUsize, Ordering};
+
+use super::layout::field_offset;
+use super::native::{self, hold, let_go, lock_spinlock, unlock_spinlock};
+use super::processes::{
+    HeldTask, append_for_readers, detach_pid, free_pids, unlink, unlink_for_readers,
+    write_lock_tasklist, write_unlock_tasklist,
+};
+
+pub fn start_in_process(id: u32, entry: usize, stack: usize, argument: usize, tls: usize) -> c_int {
+    let Some(layout) = describe_thread_group() else {
+        return -ENOSYS;
+    };
+    let Some(task) = HeldTask::with_id(id) else {
+        return -ESRCH;
+    };
+    let Some(leader) = task.leader() else {
+        return -ESRCH;
+    };
+    let mm = unsafe { _get_task_mm(task.address() as *mut c_void) };
+    if mm.is_null() {
+        return -ESRCH;
+    }
+
+    let birth = native::alloc(size_of::<Birth>()) as *mut Birth;
+    unsafe {
+        birth.write(Birth { entry, stack, argument, tls, leader: leader.into_address(), layout });
+
+        _kthread_use_mm(mm);
+        let started = _user_mode_thread(thread_entry(), birth as *mut c_void, 0);
+        _kthread_unuse_mm(mm);
+        _mmput(mm);
+
+        if started < 0 {
+            drop(HeldTask::from_address((*birth).leader));
+            native::free(birth as *mut u8, size_of::<Birth>());
+        }
+
+        started
+    }
+}
+
+struct Birth {
+    entry: usize,
+    stack: usize,
+    argument: usize,
+    tls: usize,
+    leader: usize,
+    layout: ThreadGroupLayout,
+}
+
+fn thread_entry() -> native::KthreadFn {
+    native::kernel_thread_function(&THREAD_ENTRY, frida_cb_user_thread)
+}
+
+static THREAD_ENTRY: AtomicUsize = AtomicUsize::new(0);
+
+unsafe extern "C" fn frida_cb_user_thread(argument: *mut c_void) -> c_int {
+    let birth = unsafe { (argument as *const Birth).read() };
+    native::free(argument as *mut u8, size_of::<Birth>());
+
+    let leader = unsafe { HeldTask::from_address(birth.leader) };
+    let child = native::current_task() as usize;
+    join_thread_group(child, leader.address(), &birth.layout);
+    adopt_context_of(child, leader.address(), &birth.layout);
+    drop(leader);
+
+    super::injection::hand_userland_to(birth.entry, birth.stack, birth.argument);
+    use_thread_pointer(child, birth.tls, &birth.layout);
+
+    0
+}
+
+fn join_thread_group(child: usize, leader: usize, layout: &ThreadGroupLayout) {
+    let task = &layout.task;
+    let old_signal = read_word(child + task.signal);
+    let old_sighand = read_word(child + task.sighand);
+    let signal = read_word(leader + task.signal);
+    let sighand = read_word(leader + task.sighand);
+    let mut freed = [ptr::null_mut::<c_void>(); PIDTYPE_MAX];
+
+    let flags = write_lock_tasklist();
+
+    unsafe {
+        detach_pid(freed.as_mut_ptr(), child as *mut c_void, PIDTYPE_SID);
+        detach_pid(freed.as_mut_ptr(), child as *mut c_void, PIDTYPE_PGID);
+        detach_pid(freed.as_mut_ptr(), child as *mut c_void, PIDTYPE_TGID);
+
+        unlink_for_readers(child + task.tasks);
+        unlink(child + task.sibling);
+        unlink_for_readers(child + task.thread_node);
+    }
+
+    let parent = read_word(leader + task.real_parent);
+    write_word(child + task.real_parent, parent);
+    write_word(child + task.parent, parent);
+    write_word(child + task.group_leader, leader);
+    write_i32(child + task.tgid, read_i32(leader + task.tgid));
+    write_i32(child + task.exit_signal, -1);
+    write_word(child + task.signal, signal);
+    write_word(child + task.sighand, sighand);
+
+    hold(sighand + layout.sighand.count);
+
+    lock_spinlock(sighand + layout.sighand.siglock);
+
+    write_i32(signal + layout.signal.nr_threads, read_i32(signal + layout.signal.nr_threads) + 1);
+    if let Some(quick_threads) = layout.signal.quick_threads {
+        write_i32(signal + quick_threads, read_i32(signal + quick_threads) + 1);
+    }
+    hold(signal + layout.signal.live);
+    hold(signal + layout.signal.sigcnt);
+
+    unsafe {
+        (layout.calls.task_join_group_stop)(child as *mut c_void);
+
+        if let Some(thread_group) = task.thread_group {
+            append_for_readers(child + thread_group, leader + thread_group);
+        }
+        append_for_readers(child + task.thread_node, signal + layout.signal.thread_head);
+    }
+
+    unlock_spinlock(sighand + layout.sighand.siglock);
+
+    write_unlock_tasklist(flags);
+
+    unsafe {
+        free_pids(freed.as_mut_ptr());
+
+        (layout.calls.cleanup_sighand)(old_sighand as *mut c_void);
+        if let_go(old_signal + layout.signal.sigcnt) {
+            _kmem_cache_free(*_signal_cachep, old_signal as *mut c_void);
+        }
+    }
+}
+
+fn adopt_context_of(child: usize, leader: usize, layout: &ThreadGroupLayout) {
+    let task = &layout.task;
+
+    let files = read_word(leader + task.files);
+    let old_files = read_word(child + task.files);
+    hold(files + layout.files_count);
+    lock_spinlock(child + task.alloc_lock);
+    write_word(child + task.files, files);
+    unlock_spinlock(child + task.alloc_lock);
+    unsafe { (layout.calls.put_files_struct)(old_files as *mut c_void) };
+
+    let fs = read_word(leader + task.fs);
+    let old_fs = read_word(child + task.fs);
+    lock_spinlock(fs + layout.fs.lock);
+    write_i32(fs + layout.fs.users, read_i32(fs + layout.fs.users) + 1);
+    unlock_spinlock(fs + layout.fs.lock);
+    lock_spinlock(child + task.alloc_lock);
+    write_word(child + task.fs, fs);
+    unlock_spinlock(child + task.alloc_lock);
+    lock_spinlock(old_fs + layout.fs.lock);
+    let users_left = read_i32(old_fs + layout.fs.users) - 1;
+    write_i32(old_fs + layout.fs.users, users_left);
+    unlock_spinlock(old_fs + layout.fs.lock);
+    if users_left == 0 {
+        unsafe { (layout.calls.free_fs_struct)(old_fs as *mut c_void) };
+    }
+
+    let real_cred = read_word(leader + task.real_cred);
+    let cred = read_word(leader + task.cred);
+    let old_real_cred = read_word(child + task.real_cred);
+    let old_cred = read_word(child + task.cred);
+    hold(real_cred + layout.cred_usage);
+    write_word(child + task.real_cred, real_cred);
+    hold(cred + layout.cred_usage);
+    unsafe { AtomicUsize::from_ptr((child + task.cred) as *mut usize).store(cred, Ordering::Release) };
+    release_cred(old_cred, layout);
+    release_cred(old_real_cred, layout);
+
+    let namespaces = read_word(leader + task.nsproxy);
+    hold(namespaces + layout.nsproxy_count);
+    unsafe { (layout.calls.switch_task_namespaces)(child as *mut c_void, namespaces as *mut c_void) };
+
+    if let Some(undo) = &layout.semaphore_undo {
+        let undo_list = read_word(leader + undo.list);
+        if undo_list != 0 {
+            hold(undo_list + undo.refcnt);
+            write_word(child + undo.list, undo_list);
+        }
+    }
+
+    #[cfg(target_arch = "aarch64")]
+    if let Some(lengths) = &layout.vector_lengths {
+        adopt_vector_lengths(child, leader, lengths);
+    }
+}
+
+fn release_cred(cred: usize, layout: &ThreadGroupLayout) {
+    if let_go(cred + layout.cred_usage) {
+        unsafe { (layout.calls.put_cred)(cred as *mut c_void) };
+    }
+}
+
+#[cfg(target_arch = "aarch64")]
+fn adopt_vector_lengths(child: usize, leader: usize, lengths: &VectorLengthLayout) {
+    for array in [lengths.current, lengths.on_exec] {
+        for kind in 0..VECTOR_KINDS {
+            let at = array + kind * size_of::<u32>();
+            write_i32(child + at, read_i32(leader + at));
+        }
+    }
+}
+
+#[cfg(target_arch = "x86_64")]
+fn use_thread_pointer(child: usize, tls: usize, layout: &ThreadGroupLayout) {
+    const MSR_FS_BASE: u32 = 0xc000_0100;
+
+    write_word(child + layout.task.thread_pointer, tls);
+    unsafe {
+        core::arch::asm!("wrmsr", in("ecx") MSR_FS_BASE, in("eax") tls as u32,
+            in("edx") (tls >> 32) as u32, options(nostack));
+    }
+}
+
+#[cfg(target_arch = "aarch64")]
+fn use_thread_pointer(child: usize, tls: usize, layout: &ThreadGroupLayout) {
+    write_word(child + layout.task.thread_pointer, tls);
+    unsafe { core::arch::asm!("msr tpidr_el0, {}", in(reg) tls, options(nostack)) };
+}
+
+struct ThreadGroupLayout {
+    calls: ThreadGroupCalls,
+    task: TaskLayout,
+    signal: SignalLayout,
+    sighand: SighandLayout,
+    fs: FsLayout,
+    files_count: usize,
+    cred_usage: usize,
+    nsproxy_count: usize,
+    semaphore_undo: Option<SemaphoreUndoLayout>,
+    #[cfg(target_arch = "aarch64")]
+    vector_lengths: Option<VectorLengthLayout>,
+}
+
+struct ThreadGroupCalls {
+    task_join_group_stop: unsafe extern "C" fn(*mut c_void),
+    cleanup_sighand: unsafe extern "C" fn(*mut c_void),
+    put_files_struct: unsafe extern "C" fn(*mut c_void),
+    free_fs_struct: unsafe extern "C" fn(*mut c_void),
+    put_cred: unsafe extern "C" fn(*mut c_void),
+    switch_task_namespaces: unsafe extern "C" fn(*mut c_void, *mut c_void),
+}
+
+struct TaskLayout {
+    tgid: usize,
+    real_parent: usize,
+    parent: usize,
+    group_leader: usize,
+    exit_signal: usize,
+    signal: usize,
+    sighand: usize,
+    tasks: usize,
+    sibling: usize,
+    thread_node: usize,
+    thread_group: Option<usize>,
+    files: usize,
+    fs: usize,
+    cred: usize,
+    real_cred: usize,
+    nsproxy: usize,
+    alloc_lock: usize,
+    thread_pointer: usize,
+}
+
+struct SignalLayout {
+    nr_threads: usize,
+    quick_threads: Option<usize>,
+    live: usize,
+    sigcnt: usize,
+    thread_head: usize,
+}
+
+struct SighandLayout {
+    count: usize,
+    siglock: usize,
+}
+
+struct FsLayout {
+    users: usize,
+    lock: usize,
+}
+
+struct SemaphoreUndoLayout {
+    list: usize,
+    refcnt: usize,
+}
+
+#[cfg(target_arch = "aarch64")]
+struct VectorLengthLayout {
+    current: usize,
+    on_exec: usize,
+}
+
+fn describe_thread_group() -> Option<ThreadGroupLayout> {
+    Some(ThreadGroupLayout {
+        calls: unsafe {
+            ThreadGroupCalls {
+                task_join_group_stop: _task_join_group_stop?,
+                cleanup_sighand: ___cleanup_sighand?,
+                put_files_struct: _put_files_struct?,
+                free_fs_struct: _free_fs_struct?,
+                put_cred: ___put_cred?,
+                switch_task_namespaces: _switch_task_namespaces?,
+            }
+        },
+        task: TaskLayout {
+            tgid: field_offset("task_struct", "tgid")?,
+            real_parent: field_offset("task_struct", "real_parent")?,
+            parent: field_offset("task_struct", "parent")?,
+            group_leader: field_offset("task_struct", "group_leader")?,
+            exit_signal: field_offset("task_struct", "exit_signal")?,
+            signal: field_offset("task_struct", "signal")?,
+            sighand: field_offset("task_struct", "sighand")?,
+            tasks: field_offset("task_struct", "tasks")?,
+            sibling: field_offset("task_struct", "sibling")?,
+            thread_node: field_offset("task_struct", "thread_node")?,
+            thread_group: field_offset("task_struct", "thread_group"),
+            files: field_offset("task_struct", "files")?,
+            fs: field_offset("task_struct", "fs")?,
+            cred: field_offset("task_struct", "cred")?,
+            real_cred: field_offset("task_struct", "real_cred")?,
+            nsproxy: field_offset("task_struct", "nsproxy")?,
+            alloc_lock: field_offset("task_struct", "alloc_lock")?,
+            thread_pointer: field_offset("task_struct", THREAD_POINTER_FIELD)?,
+        },
+        signal: SignalLayout {
+            nr_threads: field_offset("signal_struct", "nr_threads")?,
+            quick_threads: field_offset("signal_struct", "quick_threads"),
+            live: field_offset("signal_struct", "live")?,
+            sigcnt: field_offset("signal_struct", "sigcnt")?,
+            thread_head: field_offset("signal_struct", "thread_head")?,
+        },
+        sighand: SighandLayout {
+            count: field_offset("sighand_struct", "count")?,
+            siglock: field_offset("sighand_struct", "siglock")?,
+        },
+        fs: FsLayout {
+            users: field_offset("fs_struct", "users")?,
+            lock: field_offset("fs_struct", "lock").or_else(|| field_offset("fs_struct", "seq.lock"))?,
+        },
+        files_count: field_offset("files_struct", "count")?,
+        cred_usage: field_offset("cred", "usage")?,
+        nsproxy_count: field_offset("nsproxy", "count")?,
+        semaphore_undo: describe_semaphore_undo(),
+        #[cfg(target_arch = "aarch64")]
+        vector_lengths: describe_vector_lengths(),
+    })
+}
+
+fn describe_semaphore_undo() -> Option<SemaphoreUndoLayout> {
+    Some(SemaphoreUndoLayout {
+        list: field_offset("task_struct", "sysvsem.undo_list")?,
+        refcnt: field_offset("sem_undo_list", "refcnt")?,
+    })
+}
+
+#[cfg(target_arch = "aarch64")]
+fn describe_vector_lengths() -> Option<VectorLengthLayout> {
+    Some(VectorLengthLayout {
+        current: field_offset("task_struct", "thread.vl")?,
+        on_exec: field_offset("task_struct", "thread.vl_onexec")?,
+    })
+}
+
+#[cfg(target_arch = "x86_64")]
+const THREAD_POINTER_FIELD: &str = "thread.fsbase";
+#[cfg(target_arch = "aarch64")]
+const THREAD_POINTER_FIELD: &str = "thread.uw.tp_value";
+
+#[cfg(target_arch = "aarch64")]
+const VECTOR_KINDS: usize = 2;
+
+fn read_word(address: usize) -> usize {
+    unsafe { (address as *const usize).read() }
+}
+
+fn write_word(address: usize, value: usize) {
+    unsafe { (address as *mut usize).write(value) };
+}
+
+fn read_i32(address: usize) -> i32 {
+    unsafe { (address as *const i32).read() }
+}
+
+fn write_i32(address: usize, value: i32) {
+    unsafe { (address as *mut i32).write(value) };
+}
+
+const PIDTYPE_TGID: c_int = 1;
+const PIDTYPE_PGID: c_int = 2;
+const PIDTYPE_SID: c_int = 3;
+const PIDTYPE_MAX: usize = 4;
+
+const ESRCH: c_int = 3;
+const ENOSYS: c_int = 38;
+
+unsafe extern "C" {
+    static _get_task_mm: unsafe extern "C" fn(*mut c_void) -> *mut c_void;
+    static _mmput: unsafe extern "C" fn(*mut c_void);
+    static _kthread_use_mm: unsafe extern "C" fn(*mut c_void);
+    static _kthread_unuse_mm: unsafe extern "C" fn(*mut c_void);
+    static _user_mode_thread: unsafe extern "C" fn(native::KthreadFn, *mut c_void, usize) -> c_int;
+    static _task_join_group_stop: Option<unsafe extern "C" fn(*mut c_void)>;
+    static ___cleanup_sighand: Option<unsafe extern "C" fn(*mut c_void)>;
+    static _kmem_cache_free: unsafe extern "C" fn(*mut c_void, *mut c_void);
+    static _signal_cachep: *const *mut c_void;
+    static _put_files_struct: Option<unsafe extern "C" fn(*mut c_void)>;
+    static _free_fs_struct: Option<unsafe extern "C" fn(*mut c_void)>;
+    static ___put_cred: Option<unsafe extern "C" fn(*mut c_void)>;
+    static _switch_task_namespaces: Option<unsafe extern "C" fn(*mut c_void, *mut c_void)>;
+}

src/barebone/agent/src/main.rs

@@ -2,9 +2,6 @@
 // image and calls its `_start`. All of the agent lives in the library crate;
 // this target exists only to produce that link. build.rs passes `--undefined
 // _start` so the entrypoint survives --gc-sections.
-//
-// The Linux flavour is built as a static library instead and linked into a
-// kernel module by kbuild — see linux/Kbuild.
 
 #![no_std]
 #![no_main]

src/darwin/frida-helper-backend.vala

@@ -328,7 +328,6 @@ namespace Frida {
 				task = task_for_pid (pid);
 				resume_process (task);
 			} catch (Error e) {
-			} catch (Efni.Error e) {
 			} finally {
 				if (task != 0)
 					deallocate_port (task);
@@ -341,7 +340,6 @@ namespace Frida {
 				task = task_for_pid (pid);
 				resume_process_fast (task);
 			} catch (Error e) {
-			} catch (Efni.Error e) {
 			} finally {
 				if (task != 0)
 					deallocate_port (task);
@@ -607,8 +605,6 @@ namespace Frida {
 				policy_softener.retain (pid);
 			} catch (Error e) {
 				assert_not_reached ();
-			} catch (Efni.Error e) {
-				assert_not_reached ();
 			}
 
 			if (mapped_module != null)
@@ -741,8 +737,6 @@ namespace Frida {
 					yield disable_spawn_gating (cancellable);
 				} catch (Error e) {
 					assert_not_reached ();
-				} catch (Efni.Error e) {
-					assert_not_reached ();
 				}
 			}
 		}

src/freebsd/binjector.vala

@@ -17,8 +17,6 @@ namespace Frida {
 						_resource_store = new ResourceStore ();
 					} catch (Error e) {
 						assert_not_reached ();
-					} catch (Efni.Error e) {
-						assert_not_reached ();
 					}
 				}
 				return _resource_store;

src/linux/linux-host-session.vala

@@ -109,7 +109,6 @@ namespace Frida {
 					var client = yield get_android_helper_client (cancellable);
 					yield client.transport.close (cancellable);
 				} catch (Error e) {
-				} catch (Efni.Error e) {
 				}
 			}
 #endif
@@ -578,8 +577,6 @@ namespace Frida {
 				yield promise.future.wait_async (cancellable);
 			} catch (Error e) {
 				assert_not_reached ();
-			} catch (Efni.Error e) {
-				assert_not_reached ();
 			}
 
 			worker_thread.join ();
@@ -1448,7 +1445,6 @@ namespace Frida {
 						Posix.kill ((Posix.pid_t) pid, Posix.Signal.CONT);
 					}
 				} catch (Error e) {
-				} catch (Efni.Error e) {
 				}
 			}
 			zymbiote_patches.clear ();
@@ -2345,7 +2341,6 @@ namespace Frida {
 					var crash = parse_java_report (entry.message);
 					on_crash_received (crash);
 				} catch (Error e) {
-				} catch (Efni.Error e) {
 				}
 
 				return;
@@ -2386,7 +2381,6 @@ namespace Frida {
 				var crash = parse_native_report (report);
 				on_crash_received (crash);
 			} catch (Error e) {
-			} catch (Efni.Error e) {
 			}
 		}
 

src/package-manager.vala

@@ -1246,7 +1246,7 @@ namespace Frida {
 		}
 
 		private static async Json.Reader? try_open_already_installed_manifest (PackageNode node, File dir, Cancellable? cancellable)
-				throws Efni.Error, IOError {
+				throws IOError {
 			try {
 				Json.Reader r = yield load_json (dir.get_child ("package.json"), cancellable);
 
@@ -1264,6 +1264,8 @@ namespace Frida {
 				return r;
 			} catch (Error e) {
 				return null;
+			} catch (Efni.Error e) {
+				return null;
 			}
 		}
 
@@ -2255,6 +2257,8 @@ namespace Frida {
 					yield tar_reader.feed (buffer[:n], cancellable);
 				} catch (Error e) {
 					throw new Error.PROTOCOL ("Unable to extract tarball at %s: %s", tarball_url, e.message);
+				} catch (Efni.Error e) {
+					throw new Error.PROTOCOL ("Unable to extract tarball at %s: %s", tarball_url, e.message);
 				}
 				read_total += (size_t) n;
 				report_bucket += (size_t) n;

src/prospero/binjector.vala

@@ -17,8 +17,6 @@ namespace Frida {
 						_resource_store = new ResourceStore ();
 					} catch (Error e) {
 						assert_not_reached ();
-					} catch (Efni.Error e) {
-						assert_not_reached ();
 					}
 				}
 				return _resource_store;
@@ -844,7 +842,6 @@ namespace Frida {
 				try {
 					protect_code (load_bias + segment.vm_address, (size_t) round_page (segment.vm_size));
 				} catch (Error e) {
-				} catch (Efni.Error e) {
 				}
 				return true;
 			});

src/qnx/qinjector.vala

@@ -13,8 +13,6 @@ namespace Frida {
 						_resource_store = new ResourceStore ();
 					} catch (Error e) {
 						assert_not_reached ();
-					} catch (Efni.Error e) {
-						assert_not_reached ();
 					}
 				}
 				return _resource_store;

src/simmy/simmy-host-session.vala

@@ -280,7 +280,6 @@ namespace Frida {
 				try {
 					icons = yield springboard_agent.fetch_application_icons (app_ids, cancellable);
 				} catch (Error e) {
-				} catch (Efni.Error e) {
 				}
 			}