dyncvoke-spoof 0.1.1

//! Call-stack spoofing for Dyncvoke.
//!
//! Public API is [`spoof!`] (direct call) and [`spoof_syscall!`] (indirect
//! syscall). Mode is a compile-time choice:
//!
//! - **Synthetic** (default, `dyncvoke` feature `spoof`). Builds
//!   `RtlUserThreadStart -> BaseThreadInitThunk -> gadget frames -> target`.
//!   Works from any thread, including pool threads.
//! - **Desync** (`dyncvoke` feature `spoof-desync`, crate feature `desync`).
//!   Splices spoofed frames onto a live `BaseThreadInitThunk` return on the
//!   current thread. Looks closer to a normal user thread. Fails on pool
//!   threads that never went through that path.
//!
//! ```ignore
//! use spoof::{spoof_syscall, AsPointer};
//! use core::ffi::c_void;
//! use core::ptr::null_mut;
//!
//! let mut addr: *mut c_void = null_mut();
//! let mut size: usize = 0x1000;
//! let status = spoof_syscall!(
//!     "NtAllocateVirtualMemory",
//!     -1isize,
//!     addr.as_ptr_mut(),
//!     0usize,
//!     size.as_ptr_mut(),
//!     0x3000u32,
//!     0x04u32,
//! ).unwrap() as i32;
//! ```
//!
//! The trampoline lives in `src/asm/{msvc,gnu}/*.asm` and reads
//! [`types::Config`] by field order. Do not reorder those fields.
//! Syscall SSNs come from [`dyncvoke_core::resolve_syscall`].

#![no_std]
#![cfg_attr(docsrs, feature(doc_cfg))]
extern crate alloc;

#[cfg(not(windows))]
compile_error!("dyncvoke-spoof is Windows-only");

use alloc::string::String;
use alloc::vec::Vec;
use core::ffi::c_void;

use data::lc;

pub mod pe;
pub mod types;
pub mod unwind;
pub mod util;

use crate::pe::{function_by_rva, runtime_function_table};
use crate::types::{Config, ImageRuntimeFunction};
use crate::unwind::{rbp_offset, stack_frame};
#[cfg(not(feature = "desync"))]
use crate::unwind::ignoring_set_fpreg;
use crate::util::{find_gadget, find_valid_instruction_offset, shuffle};

#[cfg(feature = "desync")]
use crate::util::find_base_thread_return_address;

// Compile-time-selected asm trampoline. Synthetic mode calls SpoofSynthetic,
// desync calls Spoof. The asm files compile both — the linker drops the
// unused symbol.
#[cfg(feature = "desync")]
unsafe extern "C" {
    fn Spoof(config: &mut Config) -> *mut c_void;
}

#[cfg(not(feature = "desync"))]
unsafe extern "C" {
    fn SpoofSynthetic(config: &mut Config) -> *mut c_void;
}

/// Spoof a direct function call. Resolves the spoofing scaffolding once
/// and then jumps to `addr` with the supplied args under a fake stack.
///
/// # Examples
///
/// ```ignore
/// let kernel32 = dyncvoke_core::get_module_base_address("kernel32.dll");
/// let virtual_alloc = dyncvoke_core::get_function_address(kernel32, "VirtualAlloc");
/// let addr = spoof::spoof!(
///     virtual_alloc,
///     core::ptr::null_mut::<core::ffi::c_void>(),
///     1 << 12,
///     0x3000u32,
///     0x04u32
/// )?;
/// ```
#[macro_export]
macro_rules! spoof {
    ($addr:expr, $($arg:expr),+ $(,)?) => {
        unsafe {
            $crate::__private::spoof(
                $addr as *mut ::core::ffi::c_void,
                &[$(::core::mem::transmute($arg as usize)),*],
                $crate::SpoofKind::Function,
            )
        }
    };
}

/// Spoof an indirect syscall. SSN is resolved via dyncvoke_core's Tartarus
/// Gate, then the syscall instruction inside ntdll is dispatched under a
/// fake stack.
///
/// # Examples
///
/// ```ignore
/// use spoof::{spoof_syscall, AsPointer};
/// let mut addr = core::ptr::null_mut::<core::ffi::c_void>();
/// let mut size = (1 << 12) as usize;
/// let status = spoof_syscall!(
///     "NtAllocateVirtualMemory",
///     -1isize,
///     addr.as_ptr_mut(),
///     0usize,
///     size.as_ptr_mut(),
///     0x3000u32,
///     0x04u32
/// )? as i32;
/// ```
#[macro_export]
macro_rules! spoof_syscall {
    ($name:expr, $($arg:expr),* $(,)?) => {
        unsafe {
            $crate::__private::spoof(
                ::core::ptr::null_mut(),
                &[$(::core::mem::transmute($arg as usize)),*],
                $crate::SpoofKind::Syscall($name),
            )
        }
    };
}

/// Selects which mode the public entry uses.
pub enum SpoofKind<'a> {
    /// Spoof a direct function call.
    Function,
    /// Spoof an indirect syscall by ntdll export name.
    Syscall(&'a str),
}

/// Where a failure happened. Carried by SpoofError variants as a non-string
/// tag so the error path doesn't leak plaintext module/function names into
/// `.rdata`. Display impl decrypts via obfstr at format time only.
#[derive(Debug, Clone, Copy)]
pub enum SpoofTag {
    Kernelbase,
    Kernel32,
    Ntdll,
    RtlUserThreadStart,
    BaseThreadInitThunk,
    AddRspGadget,
    JmpRbxGadget,
}

/// Spoofing errors. Variants carry [`SpoofTag`] rather than literal strings.
#[derive(Debug)]
pub enum SpoofError {
    TooManyArguments,
    NullFunctionAddress,
    ModuleNotLoaded(SpoofTag),
    FunctionNotFound(SpoofTag),
    PdataMissing(SpoofTag),
    UnwindInfoMissing(SpoofTag),
    PrologueNotFound,
    PushRbpPrologueNotFound,
    GadgetNotFound(SpoofTag),
    BaseThreadReturnNotFound,
    SsnResolutionFailed,
}

impl core::fmt::Display for SpoofError {
    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
        let tag_str = |t: SpoofTag| match t {
            SpoofTag::Kernelbase => lc!("kernelbase.dll"),
            SpoofTag::Kernel32 => lc!("kernel32.dll"),
            SpoofTag::Ntdll => lc!("ntdll.dll"),
            SpoofTag::RtlUserThreadStart => lc!("RtlUserThreadStart"),
            SpoofTag::BaseThreadInitThunk => lc!("BaseThreadInitThunk"),
            SpoofTag::AddRspGadget => lc!("add rsp,0x58; ret"),
            SpoofTag::JmpRbxGadget => lc!("jmp [rbx]"),
        };
        match self {
            SpoofError::TooManyArguments => write!(f, "{}", lc!("too many arguments")),
            SpoofError::NullFunctionAddress => write!(f, "{}", lc!("null function address")),
            SpoofError::ModuleNotLoaded(t) => write!(f, "{}: {}", lc!("module"), tag_str(*t)),
            SpoofError::FunctionNotFound(t) => write!(f, "{}: {}", lc!("export"), tag_str(*t)),
            SpoofError::PdataMissing(t) => write!(f, "{}: {}", lc!("pdata"), tag_str(*t)),
            SpoofError::UnwindInfoMissing(t) => write!(f, "{}: {}", lc!("unwind"), tag_str(*t)),
            SpoofError::PrologueNotFound => write!(f, "{}", lc!("no prologue")),
            SpoofError::PushRbpPrologueNotFound => write!(f, "{}", lc!("no rbp prologue")),
            SpoofError::GadgetNotFound(t) => write!(f, "{}: {}", lc!("gadget"), tag_str(*t)),
            SpoofError::BaseThreadReturnNotFound => write!(f, "{}", lc!("no thread return")),
            SpoofError::SsnResolutionFailed => write!(f, "{}", lc!("ssn fail")),
        }
    }
}

impl core::error::Error for SpoofError {}

/// Ergonomic trait that lets call sites write `addr.as_ptr_mut()` in macro
/// args instead of `&mut addr as *mut _ as *mut c_void`.
pub trait AsPointer {
    fn as_ptr_const(&self) -> *const c_void;
    fn as_ptr_mut(&mut self) -> *mut c_void;
}

impl<T> AsPointer for T {
    #[inline(always)]
    fn as_ptr_const(&self) -> *const c_void {
        self as *const _ as *const c_void
    }
    #[inline(always)]
    fn as_ptr_mut(&mut self) -> *mut c_void {
        self as *mut _ as *mut c_void
    }
}

#[doc(hidden)]
pub mod __private {
    use super::*;
    #[allow(unused_imports)]
    use alloc::string::String;

    /// Synthetic-mode entry. Builds the full scaffolding from scratch and
    /// hands it to the asm trampoline. Works on any thread.
    #[cfg(not(feature = "desync"))]
    pub unsafe fn spoof(
        addr: *mut c_void,
        args: &[*const c_void],
        kind: SpoofKind,
    ) -> Result<*mut c_void, SpoofError> {
        if args.len() > 11 {
            return Err(SpoofError::TooManyArguments);
        }
        if matches!(kind, SpoofKind::Function) && addr.is_null() {
            return Err(SpoofError::NullFunctionAddress);
        }

        let mut config = Config::default();

        let kernelbase = dyncvoke_core::get_module_base_address(&lc!("kernelbase.dll"));
        if kernelbase == 0 {
            return Err(SpoofError::ModuleNotLoaded(SpoofTag::Kernelbase));
        }
        let ntdll = dyncvoke_core::get_module_base_address(&lc!("ntdll.dll"));
        if ntdll == 0 {
            return Err(SpoofError::ModuleNotLoaded(SpoofTag::Ntdll));
        }
        let kernel32 = dyncvoke_core::get_module_base_address(&lc!("kernel32.dll"));
        if kernel32 == 0 {
            return Err(SpoofError::ModuleNotLoaded(SpoofTag::Kernel32));
        }

        let rtl_user_addr =
            dyncvoke_core::get_function_address(ntdll, &lc!("RtlUserThreadStart"));
        if rtl_user_addr == 0 {
            return Err(SpoofError::FunctionNotFound(SpoofTag::RtlUserThreadStart));
        }
        let base_thread_addr =
            dyncvoke_core::get_function_address(kernel32, &lc!("BaseThreadInitThunk"));
        if base_thread_addr == 0 {
            return Err(SpoofError::FunctionNotFound(SpoofTag::BaseThreadInitThunk));
        }

        config.rtl_user_addr = rtl_user_addr as *const c_void;
        config.base_thread_addr = base_thread_addr as *const c_void;

        let kb_table = runtime_function_table(kernelbase as *mut c_void)
            .ok_or(SpoofError::PdataMissing(SpoofTag::Kernelbase))?;
        let ntdll_table = runtime_function_table(ntdll as *mut c_void)
            .ok_or(SpoofError::PdataMissing(SpoofTag::Ntdll))?;
        let k32_table = runtime_function_table(kernel32 as *mut c_void)
            .ok_or(SpoofError::PdataMissing(SpoofTag::Kernel32))?;

        let rtl_user_runtime =
            function_by_rva(ntdll_table, (rtl_user_addr - ntdll) as u32)
                .ok_or(SpoofError::UnwindInfoMissing(SpoofTag::RtlUserThreadStart))?;
        let base_thread_runtime =
            function_by_rva(k32_table, (base_thread_addr - kernel32) as u32)
                .ok_or(SpoofError::UnwindInfoMissing(SpoofTag::BaseThreadInitThunk))?;

        config.rtl_user_thread_size =
            ignoring_set_fpreg(ntdll as *mut c_void, rtl_user_runtime)
                .ok_or(SpoofError::UnwindInfoMissing(SpoofTag::RtlUserThreadStart))? as u64;
        config.base_thread_size =
            ignoring_set_fpreg(kernel32 as *mut c_void, base_thread_runtime)
                .ok_or(SpoofError::UnwindInfoMissing(SpoofTag::BaseThreadInitThunk))? as u64;

        let first_prolog = find_prolog(kernelbase as *mut c_void, kb_table)
            .ok_or(SpoofError::PrologueNotFound)?;
        config.first_frame_fp =
            (first_prolog.frame + first_prolog.offset as u64) as *const c_void;
        config.first_frame_size = first_prolog.stack_size as u64;

        let second_prolog = find_push_rbp(kernelbase as *mut c_void, kb_table)
            .ok_or(SpoofError::PushRbpPrologueNotFound)?;
        config.second_frame_fp =
            (second_prolog.frame + second_prolog.offset as u64) as *const c_void;
        config.second_frame_size = second_prolog.stack_size as u64;
        config.rbp_stack_offset = second_prolog.rbp_offset as u64;

        let (add_rsp_addr, size) = find_gadget(
            kernelbase as *mut c_void,
            &[0x48, 0x83, 0xC4, 0x58, 0xC3],
            kb_table,
            Some(0x58),
        )
        .ok_or(SpoofError::GadgetNotFound(SpoofTag::AddRspGadget))?;
        config.add_rsp_gadget = add_rsp_addr as *const c_void;
        config.add_rsp_frame_size = size as u64;

        let (jmp_rbx_addr, size) =
            find_gadget(kernelbase as *mut c_void, &[0xFF, 0x23], kb_table, None)
                .ok_or(SpoofError::GadgetNotFound(SpoofTag::JmpRbxGadget))?;
        config.jmp_rbx_gadget = jmp_rbx_addr as *const c_void;
        config.jmp_rbx_frame_size = size as u64;

        config.number_args = args.len() as u64;
        write_args(&mut config, args);

        match kind {
            SpoofKind::Function => config.spoof_function = addr as *const c_void,
            SpoofKind::Syscall(name) => {
                let (ssn, syscall_addr) = dyncvoke_core::resolve_syscall(name)
                    .map_err(|_| SpoofError::SsnResolutionFailed)?;
                config.is_syscall = 1;
                config.ssn = ssn as u32;
                config.spoof_function = syscall_addr as *const c_void;
            }
        }

        Ok(SpoofSynthetic(&mut config))
    }

    /// Desync-mode entry. Reuses the current thread's real
    /// BaseThreadInitThunk return record. Won't work on pool threads.
    #[cfg(feature = "desync")]
    pub unsafe fn spoof(
        addr: *mut c_void,
        args: &[*const c_void],
        kind: SpoofKind,
    ) -> Result<*mut c_void, SpoofError> {
        if args.len() > 11 {
            return Err(SpoofError::TooManyArguments);
        }
        if matches!(kind, SpoofKind::Function) && addr.is_null() {
            return Err(SpoofError::NullFunctionAddress);
        }

        let mut config = Config::default();

        let kernelbase = dyncvoke_core::get_module_base_address(&lc!("kernelbase.dll"));
        if kernelbase == 0 {
            return Err(SpoofError::ModuleNotLoaded(SpoofTag::Kernelbase));
        }
        let kernel32 = dyncvoke_core::get_module_base_address(&lc!("kernel32.dll"));
        if kernel32 == 0 {
            return Err(SpoofError::ModuleNotLoaded(SpoofTag::Kernel32));
        }

        let base_thread_addr =
            dyncvoke_core::get_function_address(kernel32, &lc!("BaseThreadInitThunk"));
        if base_thread_addr == 0 {
            return Err(SpoofError::FunctionNotFound(SpoofTag::BaseThreadInitThunk));
        }

        let k32_table = runtime_function_table(kernel32 as *mut c_void)
            .ok_or(SpoofError::PdataMissing(SpoofTag::Kernel32))?;
        let base_thread_runtime =
            function_by_rva(k32_table, (base_thread_addr - kernel32) as u32)
                .ok_or(SpoofError::UnwindInfoMissing(SpoofTag::BaseThreadInitThunk))?;

        let base_thread_size =
            (base_thread_runtime.EndAddress - base_thread_runtime.BeginAddress) as usize;
        let stack_slot = find_base_thread_return_address(
            kernel32 as *mut c_void,
            base_thread_addr as *mut c_void,
            base_thread_size,
        )
        .ok_or(SpoofError::BaseThreadReturnNotFound)?;
        config.return_address = stack_slot as *const c_void;

        let kb_table = runtime_function_table(kernelbase as *mut c_void)
            .ok_or(SpoofError::PdataMissing(SpoofTag::Kernelbase))?;

        let first_prolog = find_prolog(kernelbase as *mut c_void, kb_table)
            .ok_or(SpoofError::PrologueNotFound)?;
        config.first_frame_fp =
            (first_prolog.frame + first_prolog.offset as u64) as *const c_void;
        config.first_frame_size = first_prolog.stack_size as u64;

        let second_prolog = find_push_rbp(kernelbase as *mut c_void, kb_table)
            .ok_or(SpoofError::PushRbpPrologueNotFound)?;
        config.second_frame_fp =
            (second_prolog.frame + second_prolog.offset as u64) as *const c_void;
        config.second_frame_size = second_prolog.stack_size as u64;
        config.rbp_stack_offset = second_prolog.rbp_offset as u64;

        let (add_rsp_addr, size) = find_gadget(
            kernelbase as *mut c_void,
            &[0x48, 0x83, 0xC4, 0x58, 0xC3],
            kb_table,
            Some(0x58),
        )
        .ok_or(SpoofError::GadgetNotFound(SpoofTag::AddRspGadget))?;
        config.add_rsp_gadget = add_rsp_addr as *const c_void;
        config.add_rsp_frame_size = size as u64;

        let (jmp_rbx_addr, size) =
            find_gadget(kernelbase as *mut c_void, &[0xFF, 0x23], kb_table, None)
                .ok_or(SpoofError::GadgetNotFound(SpoofTag::JmpRbxGadget))?;
        config.jmp_rbx_gadget = jmp_rbx_addr as *const c_void;
        config.jmp_rbx_frame_size = size as u64;

        config.number_args = args.len() as u64;
        write_args(&mut config, args);

        match kind {
            SpoofKind::Function => config.spoof_function = addr as *const c_void,
            SpoofKind::Syscall(name) => {
                let (ssn, syscall_addr) = dyncvoke_core::resolve_syscall(name)
                    .map_err(|_| SpoofError::SsnResolutionFailed)?;
                config.is_syscall = 1;
                config.ssn = ssn as u32;
                config.spoof_function = syscall_addr as *const c_void;
            }
        }

        Ok(Spoof(&mut config))
    }

    #[inline]
    fn write_args(cfg: &mut Config, args: &[*const c_void]) {
        for (i, &a) in args.iter().enumerate() {
            match i {
                0 => cfg.arg01 = a,
                1 => cfg.arg02 = a,
                2 => cfg.arg03 = a,
                3 => cfg.arg04 = a,
                4 => cfg.arg05 = a,
                5 => cfg.arg06 = a,
                6 => cfg.arg07 = a,
                7 => cfg.arg08 = a,
                8 => cfg.arg09 = a,
                9 => cfg.arg10 = a,
                10 => cfg.arg11 = a,
                _ => break,
            }
        }
    }
}

/// Selected decoy frame metadata.
#[derive(Copy, Clone, Default)]
struct Prolog {
    frame: u64,
    stack_size: u32,
    offset: u32,
    rbp_offset: u32,
}

/// Scan kernelbase's runtime function table for a function whose unwind
/// info describes a spoof-compatible RSP-based frame. Shuffles the survivors
/// and returns the first one — picks a different decoy each run.
fn find_prolog(module_base: *mut c_void, runtime_table: &[ImageRuntimeFunction]) -> Option<Prolog> {
    let mut prologs: Vec<Prolog> = runtime_table
        .iter()
        .filter_map(|runtime| {
            let (is_valid, stack_size) = unsafe { stack_frame(module_base, runtime) }?;
            if !is_valid {
                return None;
            }
            let offset = find_valid_instruction_offset(module_base, runtime)?;
            let frame = module_base as u64 + runtime.BeginAddress as u64;
            Some(Prolog {
                frame,
                stack_size,
                offset,
                ..Default::default()
            })
        })
        .collect();

    if prologs.is_empty() {
        return None;
    }
    shuffle(&mut prologs);
    prologs.first().copied()
}

/// Same idea but for an RBP-pushing prologue — these are needed for the
/// inner spoofed frame where the unwinder will reconstruct rbp from the
/// stored slot.
fn find_push_rbp(
    module_base: *mut c_void,
    runtime_table: &[ImageRuntimeFunction],
) -> Option<Prolog> {
    let mut prologs: Vec<Prolog> = runtime_table
        .iter()
        .filter_map(|runtime| {
            let (rbp_off, stack_size) = unsafe { rbp_offset(module_base, runtime) }?;
            if rbp_off == 0 || stack_size == 0 || stack_size <= rbp_off {
                return None;
            }
            let offset = find_valid_instruction_offset(module_base, runtime)?;
            let frame = module_base as u64 + runtime.BeginAddress as u64;
            Some(Prolog {
                frame,
                stack_size,
                offset,
                rbp_offset: rbp_off,
            })
        })
        .collect();

    if prologs.is_empty() {
        return None;
    }
    // First match is consistently unsuitable on most Windows builds; drop it.
    prologs.remove(0);
    if prologs.is_empty() {
        return None;
    }
    shuffle(&mut prologs);
    prologs.first().copied()
}