ralloc @ 867c8090

-Subproject commit 867c809039aef77dbce22e98b1009b8995dfa868

redox-rt/Cargo.toml

+[package]
+name = "redox-rt"
+authors = ["4lDO2 <4lDO2@protonmail.com>"]
+version = "0.1.0"
+edition = "2021"
+license = "MIT"
+description = "Libc-independent runtime for Redox"
+
+# See more keys and their definitions at https://doc.rust-lang.org/cargo/reference/manifest.html
+
+[dependencies]
+bitflags = "2"
+goblin = { version = "0.7", default-features = false, features = ["elf32", "elf64", "endian_fd"] }
+plain = "0.2"
+redox_syscall = "0.5.8"
+
+generic-rt = { path = "../generic-rt" }

redox-rt/src/arch/aarch64.rs

+use core::{mem::offset_of, ptr::NonNull};
+
+use syscall::{data::*, error::*};
+
+use crate::{
+    proc::{fork_inner, FdGuard},
+    signal::{inner_c, PosixStackt, RtSigarea, SigStack, PROC_CONTROL_STRUCT},
+    RtTcb, Tcb,
+};
+
+// Setup a stack starting from the very end of the address space, and then growing downwards.
+pub(crate) const STACK_TOP: usize = 1 << 47;
+pub(crate) const STACK_SIZE: usize = 1024 * 1024;
+
+#[derive(Debug, Default)]
+#[repr(C)]
+pub struct SigArea {
+    pub altstack_top: usize,
+    pub altstack_bottom: usize,
+    pub tmp_x1_x2: [usize; 2],
+    pub tmp_x3_x4: [usize; 2],
+    pub tmp_x5_x6: [usize; 2],
+    pub tmp_sp: usize,
+    pub onstack: u64,
+    pub disable_signals_depth: u64,
+    pub pctl: usize, // TODO: remove
+    pub last_sig_was_restart: bool,
+    pub last_sigstack: Option<NonNull<SigStack>>,
+    pub tmp_rt_inf: RtSigInfo,
+    pub tmp_id_inf: u64,
+}
+#[repr(C)]
+#[derive(Debug, Default)]
+pub struct ArchIntRegs {
+    pub x30: usize,
+    pub x29: usize,
+    pub x28: usize,
+    pub x27: usize,
+    pub x26: usize,
+    pub x25: usize,
+    pub x24: usize,
+    pub x23: usize,
+    pub x22: usize,
+    pub x21: usize,
+    pub x20: usize,
+    pub x19: usize,
+    pub x18: usize,
+    pub x17: usize,
+    pub x16: usize,
+    pub x15: usize,
+    pub x14: usize,
+    pub x13: usize,
+    pub x12: usize,
+    pub x11: usize,
+    pub x10: usize,
+    pub x9: usize,
+    pub x8: usize,
+    pub x7: usize,
+    pub x6: usize,
+    pub x5: usize,
+    pub x4: usize,
+    pub x3: usize,
+    pub x2: usize,
+    pub x1: usize,
+
+    pub sp: usize,
+    pub nzcv: usize, // user-accessible PSTATE bits
+
+    pub pc: usize,
+    pub x0: usize,
+}
+
+/// Deactive TLS, used before exec() on Redox to not trick target executable into thinking TLS
+/// is already initialized as if it was a thread.
+pub unsafe fn deactivate_tcb(open_via_dup: usize) -> Result<()> {
+    let mut env = syscall::EnvRegisters::default();
+
+    let file = FdGuard::new(syscall::dup(open_via_dup, b"regs/env")?);
+
+    env.tpidr_el0 = 0;
+
+    let _ = syscall::write(*file, &mut env)?;
+    Ok(())
+}
+
+pub fn copy_env_regs(cur_pid_fd: usize, new_pid_fd: usize) -> Result<()> {
+    // Copy environment registers.
+    {
+        let cur_env_regs_fd = FdGuard::new(syscall::dup(cur_pid_fd, b"regs/env")?);
+        let new_env_regs_fd = FdGuard::new(syscall::dup(new_pid_fd, b"regs/env")?);
+
+        let mut env_regs = syscall::EnvRegisters::default();
+        let _ = syscall::read(*cur_env_regs_fd, &mut env_regs)?;
+        let _ = syscall::write(*new_env_regs_fd, &env_regs)?;
+    }
+
+    Ok(())
+}
+
+unsafe extern "C" fn fork_impl(initial_rsp: *mut usize) -> usize {
+    Error::mux(fork_inner(initial_rsp))
+}
+
+unsafe extern "C" fn child_hook(cur_filetable_fd: usize, new_pid_fd: usize) {
+    let _ = syscall::close(cur_filetable_fd);
+    // TODO: Currently pidfd == threadfd, but this will not be the case later.
+    RtTcb::current()
+        .thr_fd
+        .get()
+        .write(Some(FdGuard::new(new_pid_fd)));
+}
+
+asmfunction!(__relibc_internal_fork_wrapper -> usize: ["
+    stp     x29, x30, [sp, #-16]!
+    stp     x27, x28, [sp, #-16]!
+    stp     x25, x26, [sp, #-16]!
+    stp     x23, x24, [sp, #-16]!
+    stp     x21, x22, [sp, #-16]!
+    stp     x19, x20, [sp, #-16]!
+
+    sub sp, sp, #32
+
+    //TODO: store floating point regs
+
+    mov x0, sp
+    bl {fork_impl}
+
+    add sp, sp, #32
+    ldp     x19, x20, [sp], #16
+    ldp     x21, x22, [sp], #16
+    ldp     x23, x24, [sp], #16
+    ldp     x25, x26, [sp], #16
+    ldp     x27, x28, [sp], #16
+    ldp     x29, x30, [sp], #16
+    ret
+"] <= [fork_impl = sym fork_impl]);
+
+asmfunction!(__relibc_internal_fork_ret: ["
+    ldp x0, x1, [sp]
+    bl {child_hook}
+
+    //TODO: load floating point regs
+
+    mov x0, xzr
+
+    add sp, sp, #32
+    ldp     x19, x20, [sp], #16
+    ldp     x21, x22, [sp], #16
+    ldp     x23, x24, [sp], #16
+    ldp     x25, x26, [sp], #16
+    ldp     x27, x28, [sp], #16
+    ldp     x29, x30, [sp], #16
+
+    ret
+"] <= [child_hook = sym child_hook]);
+
+// https://devblogs.microsoft.com/oldnewthing/20220811-00/?p=106963
+asmfunction!(__relibc_internal_sigentry: ["
+    // Clear any active reservation.
+    clrex
+
+    // The old pc and x0 are saved in the sigcontrol struct.
+    mrs x0, tpidr_el0 // ABI ptr
+    ldr x0, [x0] // TCB ptr
+
+    // Save x1-x6 and sp
+    stp x1, x2, [x0, #{tcb_sa_off} + {sa_tmp_x1_x2}]
+    stp x3, x4, [x0, #{tcb_sa_off} + {sa_tmp_x3_x4}]
+    stp x5, x6, [x0, #{tcb_sa_off} + {sa_tmp_x5_x6}]
+    mov x1, sp
+    str x1, [x0, #{tcb_sa_off} + {sa_tmp_sp}]
+
+    ldr x6, [x0, #{tcb_sa_off} + {sa_pctl}]
+1:
+    // Load x1 with the thread's bits
+    add x5, x0, #{tcb_sc_off} + {sc_word}
+    ldaxr x1, [x5]
+
+    // First check if there are standard thread signals,
+    and x4, x1, x1, lsr #32 // x4 := x1 & (x1 >> 32)
+    cbnz x4, 3f // jump if x4 != 0
+    clrex
+
+    // and if not, load process pending bitset.
+    add x5, x6, #{pctl_pending}
+    ldaxr x2, [x5]
+
+    // Check if there are standard proc signals:
+    lsr x3, x1, #32 // mask
+    and w3, w2, w3 // pending unblocked proc
+    cbz w3, 4f // skip 'fetch_andn' step if zero
+
+    // If there was one, find which one, and try clearing the bit (last value in x3, addr in x6)
+    // this picks the MSB rather than the LSB, unlike x86. POSIX does not require any specific
+    // ordering though.
+    clz w3, w3
+    mov w4, #31
+    sub w3, w4, w3
+    // x3 now contains the sig_idx
+
+    mov x4, #1
+    lsl x4, x4, x3 // bit to remove
+
+    sub x4, x2, x4 // bit was certainly set, so sub is allowed
+    // x4 is now the new mask to be set
+    add x5, x6, #{pctl_pending}
+
+    add x2, x5, #{pctl_sender_infos}
+    add x2, x2, w3, uxtb 3
+    ldar x2, [x2]
+
+    // Try clearing the bit, retrying on failure.
+    stxr w1, x4, [x5] // try setting pending set to x4, set w1 := 0 on success
+    cbnz w1, 1b // retry everything if this fails
+    mov x1, x3
+    b 2f
+4:
+    // Check for realtime signals, thread/proc.
+    clrex
+
+    // Load the pending set again. TODO: optimize this?
+    add x1, x6, #{pctl_pending}
+    ldaxr x2, [x1]
+    lsr x2, x2, #32
+
+    add x5, x0, #{tcb_sc_off} + {sc_word} + 8
+    ldar x1, [x5]
+
+    orr x2, x1, x2
+    and x2, x2, x2, lsr #32
+    cbz x2, 7f
+
+    rbit x3, x2
+    clz x3, x3
+    mov x4, #31
+    sub x2, x4, x3
+    // x2 now contains sig_idx - 32
+
+    // If realtime signal was directed at thread, handle it as an idempotent signal.
+    lsr x3, x1, x2
+    tbnz x3, #0, 5f
+
+    mov x5, x0
+    mov x4, x8
+    mov x8, #{SYS_SIGDEQUEUE}
+    mov x0, x1
+    add x1, x0, #{tcb_sa_off} + {sa_tmp_rt_inf}
+    svc 0
+    mov x0, x5
+    mov x8, x4
+    cbnz x0, 1b
+
+    b 2f
+5:
+    // A realtime signal was sent to this thread, try clearing its bit.
+    // x3 contains last rt signal word, x2 contains rt_idx
+    clrex
+
+    // Calculate the absolute sig_idx
+    add x1, x3, 32
+
+    // Load si_pid and si_uid
+    add x2, x0, #{tcb_sc_off} + {sc_sender_infos}
+    add x2, x2, w1, uxtb #3
+    ldar x2, [x2]
+
+    add x3, x0, #{tcb_sc_off} + {sc_word} + 8
+    ldxr x2, [x3]
+
+    // Calculate new mask
+    mov x4, #1
+    lsl x4, x4, x2
+    sub x2, x2, x4 // remove bit
+
+    stxr w5, x2, [x3]
+    cbnz w5, 1b
+    str x2, [x0, #{tcb_sa_off} + {sa_tmp_id_inf}]
+    b 2f
+3:
+    // A standard signal was sent to this thread, try clearing its bit.
+    clz x1, x1
+    mov x2, #31
+    sub x1, x2, x1
+
+    // Load si_pid and si_uid
+    add x2, x0, #{tcb_sc_off} + {sc_sender_infos}
+    add x2, x2, w1, uxtb #3
+    ldar x2, [x2]
+
+    // Clear bit from mask
+    mov x3, #1
+    lsl x3, x3, x1
+    sub x4, x4, x3
+
+    // Try updating the mask
+    stxr w3, x1, [x5]
+    cbnz w3, 1b
+
+    str x2, [x0, #{tcb_sa_off} + {sa_tmp_id_inf}]
+2:
+    ldr x3, [x0, #{tcb_sa_off} + {sa_pctl}]
+    add x2, x2, {pctl_actions}
+    add x2, x3, w1, uxtb #4 // actions_base + sig_idx * sizeof Action
+    // TODO: NOT ATOMIC (tearing allowed between regs)!
+    ldxp x2, x3, [x2]
+    clrex
+
+    // Calculate new sp wrt redzone and alignment
+    mov x4, sp
+    sub x4, x4, {REDZONE_SIZE}
+    and x4, x4, -{STACK_ALIGN}
+    mov sp, x4
+
+    // skip sigaltstack step if SA_ONSTACK is clear
+    // tbz x2, #{SA_ONSTACK_BIT}, 2f
+
+    ldr x2, [x0, #{tcb_sc_off} + {sc_saved_pc}]
+    ldr x3, [x0, #{tcb_sc_off} + {sc_saved_x0}]
+    stp x2, x3, [sp, #-16]!
+
+    ldr x2, [x0, #{tcb_sa_off} + {sa_tmp_sp}]
+    mrs x3, nzcv
+    stp x2, x3, [sp, #-16]!
+
+    ldp x2, x3, [x0, #{tcb_sa_off} + {sa_tmp_x1_x2}]
+    stp x2, x3, [sp, #-16]!
+    ldp x3, x4, [x0, #{tcb_sa_off} + {sa_tmp_x3_x4}]
+    stp x4, x3, [sp, #-16]!
+    ldp x5, x6, [x0, #{tcb_sa_off} + {sa_tmp_x5_x6}]
+    stp x6, x5, [sp, #-16]!
+
+    stp x8, x7, [sp, #-16]!
+    stp x10, x9, [sp, #-16]!
+    stp x12, x11, [sp, #-16]!
+    stp x14, x13, [sp, #-16]!
+    stp x16, x15, [sp, #-16]!
+    stp x18, x17, [sp, #-16]!
+    stp x20, x19, [sp, #-16]!
+    stp x22, x21, [sp, #-16]!
+    stp x24, x23, [sp, #-16]!
+    stp x26, x25, [sp, #-16]!
+    stp x28, x27, [sp, #-16]!
+    stp x30, x29, [sp, #-16]!
+
+    str w1, [sp, #-4]
+    sub sp, sp, #64
+
+    mov x0, sp
+    bl {inner}
+
+    add sp, sp, #64
+
+    ldp x30, x29, [sp], #16
+    ldp x28, x27, [sp], #16
+    ldp x26, x25, [sp], #16
+    ldp x24, x23, [sp], #16
+    ldp x22, x21, [sp], #16
+    ldp x20, x19, [sp], #16
+    ldp x18, x17, [sp], #16
+    ldp x16, x15, [sp], #16
+    ldp x14, x13, [sp], #16
+    ldp x12, x11, [sp], #16
+    ldp x10, x9, [sp], #16
+    ldp x8, x7, [sp], #16
+    ldp x6, x5, [sp], #16
+    ldp x4, x3, [sp], #16
+    ldp x2, x1, [sp], #16
+
+    ldr x0, [sp, #8]
+    msr nzcv, x0
+
+8:
+    // x18 is reserved by ABI as 'platform register', so clobbering it should be safe.
+    mov x18, sp
+    ldr x0, [x18]
+    mov sp, x0
+
+    ldp x18, x0, [x18, #16]
+    br x18
+7:
+    // Spurious signal, i.e. all bitsets were 0 at the time they were checked
+    clrex
+
+    ldr x1, [x0, #{tcb_sc_off} + {sc_flags}]
+    and x1, x1, ~1
+    str x1, [x0, #{tcb_sc_off} + {sc_flags}]
+
+    ldp x1, x2, [x0, #{tcb_sa_off} + {sa_tmp_x1_x2}]
+    ldp x3, x4, [x0, #{tcb_sa_off} + {sa_tmp_x3_x4}]
+    ldp x5, x6, [x0, #{tcb_sa_off} + {sa_tmp_x5_x6}]
+    ldr x18, [x0, #{tcb_sc_off} + {sc_saved_pc}]
+    ldr x0, [x0, #{tcb_sc_off} + {sc_saved_x0}]
+    br x18
+"] <= [
+    pctl_pending = const (offset_of!(SigProcControl, pending)),
+    pctl_actions = const (offset_of!(SigProcControl, actions)),
+    pctl_sender_infos = const (offset_of!(SigProcControl, sender_infos)),
+    tcb_sc_off = const (offset_of!(crate::Tcb, os_specific) + offset_of!(RtSigarea, control)),
+    tcb_sa_off = const (offset_of!(crate::Tcb, os_specific) + offset_of!(RtSigarea, arch)),
+    sa_tmp_x1_x2 = const offset_of!(SigArea, tmp_x1_x2),
+    sa_tmp_x3_x4 = const offset_of!(SigArea, tmp_x3_x4),
+    sa_tmp_x5_x6 = const offset_of!(SigArea, tmp_x5_x6),
+    sa_tmp_sp = const offset_of!(SigArea, tmp_sp),
+    sa_tmp_rt_inf = const offset_of!(SigArea, tmp_rt_inf),
+    sa_tmp_id_inf = const offset_of!(SigArea, tmp_id_inf),
+    sa_pctl = const offset_of!(SigArea, pctl),
+    sc_saved_pc = const offset_of!(Sigcontrol, saved_ip),
+    sc_saved_x0 = const offset_of!(Sigcontrol, saved_archdep_reg),
+    sc_sender_infos = const offset_of!(Sigcontrol, sender_infos),
+    sc_word = const offset_of!(Sigcontrol, word),
+    sc_flags = const offset_of!(Sigcontrol, control_flags),
+    inner = sym inner_c,
+
+    SA_ONSTACK_BIT = const 58, // (1 << 58) >> 32 = 0x0400_0000
+    SYS_SIGDEQUEUE = const syscall::SYS_SIGDEQUEUE,
+    STACK_ALIGN = const 16,
+    REDZONE_SIZE = const 128,
+]);
+
+asmfunction!(__relibc_internal_rlct_clone_ret: ["
+    # Load registers
+    ldp x8, x0, [sp], #16
+    ldp x1, x2, [sp], #16
+    ldp x3, x4, [sp], #16
+
+    # Call entry point
+    blr x8
+
+    ret
+"] <= []);
+
+pub fn current_sp() -> usize {
+    let sp: usize;
+    unsafe {
+        core::arch::asm!("mov {}, sp", out(reg) sp);
+    }
+    sp
+}
+
+pub unsafe fn manually_enter_trampoline() {
+    let ctl = &Tcb::current().unwrap().os_specific.control;
+
+    ctl.saved_archdep_reg.set(0);
+    let ip_location = &ctl.saved_ip as *const _ as usize;
+
+    core::arch::asm!("
+        bl 2f
+        b 3f
+    2:
+        str lr, [x0]
+        b __relibc_internal_sigentry
+    3:
+    ", inout("x0") ip_location => _, out("lr") _);
+}
+
+pub unsafe fn arch_pre(stack: &mut SigStack, os: &mut SigArea) -> PosixStackt {
+    PosixStackt {
+        sp: core::ptr::null_mut(), // TODO
+        size: 0,                   // TODO
+        flags: 0,                  // TODO
+    }
+}

redox-rt/src/arch/i686.rs

+use core::{mem::offset_of, ptr::NonNull, sync::atomic::Ordering};
+
+use syscall::*;
+
+use crate::{
+    proc::{fork_inner, FdGuard},
+    signal::{inner_fastcall, PosixStackt, RtSigarea, SigStack, PROC_CONTROL_STRUCT},
+    RtTcb,
+};
+
+// Setup a stack starting from the very end of the address space, and then growing downwards.
+pub(crate) const STACK_TOP: usize = 1 << 31;
+pub(crate) const STACK_SIZE: usize = 1024 * 1024;
+
+#[derive(Debug, Default)]
+#[repr(C)]
+pub struct SigArea {
+    pub altstack_top: usize,
+    pub altstack_bottom: usize,
+    pub tmp_eip: usize,
+    pub tmp_esp: usize,
+    pub tmp_eax: usize,
+    pub tmp_ecx: usize,
+    pub tmp_edx: usize,
+    pub tmp_rt_inf: RtSigInfo,
+    pub tmp_id_inf: u64,
+    pub tmp_mm0: u64,
+    pub pctl: usize, // TODO: reference pctl directly
+    pub disable_signals_depth: u64,
+    pub last_sig_was_restart: bool,
+    pub last_sigstack: Option<NonNull<SigStack>>,
+}
+#[derive(Debug, Default)]
+#[repr(C, align(16))]
+pub struct ArchIntRegs {
+    pub fxsave: [u16; 29],
+
+    // ensure fxsave region is 16 byte aligned
+    pub _pad: [usize; 2], // fxsave "available" +0
+
+    pub ebp: usize, // fxsave "available" +8
+    pub esi: usize, // avail +12
+    pub edi: usize, // avail +16
+    pub ebx: usize, // avail +20
+    pub eax: usize, // avail +24
+    pub ecx: usize, // avail +28
+    pub edx: usize, // avail +32
+
+    pub eflags: usize, // avail +36
+    pub eip: usize,    // avail +40
+    pub esp: usize,    // avail +44
+}
+
+/// Deactive TLS, used before exec() on Redox to not trick target executable into thinking TLS
+/// is already initialized as if it was a thread.
+pub unsafe fn deactivate_tcb(open_via_dup: usize) -> Result<()> {
+    let mut env = syscall::EnvRegisters::default();
+
+    let file = FdGuard::new(syscall::dup(open_via_dup, b"regs/env")?);
+
+    env.fsbase = 0;
+    env.gsbase = 0;
+
+    let _ = syscall::write(*file, &mut env)?;
+    Ok(())
+}
+
+pub fn copy_env_regs(cur_pid_fd: usize, new_pid_fd: usize) -> Result<()> {
+    // Copy environment registers.
+    {
+        let cur_env_regs_fd = FdGuard::new(syscall::dup(cur_pid_fd, b"regs/env")?);
+        let new_env_regs_fd = FdGuard::new(syscall::dup(new_pid_fd, b"regs/env")?);
+
+        let mut env_regs = syscall::EnvRegisters::default();
+        let _ = syscall::read(*cur_env_regs_fd, &mut env_regs)?;
+        let _ = syscall::write(*new_env_regs_fd, &env_regs)?;
+    }
+
+    Ok(())
+}
+
+unsafe extern "cdecl" fn fork_impl(initial_rsp: *mut usize) -> usize {
+    Error::mux(fork_inner(initial_rsp))
+}
+
+unsafe extern "cdecl" fn child_hook(cur_filetable_fd: usize, new_pid_fd: usize) {
+    let _ = syscall::close(cur_filetable_fd);
+    crate::child_hook_common(FdGuard::new(new_pid_fd));
+}
+
+asmfunction!(__relibc_internal_fork_wrapper -> usize: ["
+    push ebp
+    mov ebp, esp
+
+    // Push preserved registers
+    push ebx
+    push edi
+    push esi
+    push ebp
+
+    sub esp, 32
+
+    //TODO stmxcsr [esp+16]
+    fnstcw [esp+24]
+
+    push esp
+    call {fork_impl}
+    pop esp
+    jmp 2f
+"] <= [fork_impl = sym fork_impl]);
+
+asmfunction!(__relibc_internal_fork_ret: ["
+    // Arguments already on the stack
+    call {child_hook}
+
+    //TODO ldmxcsr [esp+16]
+    fldcw [esp+24]
+
+    xor eax, eax
+
+    .p2align 4
+2:
+    add esp, 32
+
+    // Pop preserved registers
+    pop ebp
+    pop esi
+    pop edi
+    pop ebx
+
+    pop ebp
+    ret
+"] <= [child_hook = sym child_hook]);
+asmfunction!(__relibc_internal_sigentry: ["
+    // Save some registers
+    mov gs:[{tcb_sa_off} + {sa_tmp_esp}], esp
+    mov gs:[{tcb_sa_off} + {sa_tmp_eax}], eax
+    mov gs:[{tcb_sa_off} + {sa_tmp_edx}], edx
+    mov gs:[{tcb_sa_off} + {sa_tmp_ecx}], ecx
+1:
+    // Read standard signal word - first for this thread
+    mov edx, gs:[{tcb_sc_off} + {sc_word} + 4]
+    mov eax, gs:[{tcb_sc_off} + {sc_word}]
+    and eax, edx
+    bsf eax, eax
+    jnz 9f
+
+    mov ecx, gs:[{tcb_sa_off} + {sa_pctl}]
+
+    // Read standard signal word - for the process
+    mov eax, [ecx + {pctl_pending}]
+    and eax, edx
+    bsf eax, eax
+    jz 3f
+
+    // Read si_pid and si_uid, atomically.
+    movq gs:[{tcb_sa_off} + {sa_tmp_mm0}], mm0
+    movq mm0, [ecx + {pctl_sender_infos} + eax * 8]
+    movq gs:[{tcb_sa_off} + {sa_tmp_id_inf}], mm0
+    movq mm0, gs:[{tcb_sa_off} + {sa_tmp_mm0}]
+
+    // Try clearing the pending bit, otherwise retry if another thread did that first
+    lock btr [ecx + {pctl_pending}], eax
+    jnc 1b
+    jmp 2f
+3:
+    // Read realtime thread and process signal word together
+    mov edx, [ecx + {pctl_pending} + 4]
+    mov eax, gs:[{tcb_sc_off} + {sc_word} + 8]
+    or eax, edx
+    and eax, gs:[{tcb_sc_off} + {sc_word} + 12]
+    jz 7f // spurious signal
+    bsf eax, eax
+
+    // If thread was specifically targeted, send the signal to it first.
+    bt edx, eax
+    jc 8f
+
+    mov edx, ebx
+    lea ecx, [eax+32]
+    mov eax, {SYS_SIGDEQUEUE}
+    mov edx, gs:[0]
+    add edx, {tcb_sa_off} + {sa_tmp_rt_inf}
+    int 0x80
+    mov ebx, edx
+    test eax, eax
+    jnz 1b
+
+    mov eax, ecx
+    jmp 2f
+8:
+    add eax, 32
+9:
+    // Read si_pid and si_uid, atomically.
+    movq gs:[{tcb_sa_off} + {sa_tmp_mm0}], mm0
+    movq mm0, gs:[{tcb_sc_off} + {sc_sender_infos} + eax * 8]
+    movq gs:[{tcb_sa_off} + {sa_tmp_id_inf}], mm0
+    movq mm0, gs:[{tcb_sa_off} + {sa_tmp_mm0}]
+    mov edx, eax
+    shr edx, 5
+    mov ecx, eax
+    and ecx, 31
+    lock btr gs:[{tcb_sc_off} + {sc_word} + edx * 8], ecx
+
+    add eax, 64
+2:
+    and esp, -{STACK_ALIGN}
+
+    mov edx, eax
+    add edx, edx
+    bt dword ptr [{pctl} + {pctl_actions} + edx * 8 + 4], 28
+    jnc 4f
+
+    mov edx, gs:[{tcb_sa_off} + {sa_altstack_top}]
+    cmp esp, edx
+    ja 3f
+
+    cmp esp, gs:[{tcb_sa_off} + {sa_altstack_bottom}]
+    jnbe 4f
+3:
+    mov esp, edx
+4:
+    // Now that we have a stack, we can finally start populating the signal stack.
+    push dword ptr gs:[{tcb_sa_off} + {sa_tmp_esp}]
+    push dword ptr gs:[{tcb_sc_off} + {sc_saved_eip}]
+    push dword ptr gs:[{tcb_sc_off} + {sc_saved_eflags}]
+
+    push dword ptr gs:[{tcb_sa_off} + {sa_tmp_edx}]
+    push dword ptr gs:[{tcb_sa_off} + {sa_tmp_ecx}]
+    push dword ptr gs:[{tcb_sa_off} + {sa_tmp_eax}]
+    push ebx
+    push edi
+    push esi
+    push ebp
+
+    sub esp, 2 * 4 + 29 * 16
+    fxsave [esp]
+
+    mov [esp - 4], eax
+    sub esp, 48
+
+    mov ecx, esp
+    call {inner}
+
+    fxrstor [esp + 48]
+    add esp, 48 + 29 * 16 + 2 * 4
+
+    pop ebp
+    pop esi
+    pop edi
+    pop ebx
+    pop eax
+    pop ecx
+    pop edx
+
+    popfd
+    pop dword ptr gs:[{tcb_sa_off} + {sa_tmp_eip}]
+
+    .globl __relibc_internal_sigentry_crit_first
+__relibc_internal_sigentry_crit_first:
+    pop esp
+
+    .globl __relibc_internal_sigentry_crit_second
+__relibc_internal_sigentry_crit_second:
+    jmp dword ptr gs:[{tcb_sa_off} + {sa_tmp_eip}]
+7:
+    mov eax, gs:[0]
+    lea esp, [eax + {tcb_sc_off} + {sc_saved_eflags}]
+    popfd
+
+    mov esp, gs:[{tcb_sa_off} + {sa_tmp_esp}]
+
+    mov eax, gs:[{tcb_sc_off} + {sc_saved_eip}]
+    mov gs:[{tcb_sa_off} + {sa_tmp_eip}], eax
+
+    mov eax, gs:[{tcb_sa_off} + {sa_tmp_eax}]
+    mov ecx, gs:[{tcb_sa_off} + {sa_tmp_ecx}]
+    mov edx, gs:[{tcb_sa_off} + {sa_tmp_edx}]
+
+    and dword ptr gs:[{tcb_sc_off} + {sc_control}], ~1
+    .globl __relibc_internal_sigentry_crit_third
+__relibc_internal_sigentry_crit_third:
+    jmp dword ptr gs:[{tcb_sa_off} + {sa_tmp_eip}]
+"] <= [
+    inner = sym inner_fastcall,
+    sa_tmp_eip = const offset_of!(SigArea, tmp_eip),
+    sa_tmp_esp = const offset_of!(SigArea, tmp_esp),
+    sa_tmp_eax = const offset_of!(SigArea, tmp_eax),
+    sa_tmp_ecx = const offset_of!(SigArea, tmp_ecx),
+    sa_tmp_edx = const offset_of!(SigArea, tmp_edx),
+    sa_tmp_mm0 = const offset_of!(SigArea, tmp_mm0),
+    sa_tmp_rt_inf = const offset_of!(SigArea, tmp_rt_inf),
+    sa_tmp_id_inf = const offset_of!(SigArea, tmp_id_inf),
+    sa_altstack_top = const offset_of!(SigArea, altstack_top),
+    sa_altstack_bottom = const offset_of!(SigArea, altstack_bottom),
+    sa_pctl = const offset_of!(SigArea, pctl),
+    sc_control = const offset_of!(Sigcontrol, control_flags),
+    sc_saved_eflags = const offset_of!(Sigcontrol, saved_archdep_reg),
+    sc_saved_eip = const offset_of!(Sigcontrol, saved_ip),
+    sc_word = const offset_of!(Sigcontrol, word),
+    sc_sender_infos = const offset_of!(Sigcontrol, sender_infos),
+    tcb_sa_off = const offset_of!(crate::Tcb, os_specific) + offset_of!(RtSigarea, arch),
+    tcb_sc_off = const offset_of!(crate::Tcb, os_specific) + offset_of!(RtSigarea, control),
+    pctl_actions = const offset_of!(SigProcControl, actions),
+    pctl_sender_infos = const offset_of!(SigProcControl, sender_infos),
+    pctl_pending = const offset_of!(SigProcControl, pending),
+    pctl = sym PROC_CONTROL_STRUCT,
+    STACK_ALIGN = const 16,
+    SYS_SIGDEQUEUE = const syscall::SYS_SIGDEQUEUE,
+]);
+
+asmfunction!(__relibc_internal_rlct_clone_ret -> usize: ["
+    # Load registers
+    pop eax
+
+    sub esp, 8
+
+    mov DWORD PTR [esp], 0x00001F80
+    # TODO: ldmxcsr [esp]
+    mov WORD PTR [esp], 0x037F
+    fldcw [esp]
+
+    add esp, 8
+
+    # Call entry point
+    call eax
+
+    ret
+"] <= []);
+extern "C" {
+    fn __relibc_internal_sigentry_crit_first();
+    fn __relibc_internal_sigentry_crit_second();
+    fn __relibc_internal_sigentry_crit_third();
+}
+pub unsafe fn arch_pre(stack: &mut SigStack, area: &mut SigArea) -> PosixStackt {
+    if stack.regs.eip == __relibc_internal_sigentry_crit_first as usize {
+        let stack_ptr = stack.regs.esp as *const usize;
+        stack.regs.esp = stack_ptr.read();
+        stack.regs.eip = stack_ptr.sub(1).read();
+    } else if stack.regs.eip == __relibc_internal_sigentry_crit_second as usize
+        || stack.regs.eip == __relibc_internal_sigentry_crit_third as usize
+    {
+        stack.regs.eip = area.tmp_eip;
+    }
+    PosixStackt {
+        sp: stack.regs.esp as *mut (),
+        size: 0,  // TODO
+        flags: 0, // TODO
+    }
+}
+#[no_mangle]
+pub unsafe fn manually_enter_trampoline() {
+    let c = &crate::Tcb::current().unwrap().os_specific.control;
+    c.control_flags.store(
+        c.control_flags.load(Ordering::Relaxed) | syscall::flag::INHIBIT_DELIVERY.bits(),
+        Ordering::Release,
+    );
+    c.saved_archdep_reg.set(0); // TODO: Just reset DF on x86?
+
+    core::arch::asm!("
+        call 2f
+        jmp 3f
+    2:
+        pop dword ptr gs:[{tcb_sc_off} + {sc_saved_eip}]
+        jmp __relibc_internal_sigentry
+    3:
+    ",
+        tcb_sc_off = const offset_of!(crate::Tcb, os_specific) + offset_of!(RtSigarea, control),
+        sc_saved_eip = const offset_of!(Sigcontrol, saved_ip),
+    );
+}
+/// Get current stack pointer, weak granularity guarantees.
+pub fn current_sp() -> usize {
+    let sp: usize;
+    unsafe {
+        core::arch::asm!("mov {}, esp", out(reg) sp);
+    }
+    sp
+}

redox-rt/src/arch/mod.rs

+#[cfg(target_arch = "aarch64")]
+pub use self::aarch64::*;
+#[cfg(target_arch = "aarch64")]
+pub mod aarch64;
+
+#[cfg(target_arch = "x86")]
+pub use self::i686::*;
+#[cfg(target_arch = "x86")]
+pub mod i686;
+
+#[cfg(target_arch = "x86_64")]
+pub use self::x86_64::*;
+#[cfg(target_arch = "x86_64")]
+pub mod x86_64;
+
+#[cfg(target_arch = "riscv64")]
+pub use self::riscv64::*;
+#[cfg(target_arch = "riscv64")]
+pub mod riscv64;

redox-rt/src/arch/riscv64.rs

+use crate::{
+    proc::{fork_inner, FdGuard},
+    signal::{get_sigaltstack, inner_c, PosixStackt, RtSigarea, SigStack},
+    RtTcb, Tcb,
+};
+use core::{mem::offset_of, ptr::NonNull, sync::atomic::Ordering};
+use syscall::{data::*, error::*};
+
+// Setup a stack starting from the very end of the address space, and then growing downwards.
+pub(crate) const STACK_TOP: usize = 1 << 47;
+pub(crate) const STACK_SIZE: usize = 1024 * 1024;
+
+#[derive(Debug, Default)]
+#[repr(C)]
+pub struct SigArea {
+    pub tmp_sp: u64,
+    pub tmp_t1: u64,
+    pub tmp_t2: u64,
+    pub tmp_t3: u64,
+    pub tmp_t4: u64,
+    pub tmp_a0: u64,
+    pub tmp_a1: u64,
+    pub tmp_a2: u64,
+    pub tmp_a7: u64,
+
+    pub pctl: usize, // TODO: remove
+    pub tmp_ip: u64,
+    pub tmp_rt_inf: RtSigInfo,
+    pub tmp_id_inf: u64,
+    pub altstack_top: usize,
+    pub altstack_bottom: usize,
+    pub disable_signals_depth: u64,
+    pub last_sig_was_restart: bool,
+    pub last_sigstack: Option<NonNull<SigStack>>,
+}
+#[repr(C)]
+#[derive(Debug, Default)]
+pub struct ArchIntRegs {
+    pub int_regs: [u64; 31],
+    pub pc: u64,
+    pub fp_regs: [u64; 32],
+    pub fcsr: u32,
+    _pad: u32,
+}
+
+/// Deactive TLS, used before exec() on Redox to not trick target executable into thinking TLS
+/// is already initialized as if it was a thread.
+pub unsafe fn deactivate_tcb(open_via_dup: usize) -> Result<()> {
+    let mut env = syscall::EnvRegisters::default();
+
+    let file = FdGuard::new(syscall::dup(open_via_dup, b"regs/env")?);
+
+    env.tp = 0;
+
+    let _ = syscall::write(*file, &mut env)?;
+    Ok(())
+}
+
+pub fn copy_env_regs(cur_pid_fd: usize, new_pid_fd: usize) -> Result<()> {
+    // Copy environment registers.
+    {
+        let cur_env_regs_fd = FdGuard::new(syscall::dup(cur_pid_fd, b"regs/env")?);
+        let new_env_regs_fd = FdGuard::new(syscall::dup(new_pid_fd, b"regs/env")?);
+
+        let mut env_regs = syscall::EnvRegisters::default();
+        let _ = syscall::read(*cur_env_regs_fd, &mut env_regs)?;
+        let _ = syscall::write(*new_env_regs_fd, &env_regs)?;
+    }
+
+    Ok(())
+}
+
+unsafe extern "C" fn fork_impl(initial_rsp: *mut usize) -> usize {
+    Error::mux(fork_inner(initial_rsp))
+}
+
+unsafe extern "C" fn child_hook(cur_filetable_fd: usize, new_pid_fd: usize) {
+    let _ = syscall::close(cur_filetable_fd);
+    crate::child_hook_common(FdGuard::new(new_pid_fd));
+}
+
+asmfunction!(__relibc_internal_fork_wrapper -> usize: ["
+    .attribute arch, \"rv64gc\"  # rust bug 80608
+    addi sp, sp, -200
+    sd   s0, 0(sp)
+    sd   s1, 8(sp)
+    sd   s2, 16(sp)
+    sd   s3, 24(sp)
+    sd   s4, 32(sp)
+    sd   s5, 40(sp)
+    sd   s6, 48(sp)
+    sd   s7, 56(sp)
+    sd   s8, 64(sp)
+    sd   s9, 72(sp)
+    sd   s10, 80(sp)
+    sd   s11, 88(sp)
+    sd   ra, 96(sp)
+
+    fsd  fs0, 104(sp)
+    fsd  fs1, 112(sp)
+    fsd  fs2, 120(sp)
+    fsd  fs3, 128(sp)
+    fsd  fs4, 136(sp)
+    fsd  fs5, 144(sp)
+    fsd  fs6, 152(sp)
+    fsd  fs7, 160(sp)
+    fsd  fs8, 168(sp)
+    fsd  fs9, 176(sp)
+    fsd  fs10, 184(sp)
+    fsd  fs11, 192(sp)
+
+    addi sp, sp, -32
+    mv   a0, sp
+    jal  {fork_impl}
+
+    addi sp, sp, 32
+
+    ld   s0, 0(sp)
+    ld   s1, 8(sp)
+    ld   s2, 16(sp)
+    ld   s3, 24(sp)
+    ld   s4, 32(sp)
+    ld   s5, 40(sp)
+    ld   s6, 48(sp)
+    ld   s7, 56(sp)
+    ld   s8, 64(sp)
+    ld   s9, 72(sp)
+    ld   s10, 80(sp)
+    ld   s11, 88(sp)
+    ld   ra, 96(sp)
+
+    fld  fs0, 104(sp)
+    fld  fs1, 112(sp)
+    fld  fs2, 120(sp)
+    fld  fs3, 128(sp)
+    fld  fs4, 136(sp)
+    fld  fs5, 144(sp)
+    fld  fs6, 152(sp)
+    fld  fs7, 160(sp)
+    fld  fs8, 168(sp)
+    fld  fs9, 176(sp)
+    fld  fs10, 184(sp)
+    fld  fs11, 192(sp)
+
+    addi sp, sp, 200
+    ret
+"] <= [fork_impl = sym fork_impl]);
+
+asmfunction!(__relibc_internal_fork_ret: ["
+    .attribute arch, \"rv64gc\"  # rust bug 80608
+    ld   a0, 0(sp)
+    ld   a1, 8(sp)
+    jal  {child_hook}
+
+    mv   a0, x0
+
+    addi sp, sp, 32
+
+    ld   s0, 0(sp)
+    ld   s1, 8(sp)
+    ld   s2, 16(sp)
+    ld   s3, 24(sp)
+    ld   s4, 32(sp)
+    ld   s5, 40(sp)
+    ld   s6, 48(sp)
+    ld   s7, 56(sp)
+    ld   s8, 64(sp)
+    ld   s9, 72(sp)
+    ld   s10, 80(sp)
+    ld   s11, 88(sp)
+    ld   ra, 96(sp)
+
+    fld  fs0, 104(sp)
+    fld  fs1, 112(sp)
+    fld  fs2, 120(sp)
+    fld  fs3, 128(sp)
+    fld  fs4, 136(sp)
+    fld  fs5, 144(sp)
+    fld  fs6, 152(sp)
+    fld  fs7, 160(sp)
+    fld  fs8, 168(sp)
+    fld  fs9, 176(sp)
+    fld  fs10, 184(sp)
+    fld  fs11, 192(sp)
+
+    addi sp, sp, 200
+    ret
+"] <= [child_hook = sym child_hook]);
+
+asmfunction!(__relibc_internal_sigentry: ["
+    .attribute arch, \"rv64gc\"  # rust bug 80608
+    // Save some registers
+    ld   t0, -8(tp) // Tcb
+    sd   sp, ({tcb_sa_off} + {sa_tmp_sp})(t0)
+    sd   t1, ({tcb_sa_off} + {sa_tmp_t1})(t0)
+    sd   t2, ({tcb_sa_off} + {sa_tmp_t2})(t0)
+    sd   t3, ({tcb_sa_off} + {sa_tmp_t3})(t0)
+    sd   t4, ({tcb_sa_off} + {sa_tmp_t4})(t0)
+    ld   t4, ({tcb_sa_off} + {sa_off_pctl})(t0)
+
+    // First, select signal, always pick first available bit
+99:
+    // Read first signal word
+    ld   t1, ({tcb_sc_off} + {sc_word})(t0)
+    srli t2, t1, 32 // bitset to low word
+    and  t1, t1, t2  // masked bitset in low word
+    beqz t1, 3f
+
+    // Found in first thread signal word
+    mv   t3, x0
+2:  andi t2, t1, 1
+    bnez t2, 10f
+    addi t3, t3, 1
+    srli t1, t1, 1
+    j 2b
+
+    // If no unblocked thread signal was found, check for process.
+    // This is competitive; we need to atomically check if *we* cleared the process-wide pending
+    // bit, otherwise restart.
+3:  lw   t1, {pctl_off_pending}(t4)
+    and  t1, t1, t2
+    beqz t1, 3f
+    // Found in first process signal word
+    li   t3, -1
+2:  andi t2, t1, 1
+    addi t3, t3, 1
+    srli t1, t1, 1
+    beqz t2, 2b
+    slli t1, t3, 3 // * 8 == size_of SenderInfo
+    add  t1, t1, t4
+    ld   t1, {pctl_off_sender_infos}(t1)
+    sd   t1, ({tcb_sa_off} + {sa_tmp_id_inf})(t0)
+    li   t1, 1
+    sll  t1, t1, t3
+    not  t1, t1
+    addi t2, t4, {pctl_off_pending}
+    amoand.w.aq t2, t1, (t2)
+    and  t1, t1, t2
+    bne  t1, t2, 9f
+
+3:
+    // Read second signal word - both process and thread simultaneously.
+    // This must be done since POSIX requires low realtime signals to be picked first.
+    ld   t1, ({tcb_sc_off} + {sc_word} + 8)(t0)
+    lw   t2, ({pctl_off_pending} + 4)(t4)
+    or   t4, t1, t2
+    srli t2, t1, 32
+    and  t4, t2, t4
+    beqz t4, 7f
+    li   t3, -1
+2:  andi t2, t4, 1
+    addi t3, t3, 1
+    srli t4, t4, 1
+    beqz t2, 2b
+    li   t2, 1
+    sll  t2, t2, t3
+    and  t1, t1, t2
+    addi t3, t3, 32
+    bnez t1, 10f // thread signal
+
+    // otherwise, try (competitively) dequeueing realtime signal
+    sd   a0, ({tcb_sa_off} + {sa_tmp_a0})(t0)
+    sd   a1, ({tcb_sa_off} + {sa_tmp_a1})(t0)
+    sd   a2, ({tcb_sa_off} + {sa_tmp_a2})(t0)
+    sd   a7, ({tcb_sa_off} + {sa_tmp_a7})(t0)
+    li   a0, {SYS_SIGDEQUEUE}
+    addi a1, t3, -32
+    add  a2, t0, {tcb_sa_off} + {sa_tmp_rt_inf} // out pointer of dequeued realtime sig
+    ecall
+    bnez a0, 99b // assumes error can only be EAGAIN
+    j   9f
+
+10: // thread signal. t3 holds signal number
+    srli t1, t3, 5
+    bnez t1, 2f // FIXME senderinfo?
+    sll  t2, t3, 3 // * 8 == size_of SenderInfo
+    add  t2, t2, t0
+    ld   t2, ({tcb_sc_off} + {sc_sender_infos})(t2)
+    sd   t2, ({tcb_sa_off} + {sa_tmp_id_inf})(t0)
+2:  andi t4, t3, 31
+    li   t2, 1
+    sll  t2, t2, t4
+    not  t2, t2
+    sll  t1, t1, 3
+    add  t1, t1, t0
+    addi t1, t1, {tcb_sc_off} + {sc_word}
+    amoand.w.aq x0, t2, (t1)
+    addi t3, t3, 64 // indicate signal was targeted at thread
+
+9: // process signal t3 holds signal number
+
+    // By now we have selected a signal, stored in eax (6-bit). We now need to choose whether or
+    // not to switch to the alternate signal stack. If SA_ONSTACK is clear for this signal, then
+    // skip the sigaltstack logic.
+    ld   t4, ({tcb_sa_off} + {sa_off_pctl})(t0)
+    andi t1, t3, 63
+    slli t1, t1, 4 // * 16 == size_of RawAction
+    add  t1, t1, t4
+    ld   t1, {pctl_off_actions}(t1)
+    slli t1, t1, 63-58 // SA_ONSTACK in sign bit
+    bgez t1, 3f
+
+    // If current RSP is above altstack region, switch to altstack
+    ld  t1, ({tcb_sa_off} + {sa_altstack_top})(t0)
+    bgtu sp, t1, 2f
+    ld  t2, ({tcb_sa_off} + {sa_altstack_bottom})(t0)
+    bgtu sp, t3, 3f
+2:  mv  sp, t1
+3:
+    // form mcontext on stack
+    addi sp, sp, -33 * 8
+    fsd  f0, (0 * 8)(sp)
+    fsd  f1, (1 * 8)(sp)
+    fsd  f2, (2 * 8)(sp)
+    fsd  f3, (3 * 8)(sp)
+    fsd  f4, (4 * 8)(sp)
+    fsd  f5, (5 * 8)(sp)
+    fsd  f6, (6 * 8)(sp)
+    fsd  f7, (7 * 8)(sp)
+    fsd  f8, (8 * 8)(sp)
+    fsd  f9, (9 * 8)(sp)
+    fsd  f10, (10 * 8)(sp)
+    fsd  f11, (11 * 8)(sp)
+    fsd  f12, (12 * 8)(sp)
+    fsd  f13, (13 * 8)(sp)
+    fsd  f14, (14 * 8)(sp)
+    fsd  f15, (15 * 8)(sp)
+    fsd  f16, (16 * 8)(sp)
+    fsd  f17, (17 * 8)(sp)
+    fsd  f18, (18 * 8)(sp)
+    fsd  f19, (19 * 8)(sp)
+    fsd  f20, (20 * 8)(sp)
+    fsd  f21, (21 * 8)(sp)
+    fsd  f22, (22 * 8)(sp)
+    fsd  f23, (23 * 8)(sp)
+    fsd  f24, (24 * 8)(sp)
+    fsd  f25, (25 * 8)(sp)
+    fsd  f26, (26 * 8)(sp)
+    fsd  f27, (27 * 8)(sp)
+    fsd  f28, (28 * 8)(sp)
+    fsd  f29, (29 * 8)(sp)
+    fsd  f30, (30 * 8)(sp)
+    fsd  f31, (31 * 8)(sp)
+    csrr t1, fcsr
+    sw   t1, (32 * 8)(sp)
+
+    addi sp, sp, -32 * 8
+    sd   x1, 0(sp)
+    ld   t1, ({tcb_sa_off} + {sa_tmp_sp})(t0)
+    sd   t1, (1 * 8)(sp)  // x2 is sp
+    sd   x3, (2 * 8)(sp)
+    sd   x4, (3 * 8)(sp)
+    ld   t1, ({tcb_sc_off} + {sc_saved_t0})(t0)
+    sd   t1, (4 * 8)(sp) // x5 is t0
+    ld   t1, ({tcb_sa_off} + {sa_tmp_t1})(t0)
+    sd   t1, (5 * 8)(sp) // x6 is t1
+    ld   t1, ({tcb_sa_off} + {sa_tmp_t2})(t0)
+    sd   t1, (6 * 8)(sp) // x7 is t2
+    sd   x8, (7 * 8)(sp)
+    sd   x9, (8 * 8)(sp)
+    sd   x10, (9 * 8)(sp)
+    sd   x11, (10 * 8)(sp)
+    sd   x12, (11 * 8)(sp)
+    sd   x13, (12 * 8)(sp)
+    sd   x14, (13 * 8)(sp)
+    sd   x15, (14 * 8)(sp)
+    sd   x16, (15 * 8)(sp)
+    sd   x17, (16 * 8)(sp)
+    sd   x18, (17 * 8)(sp)
+    sd   x19, (18 * 8)(sp)
+    sd   x20, (19 * 8)(sp)
+    sd   x21, (20 * 8)(sp)
+    sd   x22, (21 * 8)(sp)
+    sd   x23, (22 * 8)(sp)
+    sd   x24, (23 * 8)(sp)
+    sd   x25, (24 * 8)(sp)
+    sd   x26, (25 * 8)(sp)
+    sd   x27, (26 * 8)(sp)
+    ld   t1, ({tcb_sa_off} + {sa_tmp_t3})(t0)
+    sd   t1, (27 * 8)(sp) // t3 is x28
+    ld   t1, ({tcb_sa_off} + {sa_tmp_t4})(t0)
+    sd   t1, (28 * 8)(sp) // t4 is x29
+    sd   x30, (29 * 8)(sp)
+    sd   x31, (30 * 8)(sp)
+    ld   t1, ({tcb_sc_off} + {sc_saved_ip})(t0)
+    sd   t1, (31 * 8)(sp)
+
+    // form ucontext
+    addi sp, sp, -64
+    sw   t3, 60(sp)
+
+    mv   t0, sp
+    jal  {inner}
+
+    addi sp, sp, 64
+
+    addi t0, sp, 32 * 8
+    fld  f0, (0 * 8)(t0)
+    fld  f1, (1 * 8)(t0)
+    fld  f2, (2 * 8)(t0)
+    fld  f3, (3 * 8)(t0)
+    fld  f4, (4 * 8)(t0)
+    fld  f5, (5 * 8)(t0)
+    fld  f6, (6 * 8)(t0)
+    fld  f7, (7 * 8)(t0)
+    fld  f8, (8 * 8)(t0)
+    fld  f9, (9 * 8)(t0)
+    fld  f10, (10 * 8)(t0)
+    fld  f11, (11 * 8)(t0)
+    fld  f12, (12 * 8)(t0)
+    fld  f13, (13 * 8)(t0)
+    fld  f14, (14 * 8)(t0)
+    fld  f15, (15 * 8)(t0)
+    fld  f16, (16 * 8)(t0)
+    fld  f17, (17 * 8)(t0)
+    fld  f18, (18 * 8)(t0)
+    fld  f19, (19 * 8)(t0)
+    fld  f20, (20 * 8)(t0)
+    fld  f21, (21 * 8)(t0)
+    fld  f22, (22 * 8)(t0)
+    fld  f23, (23 * 8)(t0)
+    fld  f24, (24 * 8)(t0)
+    fld  f25, (25 * 8)(t0)
+    fld  f26, (26 * 8)(t0)
+    fld  f27, (27 * 8)(t0)
+    fld  f28, (28 * 8)(t0)
+    fld  f29, (29 * 8)(t0)
+    fld  f30, (30 * 8)(t0)
+    fld  f31, (31 * 8)(t0)
+    lw   t1, (32 * 8)(t0)
+    csrw fcsr, t1
+
+    ld   x1, 0(sp)
+    // skip sp
+    // skip gp
+    ld   x4, (3 * 8)(sp)
+    ld   x5, (4 * 8)(sp)
+    ld   x6, (5 * 8)(sp)
+    ld   x7, (6 * 8)(sp)
+    ld   x8, (7 * 8)(sp)
+    ld   x9, (8 * 8)(sp)
+    ld   x10, (9 * 8)(sp)
+    ld   x11, (10 * 8)(sp)
+    ld   x12, (11 * 8)(sp)
+    ld   x13, (12 * 8)(sp)
+    ld   x14, (13 * 8)(sp)
+    ld   x15, (14 * 8)(sp)
+    ld   x16, (15 * 8)(sp)
+    ld   x17, (16 * 8)(sp)
+    ld   x18, (17 * 8)(sp)
+    ld   x19, (18 * 8)(sp)
+    ld   x20, (19 * 8)(sp)
+    ld   x21, (20 * 8)(sp)
+    ld   x22, (21 * 8)(sp)
+    ld   x23, (22 * 8)(sp)
+    ld   x24, (23 * 8)(sp)
+    ld   x25, (24 * 8)(sp)
+    ld   x26, (25 * 8)(sp)
+    ld   x27, (26 * 8)(sp)
+    ld   x28, (27 * 8)(sp)
+    ld   x29, (28 * 8)(sp)
+    ld   x30, (29 * 8)(sp)
+    ld   x31, (30 * 8)(sp)
+    ld   gp, (31 * 8)(sp) // new IP; this clobbers register x3/gp which is ABI reserved
+    .global __relibc_internal_sigentry_crit_first
+__relibc_internal_sigentry_crit_first:
+    ld   sp, (1 * 8)(sp)
+    .global __relibc_internal_sigentry_crit_second
+__relibc_internal_sigentry_crit_second:
+    jr   gp
+7:
+    // A spurious signal occurred. Signals are still disabled here, but will need to be re-enabled.
+
+    // restore stack
+    ld   sp, ({tcb_sa_off} + {sa_tmp_sp})(t0)
+
+    // move saved IP away from control, allowing arch_pre to save us if interrupted.
+    ld t1, ({tcb_sc_off} + {sc_saved_ip})(t0)
+    sd t1, ({tcb_sa_off} + {sa_tmp_ip})(t0)
+
+    // restore regs
+    ld   t2, ({tcb_sa_off} + {sa_tmp_t2})(t0)
+    ld   t3, ({tcb_sa_off} + {sa_tmp_t3})(t0)
+    ld   t4, ({tcb_sa_off} + {sa_tmp_t4})(t0)
+
+    // move saved t0 away from control as well
+    mv   t1, t0
+    ld   t0, ({tcb_sc_off} + {sc_saved_t0})(t0)
+
+    // Re-enable signals. This code can be interrupted after this signal, so we need to define
+    // 'crit_third'.
+    ld   gp, ({tcb_sc_off} + {sc_control})(t1)
+    andi gp, gp, ~1
+    sd   gp, ({tcb_sc_off} + {sc_control})(t1)
+
+    .globl __relibc_internal_sigentry_crit_third
+__relibc_internal_sigentry_crit_third:
+    ld   gp, ({tcb_sa_off} + {sa_tmp_ip})(t1)
+    .globl __relibc_internal_sigentry_crit_fourth
+__relibc_internal_sigentry_crit_fourth:
+    ld   t1, ({tcb_sa_off} + {sa_tmp_t1})(t1)
+    .globl __relibc_internal_sigentry_crit_fifth
+__relibc_internal_sigentry_crit_fifth:
+    jr   gp
+ "] <= [
+    tcb_sc_off = const (offset_of!(crate::Tcb, os_specific) + offset_of!(RtSigarea, control)),
+    sc_word = const offset_of!(Sigcontrol, word),
+    sc_saved_t0 = const offset_of!(Sigcontrol, saved_archdep_reg),
+    sc_saved_ip = const offset_of!(Sigcontrol, saved_ip),
+    sc_sender_infos = const offset_of!(Sigcontrol, sender_infos),
+    sc_control = const offset_of!(Sigcontrol, control_flags),
+
+    tcb_sa_off = const (offset_of!(crate::Tcb, os_specific) + offset_of!(RtSigarea, arch)),
+    sa_off_pctl = const offset_of!(SigArea, pctl),
+    sa_tmp_sp = const offset_of!(SigArea, tmp_sp),
+    sa_tmp_t1 = const offset_of!(SigArea, tmp_t1),
+    sa_tmp_t2 = const offset_of!(SigArea, tmp_t2),
+    sa_tmp_t3 = const offset_of!(SigArea, tmp_t3),
+    sa_tmp_t4 = const offset_of!(SigArea, tmp_t4),
+    sa_tmp_a0 = const offset_of!(SigArea, tmp_a0),
+    sa_tmp_a1 = const offset_of!(SigArea, tmp_a1),
+    sa_tmp_a2 = const offset_of!(SigArea, tmp_a2),
+    sa_tmp_a7 = const offset_of!(SigArea, tmp_a7),
+    sa_tmp_ip = const offset_of!(SigArea, tmp_ip),
+    sa_tmp_id_inf = const offset_of!(SigArea, tmp_id_inf),
+    sa_tmp_rt_inf = const offset_of!(SigArea, tmp_rt_inf),
+    sa_altstack_top = const offset_of!(SigArea, altstack_top),
+    sa_altstack_bottom = const offset_of!(SigArea, altstack_bottom),
+
+    pctl_off_actions = const offset_of!(SigProcControl, actions),
+    inner = sym inner_c,
+    pctl_off_pending = const offset_of!(SigProcControl, pending),
+    pctl_off_sender_infos = const offset_of!(SigProcControl, sender_infos),
+    SYS_SIGDEQUEUE = const syscall::SYS_SIGDEQUEUE,
+]);
+
+asmfunction!(__relibc_internal_rlct_clone_ret: ["
+    ld t0, 0(sp)
+    ld a0, 8(sp)
+    ld a1, 16(sp)
+    ld a2, 24(sp)
+    ld a3, 32(sp)
+    ld a4, 40(sp)
+    ld a5, 48(sp)
+    addi sp, sp, 56
+
+    jalr t0
+    ret
+"] <= []);
+
+pub fn current_sp() -> usize {
+    let sp: usize;
+    unsafe {
+        core::arch::asm!(
+        "mv {}, sp",
+        out(reg) sp,
+        options(nomem));
+    }
+    sp
+}
+
+pub unsafe fn manually_enter_trampoline() {
+    let ctl = &Tcb::current().unwrap().os_specific.control;
+
+    ctl.control_flags.store(
+        ctl.control_flags.load(Ordering::Relaxed) | syscall::flag::INHIBIT_DELIVERY.bits(),
+        Ordering::Release,
+    );
+    ctl.saved_archdep_reg.set(0);
+    let ip_location = &ctl.saved_ip as *const _ as usize;
+
+    core::arch::asm!("
+        jal 2f
+        j 3f
+    2:
+        sd ra, 0(t0)
+        la t0, __relibc_internal_sigentry
+        jalr x0, t0
+    3:
+    ", inout("t0") ip_location => _, out("ra") _);
+}
+
+extern "C" {
+    fn __relibc_internal_sigentry_crit_first();
+    fn __relibc_internal_sigentry_crit_second();
+    fn __relibc_internal_sigentry_crit_third();
+    fn __relibc_internal_sigentry_crit_fourth();
+    fn __relibc_internal_sigentry_crit_fifth();
+}
+pub unsafe fn arch_pre(stack: &mut SigStack, area: &mut SigArea) -> PosixStackt {
+    // It is impossible to update SP and PC atomically. Instead, we abuse the fact that
+    // signals are disabled in the prologue of the signal trampoline, which allows us to emulate
+    // atomicity inside the critical section, consisting of one instruction at 'crit_first', and
+    // one at 'crit_second', see asm.
+
+    if stack.regs.pc == __relibc_internal_sigentry_crit_first as u64 {
+        // Reexecute 'ld sp, (1 * 8)(sp)'
+        let stack_ptr = stack.regs.int_regs[1] as *const u64; // x2
+        stack.regs.int_regs[1] = stack_ptr.add(1).read();
+        // and 'jr gp' steps.
+        stack.regs.pc = stack.regs.int_regs[2];
+    } else if stack.regs.pc == __relibc_internal_sigentry_crit_second as u64
+        || stack.regs.pc == __relibc_internal_sigentry_crit_fifth as u64
+    {
+        // just reexecute the jump
+        stack.regs.pc = stack.regs.int_regs[2];
+    } else if stack.regs.pc == __relibc_internal_sigentry_crit_third as u64 {
+        // ld   gp, ({tcb_sa_off} + {sa_tmp_ip})(t1)
+        stack.regs.int_regs[2] = area.tmp_ip;
+        // ld   t1, ({tcb_sa_off} + {sa_tmp_t1})(t1)
+        stack.regs.int_regs[5] = area.tmp_t1;
+        // j    gp
+        stack.regs.pc = stack.regs.int_regs[2];
+    } else if stack.regs.pc == __relibc_internal_sigentry_crit_fourth as u64 {
+        // ld   t1, ({tcb_sa_off} + {sa_tmp_t1})(t1)
+        stack.regs.int_regs[5] = area.tmp_t1;
+        // jr   gp
+        stack.regs.pc = stack.regs.int_regs[2];
+    }
+
+    get_sigaltstack(area, stack.regs.int_regs[1] as usize).into()
+}

redox-rt/src/arch/x86_64.rs

+use core::{
+    mem::offset_of,
+    ptr::NonNull,
+    sync::atomic::{AtomicU8, Ordering},
+};
+
+use syscall::{
+    data::{SigProcControl, Sigcontrol},
+    error::*,
+    RtSigInfo,
+};
+
+use crate::{
+    proc::{fork_inner, FdGuard},
+    signal::{get_sigaltstack, inner_c, PosixStackt, RtSigarea, SigStack, PROC_CONTROL_STRUCT},
+    Tcb,
+};
+
+// Setup a stack starting from the very end of the address space, and then growing downwards.
+pub(crate) const STACK_TOP: usize = 1 << 47;
+pub(crate) const STACK_SIZE: usize = 1024 * 1024;
+
+#[derive(Debug, Default)]
+#[repr(C)]
+pub struct SigArea {
+    pub tmp_rip: usize,
+    pub tmp_rsp: usize,
+    pub tmp_rax: usize,
+    pub tmp_rdx: usize,
+    pub tmp_rdi: usize,
+    pub tmp_rsi: usize,
+    pub tmp_rt_inf: RtSigInfo,
+    pub tmp_id_inf: u64,
+
+    pub altstack_top: usize,
+    pub altstack_bottom: usize,
+    pub disable_signals_depth: u64,
+    pub last_sig_was_restart: bool,
+    pub last_sigstack: Option<NonNull<SigStack>>,
+}
+
+#[repr(C, align(16))]
+#[derive(Debug, Default)]
+pub struct ArchIntRegs {
+    pub ymm_upper: [u128; 16],
+    pub fxsave: [u128; 29],
+    pub r15: usize, // fxsave "available" +0
+    pub r14: usize, // available +8
+    pub r13: usize, // available +16
+    pub r12: usize, // available +24
+    pub rbp: usize, // available +32
+    pub rbx: usize, // available +40
+    pub r11: usize, // outside fxsave, and so on
+    pub r10: usize,
+    pub r9: usize,
+    pub r8: usize,
+    pub rax: usize,
+    pub rcx: usize,
+    pub rdx: usize,
+    pub rsi: usize,
+    pub rdi: usize,
+    pub rflags: usize,
+    pub rip: usize,
+    pub rsp: usize,
+}
+
+/// Deactive TLS, used before exec() on Redox to not trick target executable into thinking TLS
+/// is already initialized as if it was a thread.
+pub unsafe fn deactivate_tcb(open_via_dup: usize) -> Result<()> {
+    let mut env = syscall::EnvRegisters::default();
+
+    let file = FdGuard::new(syscall::dup(open_via_dup, b"regs/env")?);
+
+    env.fsbase = 0;
+    env.gsbase = 0;
+
+    let _ = syscall::write(*file, &mut env)?;
+    Ok(())
+}
+
+pub fn copy_env_regs(cur_pid_fd: usize, new_pid_fd: usize) -> Result<()> {
+    // Copy environment registers.
+    {
+        let cur_env_regs_fd = FdGuard::new(syscall::dup(cur_pid_fd, b"regs/env")?);
+        let new_env_regs_fd = FdGuard::new(syscall::dup(new_pid_fd, b"regs/env")?);
+
+        let mut env_regs = syscall::EnvRegisters::default();
+        let _ = syscall::read(*cur_env_regs_fd, &mut env_regs)?;
+        let _ = syscall::write(*new_env_regs_fd, &env_regs)?;
+    }
+
+    Ok(())
+}
+
+unsafe extern "sysv64" fn fork_impl(initial_rsp: *mut usize) -> usize {
+    Error::mux(fork_inner(initial_rsp))
+}
+
+unsafe extern "sysv64" fn child_hook(cur_filetable_fd: usize, new_pid_fd: usize) {
+    let _ = syscall::close(cur_filetable_fd);
+    crate::child_hook_common(FdGuard::new(new_pid_fd));
+}
+
+asmfunction!(__relibc_internal_fork_wrapper -> usize: ["
+    push rbp
+    mov rbp, rsp
+
+    push rbx
+    push rbp
+    push r12
+    push r13
+    push r14
+    push r15
+
+    sub rsp, 32
+
+    stmxcsr [rsp+16]
+    fnstcw [rsp+24]
+
+    mov rdi, rsp
+    call {fork_impl}
+
+    add rsp, 80
+
+    pop rbp
+    ret
+
+"] <= [fork_impl = sym fork_impl]);
+asmfunction!(__relibc_internal_fork_ret: ["
+    mov rdi, [rsp]
+    mov rsi, [rsp + 8]
+    call {child_hook}
+
+    ldmxcsr [rsp + 16]
+    fldcw [rsp + 24]
+
+    xor rax, rax
+
+    add rsp, 32
+    pop r15
+    pop r14
+    pop r13
+    pop r12
+    pop rbp
+    pop rbx
+
+    pop rbp
+    ret
+"] <= [child_hook = sym child_hook]);
+asmfunction!(__relibc_internal_rlct_clone_ret: ["
+    # Load registers
+    pop rax
+    pop rdi
+    pop rsi
+    pop rdx
+    pop rcx
+    pop r8
+    pop r9
+
+    mov DWORD PTR [rsp - 8], 0x00001F80
+    ldmxcsr [rsp - 8]
+    mov WORD PTR [rsp - 8], 0x037F
+    fldcw [rsp - 8]
+
+    # Call entry point
+    call rax
+
+    ret
+"] <= []);
+
+asmfunction!(__relibc_internal_sigentry: ["
+    // Save some registers
+    mov fs:[{tcb_sa_off} + {sa_tmp_rsp}], rsp
+    mov fs:[{tcb_sa_off} + {sa_tmp_rax}], rax
+    mov fs:[{tcb_sa_off} + {sa_tmp_rdx}], rdx
+    mov fs:[{tcb_sa_off} + {sa_tmp_rdi}], rdi
+    mov fs:[{tcb_sa_off} + {sa_tmp_rsi}], rsi
+
+    // First, select signal, always pick first available bit
+1:
+    // Read standard signal word - first targeting this thread
+    mov rax, fs:[{tcb_sc_off} + {sc_word}]
+    mov rdx, rax
+    shr rdx, 32
+    and eax, edx
+    bsf eax, eax
+    jnz 2f
+
+    // If no unblocked thread signal was found, check for process.
+    // This is competitive; we need to atomically check if *we* cleared the process-wide pending
+    // bit, otherwise restart.
+    mov eax, [rip + {pctl} + {pctl_off_pending}]
+    and eax, edx
+    bsf eax, eax
+    jz 8f
+    lea rdi, [rip + {pctl} + {pctl_off_sender_infos}]
+    mov rdi, [rdi + rax * 8]
+    lock btr [rip + {pctl} + {pctl_off_pending}], eax
+    mov fs:[{tcb_sa_off} + {sa_tmp_id_inf}], rdi
+    jc 9f
+8:
+    // Read second signal word - both process and thread simultaneously.
+    // This must be done since POSIX requires low realtime signals to be picked first.
+    mov edx, fs:[{tcb_sc_off} + {sc_word} + 8]
+    mov eax, [rip + {pctl} + {pctl_off_pending} + 4]
+    or eax, edx
+    and eax, fs:[{tcb_sc_off} + {sc_word} + 12]
+    bsf eax, eax
+    jz 7f
+
+    bt edx, eax // check if signal was sent to thread specifically
+    jc 2f // if so, continue as usual
+
+    // otherwise, try (competitively) dequeueing realtime signal
+    mov esi, eax
+    mov eax, {SYS_SIGDEQUEUE}
+    mov rdi, fs:[0]
+    add rdi, {tcb_sa_off} + {sa_tmp_rt_inf} // out pointer of dequeued realtime sig
+    syscall
+    test eax, eax
+    jnz 1b // assumes error can only be EAGAIN
+    lea eax, [esi + 32]
+    jmp 9f
+2:
+    mov edx, eax
+    shr edx, 5
+    mov rdi, fs:[{tcb_sc_off} + {sc_sender_infos} + eax * 8]
+    lock btr fs:[{tcb_sc_off} + {sc_word} + edx * 4], eax
+    mov fs:[{tcb_sa_off} + {sa_tmp_id_inf}], rdi
+    add eax, 64 // indicate signal was targeted at thread
+9:
+    sub rsp, {REDZONE_SIZE}
+    and rsp, -{STACK_ALIGN}
+
+    // By now we have selected a signal, stored in eax (6-bit). We now need to choose whether or
+    // not to switch to the alternate signal stack. If SA_ONSTACK is clear for this signal, then
+    // skip the sigaltstack logic.
+    lea rdx, [rip + {pctl} + {pctl_off_actions}]
+
+    mov ecx, eax
+    and ecx, 63
+
+    // LEA doesn't support 16x, so just do two x8s.
+    lea rdx, [rdx + 8 * rcx]
+    lea rdx, [rdx + 8 * rcx]
+
+    bt qword ptr [rdx], {SA_ONSTACK_BIT}
+    jnc 4f
+
+    // Otherwise, the altstack is already active. The sigaltstack being disabled, is equivalent
+    // to setting 'top' to usize::MAX and 'bottom' to 0.
+
+    // If current RSP is above altstack region, switch to altstack
+    mov rdx, fs:[{tcb_sa_off} + {sa_altstack_top}]
+    cmp rsp, rdx
+    cmova rsp, rdx
+
+    // If current RSP is below altstack region, also switch to altstack
+    cmp rsp, fs:[{tcb_sa_off} + {sa_altstack_bottom}]
+    cmovbe rsp, rdx
+
+    .p2align 4
+4:
+    // Now that we have a stack, we can finally start initializing the signal stack!
+
+    push fs:[{tcb_sa_off} + {sa_tmp_rsp}]
+    push fs:[{tcb_sc_off} + {sc_saved_rip}]
+    push fs:[{tcb_sc_off} + {sc_saved_rflags}]
+
+    push fs:[{tcb_sa_off} + {sa_tmp_rdi}]
+    push fs:[{tcb_sa_off} + {sa_tmp_rsi}]
+    push fs:[{tcb_sa_off} + {sa_tmp_rdx}]
+    push rcx
+    push fs:[{tcb_sa_off} + {sa_tmp_rax}]
+    push r8
+    push r9
+    push r10
+    push r11
+    push rbx
+    push rbp
+    push r12
+    push r13
+    push r14
+    push r15
+    sub rsp, (29 + 16) * 16 // fxsave region minus available bytes
+    fxsave64 [rsp + 16 * 16]
+
+    // TODO: self-modifying?
+    cmp byte ptr [rip + {supports_avx}], 0
+    je 5f
+
+    // Prefer vextractf128 over vextracti128 since the former only requires AVX version 1.
+    vextractf128 [rsp + 15 * 16], ymm0, 1
+    vextractf128 [rsp + 14 * 16], ymm1, 1
+    vextractf128 [rsp + 13 * 16], ymm2, 1
+    vextractf128 [rsp + 12 * 16], ymm3, 1
+    vextractf128 [rsp + 11 * 16], ymm4, 1
+    vextractf128 [rsp + 10 * 16], ymm5, 1
+    vextractf128 [rsp + 9 * 16], ymm6, 1
+    vextractf128 [rsp + 8 * 16], ymm7, 1
+    vextractf128 [rsp + 7 * 16], ymm8, 1
+    vextractf128 [rsp + 6 * 16], ymm9, 1
+    vextractf128 [rsp + 5 * 16], ymm10, 1
+    vextractf128 [rsp + 4 * 16], ymm11, 1
+    vextractf128 [rsp + 3 * 16], ymm12, 1
+    vextractf128 [rsp + 2 * 16], ymm13, 1
+    vextractf128 [rsp + 16], ymm14, 1
+    vextractf128 [rsp], ymm15, 1
+5:
+    mov [rsp - 4], eax
+    sub rsp, 64 // alloc space for ucontext fields
+
+    mov rdi, rsp
+    call {inner}
+
+    add rsp, 64
+
+    fxrstor64 [rsp + 16 * 16]
+
+    cmp byte ptr [rip + {supports_avx}], 0
+    je 6f
+
+    vinsertf128 ymm0, ymm0, [rsp + 15 * 16], 1
+    vinsertf128 ymm1, ymm1, [rsp + 14 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 13 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 12 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 11 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 10 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 9 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 8 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 7 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 6 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 5 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 4 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 3 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 2 * 16], 1
+    vinsertf128 ymm2, ymm2, [rsp + 16], 1
+    vinsertf128 ymm2, ymm2, [rsp], 1
+6:
+    add rsp, (29 + 16) * 16
+
+    pop r15
+    pop r14
+    pop r13
+    pop r12
+    pop rbp
+    pop rbx
+    pop r11
+    pop r10
+    pop r9
+    pop r8
+    pop rax
+    pop rcx
+    pop rdx
+    pop rsi
+    pop rdi
+
+    popfq
+    pop qword ptr fs:[{tcb_sa_off} + {sa_tmp_rip}]
+
+    // x86 lacks atomic instructions for setting both the stack and instruction pointer
+    // simultaneously, except the slow microcoded IRETQ instruction. Thus, we let the arch_pre
+    // function emulate atomicity between the pop rsp and indirect jump.
+
+    .globl __relibc_internal_sigentry_crit_first
+__relibc_internal_sigentry_crit_first:
+
+    pop rsp
+
+    .globl __relibc_internal_sigentry_crit_second
+__relibc_internal_sigentry_crit_second:
+    jmp qword ptr fs:[{tcb_sa_off} + {sa_tmp_rip}]
+7:
+    // A spurious signal occurred. Signals are still disabled here, but will need to be re-enabled.
+
+    // restore flags
+    mov rax, fs:[0] // load FS base
+    // TODO: Use lahf/sahf rather than pushfq/popfq?
+    lea rsp, [rax + {tcb_sc_off} + {sc_saved_rflags}]
+    popfq
+
+    // restore stack
+    mov rsp, fs:[{tcb_sa_off} + {sa_tmp_rsp}]
+
+    // move saved RIP away from control, allowing arch_pre to save us if interrupted.
+    mov rax, fs:[{tcb_sc_off} + {sc_saved_rip}]
+    mov fs:[{tcb_sa_off} + {sa_tmp_rip}], rax
+
+    // restore regs
+    mov rax, fs:[{tcb_sa_off} + {sa_tmp_rax}]
+    mov rdx, fs:[{tcb_sa_off} + {sa_tmp_rdx}]
+
+    // Re-enable signals. This code can be interrupted after this signal, so we need to define
+    // 'crit_third'.
+    and qword ptr fs:[{tcb_sc_off} + {sc_control}], ~1
+
+    .globl __relibc_internal_sigentry_crit_third
+__relibc_internal_sigentry_crit_third:
+    jmp qword ptr fs:[{tcb_sa_off} + {sa_tmp_rip}]
+"] <= [
+    inner = sym inner_c,
+    sa_tmp_rip = const offset_of!(SigArea, tmp_rip),
+    sa_tmp_rsp = const offset_of!(SigArea, tmp_rsp),
+    sa_tmp_rax = const offset_of!(SigArea, tmp_rax),
+    sa_tmp_rdx = const offset_of!(SigArea, tmp_rdx),
+    sa_tmp_rdi = const offset_of!(SigArea, tmp_rdi),
+    sa_tmp_rsi = const offset_of!(SigArea, tmp_rsi),
+    sa_tmp_rt_inf = const offset_of!(SigArea, tmp_rt_inf),
+    sa_tmp_id_inf = const offset_of!(SigArea, tmp_id_inf),
+    sa_altstack_top = const offset_of!(SigArea, altstack_top),
+    sa_altstack_bottom = const offset_of!(SigArea, altstack_bottom),
+    sc_saved_rflags = const offset_of!(Sigcontrol, saved_archdep_reg),
+    sc_saved_rip = const offset_of!(Sigcontrol, saved_ip),
+    sc_word = const offset_of!(Sigcontrol, word),
+    sc_sender_infos = const offset_of!(Sigcontrol, sender_infos),
+    sc_control = const offset_of!(Sigcontrol, control_flags),
+    tcb_sa_off = const offset_of!(crate::Tcb, os_specific) + offset_of!(RtSigarea, arch),
+    tcb_sc_off = const offset_of!(crate::Tcb, os_specific) + offset_of!(RtSigarea, control),
+    pctl_off_actions = const offset_of!(SigProcControl, actions),
+    pctl_off_pending = const offset_of!(SigProcControl, pending),
+    pctl_off_sender_infos = const offset_of!(SigProcControl, sender_infos),
+    pctl = sym PROC_CONTROL_STRUCT,
+    supports_avx = sym SUPPORTS_AVX,
+    REDZONE_SIZE = const 128,
+    STACK_ALIGN = const 16,
+    SA_ONSTACK_BIT = const 58, // (1 << 58) >> 32 = 0x0400_0000
+    SYS_SIGDEQUEUE = const syscall::SYS_SIGDEQUEUE,
+]);
+
+extern "C" {
+    fn __relibc_internal_sigentry_crit_first();
+    fn __relibc_internal_sigentry_crit_second();
+    fn __relibc_internal_sigentry_crit_third();
+}
+/// Fixes some edge cases, and calculates the value for uc_stack.
+pub unsafe fn arch_pre(stack: &mut SigStack, area: &mut SigArea) -> PosixStackt {
+    // It is impossible to update RSP and RIP atomically on x86_64, without using IRETQ, which is
+    // almost as slow as calling a SIGRETURN syscall would be. Instead, we abuse the fact that
+    // signals are disabled in the prologue of the signal trampoline, which allows us to emulate
+    // atomicity inside the critical section, consisting of one instruction at 'crit_first', one at
+    // 'crit_second', and one at 'crit_third', see asm.
+
+    if stack.regs.rip == __relibc_internal_sigentry_crit_first as usize {
+        // Reexecute pop rsp and jump steps. This case needs to be different from the one below,
+        // since rsp has not been overwritten with the previous context's stack, just yet. At this
+        // point, we know [rsp+0] contains the saved RSP, and [rsp-8] contains the saved RIP.
+        let stack_ptr = stack.regs.rsp as *const usize;
+        stack.regs.rsp = stack_ptr.read();
+        stack.regs.rip = stack_ptr.sub(1).read();
+    } else if stack.regs.rip == __relibc_internal_sigentry_crit_second as usize
+        || stack.regs.rip == __relibc_internal_sigentry_crit_third as usize
+    {
+        // Almost finished, just reexecute the jump before tmp_rip is overwritten by this
+        // deeper-level signal.
+        stack.regs.rip = area.tmp_rip;
+    }
+
+    get_sigaltstack(area, stack.regs.rsp).into()
+}
+
+pub(crate) static SUPPORTS_AVX: AtomicU8 = AtomicU8::new(0);
+
+// __relibc will be prepended to the name, so no_mangle is fine
+#[no_mangle]
+pub unsafe fn manually_enter_trampoline() {
+    let c = &Tcb::current().unwrap().os_specific.control;
+    c.control_flags.store(
+        c.control_flags.load(Ordering::Relaxed) | syscall::flag::INHIBIT_DELIVERY.bits(),
+        Ordering::Release,
+    );
+    c.saved_archdep_reg.set(0); // TODO: Just reset DF on x86?
+
+    core::arch::asm!("
+        lea rax, [rip + 2f]
+        mov fs:[{tcb_sc_off} + {sc_saved_rip}], rax
+        jmp __relibc_internal_sigentry
+    2:
+    ",
+        out("rax") _,
+        tcb_sc_off = const offset_of!(crate::Tcb, os_specific) + offset_of!(RtSigarea, control),
+        sc_saved_rip = const offset_of!(Sigcontrol, saved_ip),
+    );
+}
+
+/// Get current stack pointer, weak granularity guarantees.
+pub fn current_sp() -> usize {
+    let sp: usize;
+    unsafe {
+        core::arch::asm!("mov {}, rsp", out(reg) sp);
+    }
+    sp
+}

redox-rt/src/lib.rs

+#![no_std]
+#![feature(
+    asm_const,
+    array_chunks,
+    int_roundings,
+    let_chains,
+    slice_ptr_get,
+    sync_unsafe_cell
+)]
+#![forbid(unreachable_patterns)]
+
+use core::cell::{SyncUnsafeCell, UnsafeCell};
+
+use generic_rt::{ExpectTlsFree, GenericTcb};
+use syscall::{Sigcontrol, O_CLOEXEC};
+
+use self::proc::FdGuard;
+
+extern crate alloc;
+
+#[macro_export]
+macro_rules! asmfunction(
+    ($name:ident $(-> $ret:ty)? : [$($asmstmt:expr),*$(,)?] <= [$($decl:ident = $(sym $symname:ident)?$(const $constval:expr)?),*$(,)?]$(,)? ) => {
+        ::core::arch::global_asm!(concat!("
+            .p2align 4
+            .section .text.", stringify!($name), ", \"ax\", @progbits
+            .globl ", stringify!($name), "
+            .type ", stringify!($name), ", @function
+        ", stringify!($name), ":
+        ", $($asmstmt, "\n",)* "
+            .size ", stringify!($name), ", . - ", stringify!($name), "
+        "), $($decl = $(sym $symname)?$(const $constval)?),*);
+
+        extern "C" {
+            pub fn $name() $(-> $ret)?;
+        }
+    }
+);
+
+pub mod arch;
+pub mod proc;
+
+// TODO: Replace auxvs with a non-stack-based interface, but keep getauxval for compatibility
+#[path = "../../src/platform/auxv_defs.rs"]
+pub mod auxv_defs;
+
+pub mod signal;
+pub mod sync;
+pub mod sys;
+pub mod thread;
+
+#[derive(Debug, Default)]
+pub struct RtTcb {
+    pub control: Sigcontrol,
+    pub arch: UnsafeCell<crate::arch::SigArea>,
+    pub thr_fd: UnsafeCell<Option<FdGuard>>,
+}
+impl RtTcb {
+    pub fn current() -> &'static Self {
+        unsafe { &Tcb::current().unwrap().os_specific }
+    }
+    pub fn thread_fd(&self) -> &FdGuard {
+        unsafe {
+            if (&*self.thr_fd.get()).is_none() {
+                self.thr_fd.get().write(Some(FdGuard::new(
+                    syscall::open("/scheme/thisproc/current/open_via_dup", O_CLOEXEC).unwrap(),
+                )));
+            }
+            (&*self.thr_fd.get()).as_ref().unwrap()
+        }
+    }
+}
+
+pub type Tcb = GenericTcb<RtTcb>;
+
+/// OS and architecture specific code to activate TLS - Redox aarch64
+#[cfg(target_arch = "aarch64")]
+pub unsafe fn tcb_activate(_tcb: &RtTcb, tls_end: usize, tls_len: usize) {
+    // Uses ABI page
+    let abi_ptr = tls_end - tls_len - 16;
+    core::ptr::write(abi_ptr as *mut usize, tls_end);
+    core::arch::asm!(
+        "msr tpidr_el0, {}",
+        in(reg) abi_ptr,
+    );
+}
+
+/// OS and architecture specific code to activate TLS - Redox x86
+#[cfg(target_arch = "x86")]
+pub unsafe fn tcb_activate(tcb: &RtTcb, tls_end: usize, _tls_len: usize) {
+    let mut env = syscall::EnvRegisters::default();
+
+    let file = FdGuard::new(
+        syscall::dup(**tcb.thread_fd(), b"regs/env")
+            .expect_notls("failed to open handle for process registers"),
+    );
+
+    let _ = syscall::read(*file, &mut env).expect_notls("failed to read gsbase");
+
+    env.gsbase = tls_end as u32;
+
+    let _ = syscall::write(*file, &env).expect_notls("failed to write gsbase");
+}
+
+/// OS and architecture specific code to activate TLS - Redox x86_64
+#[cfg(target_arch = "x86_64")]
+pub unsafe fn tcb_activate(tcb: &RtTcb, tls_end_and_tcb_start: usize, _tls_len: usize) {
+    let mut env = syscall::EnvRegisters::default();
+
+    let file = FdGuard::new(
+        syscall::dup(**tcb.thread_fd(), b"regs/env")
+            .expect_notls("failed to open handle for process registers"),
+    );
+
+    let _ = syscall::read(*file, &mut env).expect_notls("failed to read fsbase");
+
+    env.fsbase = tls_end_and_tcb_start as u64;
+
+    let _ = syscall::write(*file, &env).expect_notls("failed to write fsbase");
+}
+
+/// OS and architecture specific code to activate TLS - Redox riscv64
+#[cfg(target_arch = "riscv64")]
+pub unsafe fn tcb_activate(_tcb: &RtTcb, tls_end: usize, tls_len: usize) {
+    // tp points to static tls block
+    // FIXME limited to a single initial master
+    let tls_start = tls_end - tls_len;
+    let abi_ptr = tls_start - 8;
+    core::ptr::write(abi_ptr as *mut usize, tls_end);
+    core::arch::asm!(
+        "mv tp, {}",
+        in(reg) tls_start
+    );
+}
+
+/// Initialize redox-rt in situations where relibc is not used
+pub unsafe fn initialize_freestanding() {
+    // TODO: This code is a hack! Integrate the ld_so TCB code into generic-rt, and then use that
+    // (this function will need pointers to the ELF structs normally passed in auxvs), so the TCB
+    // is initialized properly.
+
+    // TODO: TLS
+    let page = {
+        &mut *(syscall::fmap(
+            !0,
+            &syscall::Map {
+                offset: 0,
+                size: syscall::PAGE_SIZE,
+                flags: syscall::MapFlags::PROT_READ
+                    | syscall::MapFlags::PROT_WRITE
+                    | syscall::MapFlags::MAP_PRIVATE,
+                address: 0,
+            },
+        )
+        .unwrap() as *mut Tcb)
+    };
+    page.tcb_ptr = page;
+    page.tcb_len = syscall::PAGE_SIZE;
+    page.tls_end = (page as *mut Tcb).cast();
+    // Make sure to use ptr::write to prevent dropping the existing FdGuard
+    core::ptr::write(page.os_specific.thr_fd.get(), None);
+
+    #[cfg(not(any(target_arch = "aarch64", target_arch = "riscv64")))]
+    unsafe {
+        let tcb_addr = page as *mut Tcb as usize;
+        tcb_activate(&page.os_specific, tcb_addr, 0)
+    }
+    #[cfg(target_arch = "aarch64")]
+    unsafe {
+        let abi_ptr = core::ptr::addr_of_mut!(page.tcb_ptr);
+        core::arch::asm!("msr tpidr_el0, {}", in(reg) abi_ptr);
+    }
+    #[cfg(target_arch = "riscv64")]
+    unsafe {
+        let abi_ptr = core::ptr::addr_of_mut!(page.tcb_ptr) as usize;
+        core::arch::asm!("mv tp, {}", in(reg) (abi_ptr + 8));
+    }
+    initialize();
+}
+pub unsafe fn initialize() {
+    THIS_PID
+        .get()
+        .write(Some(syscall::getpid().unwrap().try_into().unwrap()).unwrap());
+}
+
+static THIS_PID: SyncUnsafeCell<u32> = SyncUnsafeCell::new(0);
+
+unsafe fn child_hook_common(new_pid_fd: FdGuard) {
+    // TODO: Currently pidfd == threadfd, but this will not be the case later.
+    RtTcb::current().thr_fd.get().write(Some(new_pid_fd));
+    THIS_PID
+        .get()
+        .write(Some(syscall::getpid().unwrap().try_into().unwrap()).unwrap());
+}

redox-rt/src/proc.rs

+use core::{fmt::Debug, mem::size_of};
+
+use crate::{arch::*, auxv_defs::*};
+
+use alloc::{boxed::Box, collections::BTreeMap, vec};
+
+//TODO: allow use of either 32-bit or 64-bit programs
+#[cfg(target_pointer_width = "32")]
+use goblin::elf32::{
+    header::Header,
+    program_header::program_header32::{ProgramHeader, PF_W, PF_X, PT_INTERP, PT_LOAD},
+};
+#[cfg(target_pointer_width = "64")]
+use goblin::elf64::{
+    header::Header,
+    program_header::program_header64::{ProgramHeader, PF_W, PF_X, PT_INTERP, PT_LOAD},
+};
+
+use syscall::{
+    error::*,
+    flag::{MapFlags, SEEK_SET},
+    GrantDesc, GrantFlags, Map, SetSighandlerData, MAP_FIXED_NOREPLACE, MAP_SHARED, O_CLOEXEC,
+    PAGE_SIZE, PROT_EXEC, PROT_READ, PROT_WRITE,
+};
+
+pub enum FexecResult {
+    Normal {
+        addrspace_handle: FdGuard,
+    },
+    Interp {
+        path: Box<[u8]>,
+        image_file: FdGuard,
+        open_via_dup: FdGuard,
+        interp_override: InterpOverride,
+    },
+}
+pub struct InterpOverride {
+    phs: Box<[u8]>,
+    at_entry: usize,
+    at_phnum: usize,
+    at_phent: usize,
+    name: Box<[u8]>,
+    tree: BTreeMap<usize, usize>,
+}
+
+pub struct ExtraInfo<'a> {
+    pub cwd: Option<&'a [u8]>,
+    // Default scheme for the process
+    pub default_scheme: Option<&'a [u8]>,
+    // POSIX states that while sigactions are reset, ignored sigactions will remain ignored.
+    pub sigignmask: u64,
+    // POSIX also states that the sigprocmask must be preserved across execs.
+    pub sigprocmask: u64,
+    /// File mode creation mask (POSIX)
+    pub umask: u32,
+}
+
+pub fn fexec_impl<A, E>(
+    image_file: FdGuard,
+    open_via_dup: FdGuard,
+    memory_scheme_fd: &FdGuard,
+    path: &[u8],
+    args: A,
+    envs: E,
+    total_args_envs_size: usize,
+    extrainfo: &ExtraInfo,
+    mut interp_override: Option<InterpOverride>,
+) -> Result<FexecResult>
+where
+    A: IntoIterator,
+    E: IntoIterator,
+    A::Item: AsRef<[u8]>,
+    E::Item: AsRef<[u8]>,
+{
+    // Here, we do the minimum part of loading an application, which is what the kernel used to do.
+    // We load the executable into memory (albeit at different offsets in this executable), fix
+    // some misalignments, and then execute the SYS_EXEC syscall to replace the program memory
+    // entirely.
+
+    let mut header_bytes = [0_u8; size_of::<Header>()];
+    read_all(*image_file, Some(0), &mut header_bytes)?;
+    let header = Header::from_bytes(&header_bytes);
+
+    let grants_fd = {
+        let current_addrspace_fd = FdGuard::new(syscall::dup(*open_via_dup, b"addrspace")?);
+        FdGuard::new(syscall::dup(*current_addrspace_fd, b"empty")?)
+    };
+
+    // Never allow more than 1 MiB of program headers.
+    const MAX_PH_SIZE: usize = 1024 * 1024;
+    let phentsize = u64::from(header.e_phentsize) as usize;
+    let phnum = u64::from(header.e_phnum) as usize;
+    let pheaders_size = phentsize
+        .saturating_mul(phnum)
+        .saturating_add(size_of::<Header>());
+
+    if pheaders_size > MAX_PH_SIZE {
+        return Err(Error::new(E2BIG));
+    }
+    let mut phs_raw = vec![0_u8; pheaders_size];
+    phs_raw[..size_of::<Header>()].copy_from_slice(&header_bytes);
+    let phs = &mut phs_raw[size_of::<Header>()..];
+
+    // TODO: Remove clone, but this would require more as_refs and as_muts
+    let mut tree = interp_override.as_mut().map_or_else(
+        || core::iter::once((0, PAGE_SIZE)).collect::<BTreeMap<_, _>>(),
+        |o| core::mem::take(&mut o.tree),
+    );
+
+    read_all(*image_file as usize, Some(header.e_phoff as u64), phs)
+        .map_err(|_| Error::new(EIO))?;
+
+    for ph_idx in 0..phnum {
+        let ph_bytes = &phs[ph_idx * phentsize..(ph_idx + 1) * phentsize];
+        let segment: &ProgramHeader =
+            plain::from_bytes(ph_bytes).map_err(|_| Error::new(EINVAL))?;
+        let mut flags = syscall::PROT_READ;
+
+        // W ^ X. If it is executable, do not allow it to be writable, even if requested
+        if segment.p_flags & PF_X == PF_X {
+            flags |= syscall::PROT_EXEC;
+        } else if segment.p_flags & PF_W == PF_W {
+            flags |= syscall::PROT_WRITE;
+        }
+
+        match segment.p_type {
+            // PT_INTERP must come before any PT_LOAD, so we don't have to iterate twice.
+            PT_INTERP => {
+                let mut interp = vec![0_u8; segment.p_filesz as usize];
+                read_all(
+                    *image_file as usize,
+                    Some(segment.p_offset as u64),
+                    &mut interp,
+                )?;
+
+                return Ok(FexecResult::Interp {
+                    path: interp.into_boxed_slice(),
+                    image_file,
+                    open_via_dup,
+                    interp_override: InterpOverride {
+                        at_entry: header.e_entry as usize,
+                        at_phnum: phnum,
+                        at_phent: phentsize,
+                        phs: phs_raw.into_boxed_slice(),
+                        name: path.into(),
+                        tree,
+                    },
+                });
+            }
+            PT_LOAD => {
+                let voff = segment.p_vaddr as usize % PAGE_SIZE;
+                let vaddr = segment.p_vaddr as usize - voff;
+                let filesz = segment.p_filesz as usize;
+
+                let total_page_count = (segment.p_memsz as usize + voff).div_ceil(PAGE_SIZE);
+
+                // The case where segments overlap so that they share one page, is not handled.
+                // TODO: Should it be?
+
+                if segment.p_filesz > segment.p_memsz {
+                    return Err(Error::new(ENOEXEC));
+                }
+
+                allocate_remote(
+                    &grants_fd,
+                    memory_scheme_fd,
+                    vaddr,
+                    total_page_count * PAGE_SIZE,
+                    flags,
+                )?;
+                syscall::lseek(*image_file, segment.p_offset as isize, SEEK_SET)
+                    .map_err(|_| Error::new(EIO))?;
+
+                // If unaligned, read the head page separately.
+                let (first_aligned_page, remaining_filesz) = if voff > 0 {
+                    let bytes_to_next_page = PAGE_SIZE - voff;
+
+                    let (_guard, dst_page) =
+                        unsafe { MmapGuard::map_mut_anywhere(*grants_fd, vaddr, PAGE_SIZE)? };
+
+                    let length = core::cmp::min(bytes_to_next_page, filesz);
+
+                    read_all(*image_file, None, &mut dst_page[voff..][..length])?;
+
+                    (vaddr + PAGE_SIZE, filesz - length)
+                } else {
+                    (vaddr, filesz)
+                };
+
+                let remaining_page_count = remaining_filesz.div_floor(PAGE_SIZE);
+                let tail_bytes = remaining_filesz % PAGE_SIZE;
+
+                // TODO: Unless the calling process if *very* memory-constrained, the max amount of
+                // pages per iteration has no limit other than the time it takes to setup page
+                // tables.
+                //
+                // TODO: Reserve PAGES_PER_ITER "scratch pages" of virtual memory for that type of
+                // situation?
+                const PAGES_PER_ITER: usize = 64;
+
+                // TODO: Before this loop, attempt to mmap with MAP_PRIVATE directly from the image
+                // file.
+
+                for page_idx in (0..remaining_page_count).step_by(PAGES_PER_ITER) {
+                    // Use commented out lines to trigger kernel bug (FIXME).
+
+                    //let pages_in_this_group = core::cmp::min(PAGES_PER_ITER, file_page_count - page_idx * PAGES_PER_ITER);
+                    let pages_in_this_group =
+                        core::cmp::min(PAGES_PER_ITER, remaining_page_count - page_idx);
+
+                    if pages_in_this_group == 0 {
+                        break;
+                    }
+
+                    // TODO: MAP_FIXED to optimize away funmap?
+                    let (_guard, dst_memory) = unsafe {
+                        MmapGuard::map_mut_anywhere(
+                            *grants_fd,
+                            first_aligned_page + page_idx * PAGE_SIZE, // offset
+                            pages_in_this_group * PAGE_SIZE,           // size
+                        )?
+                    };
+
+                    // TODO: Are &mut [u8] and &mut [[u8; PAGE_SIZE]] interchangeable (if the
+                    // lengths are aligned, obviously)?
+
+                    read_all(*image_file, None, dst_memory)?;
+                }
+
+                if tail_bytes > 0 {
+                    let (_guard, dst_page) = unsafe {
+                        MmapGuard::map_mut_anywhere(
+                            *grants_fd,
+                            first_aligned_page + remaining_page_count * PAGE_SIZE,
+                            PAGE_SIZE,
+                        )?
+                    };
+                    read_all(*image_file, None, &mut dst_page[..tail_bytes])?;
+                }
+
+                // file_page_count..file_page_count + zero_page_count are already zero-initialized
+                // by the kernel.
+
+                if !tree
+                    .range(..=vaddr)
+                    .next_back()
+                    .filter(|(start, size)| **start + **size > vaddr)
+                    .is_some()
+                {
+                    tree.insert(vaddr, total_page_count * PAGE_SIZE);
+                }
+            }
+            _ => continue,
+        }
+    }
+
+    allocate_remote(
+        &grants_fd,
+        memory_scheme_fd,
+        STACK_TOP - STACK_SIZE,
+        STACK_SIZE,
+        MapFlags::PROT_READ | MapFlags::PROT_WRITE,
+    )?;
+    tree.insert(STACK_TOP - STACK_SIZE, STACK_SIZE);
+
+    let mut sp = STACK_TOP;
+    let mut stack_page = Option::<MmapGuard>::None;
+
+    let mut push = |word: usize| {
+        let old_page_no = sp / PAGE_SIZE;
+        sp -= size_of::<usize>();
+        let new_page_no = sp / PAGE_SIZE;
+        let new_page_off = sp % PAGE_SIZE;
+
+        let page = if let Some(ref mut page) = stack_page
+            && old_page_no == new_page_no
+        {
+            page
+        } else if let Some(ref mut stack_page) = stack_page {
+            stack_page.remap(new_page_no * PAGE_SIZE, PROT_WRITE)?;
+            stack_page
+        } else {
+            let new = MmapGuard::map(
+                *grants_fd,
+                &Map {
+                    offset: new_page_no * PAGE_SIZE,
+                    size: PAGE_SIZE,
+                    flags: PROT_WRITE,
+                    address: 0, // let kernel decide
+                },
+            )?;
+
+            stack_page.insert(new)
+        };
+
+        unsafe {
+            page.as_mut_ptr_slice()
+                .as_mut_ptr()
+                .add(new_page_off)
+                .cast::<usize>()
+                .write(word);
+        }
+
+        Ok(())
+    };
+
+    let pheaders_to_convey = if let Some(ref r#override) = interp_override {
+        &*r#override.phs
+    } else {
+        &*phs_raw
+    };
+    let pheaders_size_aligned = pheaders_to_convey.len().next_multiple_of(PAGE_SIZE);
+    let pheaders = find_free_target_addr(&tree, pheaders_size_aligned).ok_or(Error::new(ENOMEM))?;
+    tree.insert(pheaders, pheaders_size_aligned);
+    allocate_remote(
+        &grants_fd,
+        memory_scheme_fd,
+        pheaders,
+        pheaders_size_aligned,
+        MapFlags::PROT_READ | MapFlags::PROT_WRITE,
+    )?;
+    unsafe {
+        let (_guard, memory) =
+            MmapGuard::map_mut_anywhere(*grants_fd, pheaders, pheaders_size_aligned)?;
+
+        memory[..pheaders_to_convey.len()].copy_from_slice(pheaders_to_convey);
+    }
+    mprotect_remote(
+        &grants_fd,
+        pheaders,
+        pheaders_size_aligned,
+        MapFlags::PROT_READ,
+    )?;
+
+    push(0)?;
+    push(AT_NULL)?;
+    push(header.e_entry as usize)?;
+    if let Some(ref r#override) = interp_override {
+        push(AT_BASE)?;
+        push(r#override.at_entry)?;
+    }
+    push(AT_ENTRY)?;
+    push(pheaders + size_of::<Header>())?;
+    push(AT_PHDR)?;
+    push(
+        interp_override
+            .as_ref()
+            .map_or(header.e_phnum as usize, |o| o.at_phnum),
+    )?;
+    push(AT_PHNUM)?;
+    push(
+        interp_override
+            .as_ref()
+            .map_or(header.e_phentsize as usize, |o| o.at_phent),
+    )?;
+    push(AT_PHENT)?;
+
+    let total_args_envs_auxvpointee_size = total_args_envs_size
+        + extrainfo.cwd.map_or(0, |s| s.len() + 1)
+        + extrainfo.default_scheme.map_or(0, |s| s.len() + 1);
+    let args_envs_size_aligned = total_args_envs_auxvpointee_size.next_multiple_of(PAGE_SIZE);
+    let target_args_env_address =
+        find_free_target_addr(&tree, args_envs_size_aligned).ok_or(Error::new(ENOMEM))?;
+    allocate_remote(
+        &grants_fd,
+        memory_scheme_fd,
+        target_args_env_address,
+        args_envs_size_aligned,
+        MapFlags::PROT_READ | MapFlags::PROT_WRITE,
+    )?;
+    tree.insert(target_args_env_address, args_envs_size_aligned);
+
+    let mut offset = 0;
+
+    let mut argc = 0;
+
+    {
+        let mut append = |source_slice: &[u8]| {
+            // TODO
+            let address = target_args_env_address + offset;
+
+            if !source_slice.is_empty() {
+                let containing_page = address.div_floor(PAGE_SIZE) * PAGE_SIZE;
+                let displacement = address - containing_page;
+                let size = source_slice.len() + displacement;
+                let aligned_size = size.next_multiple_of(PAGE_SIZE);
+
+                let (_guard, memory) = unsafe {
+                    MmapGuard::map_mut_anywhere(*grants_fd, containing_page, aligned_size)?
+                };
+                memory[displacement..][..source_slice.len()].copy_from_slice(source_slice);
+            }
+
+            offset += source_slice.len() + 1;
+            Ok(address)
+        };
+
+        if let Some(cwd) = extrainfo.cwd {
+            push(append(cwd)?)?;
+            push(AT_REDOX_INITIAL_CWD_PTR)?;
+            push(cwd.len())?;
+            push(AT_REDOX_INITIAL_CWD_LEN)?;
+        }
+
+        if let Some(default_scheme) = extrainfo.default_scheme {
+            push(append(default_scheme)?)?;
+            push(AT_REDOX_INITIAL_DEFAULT_SCHEME_PTR)?;
+            push(default_scheme.len())?;
+            push(AT_REDOX_INITIAL_DEFAULT_SCHEME_LEN)?;
+        }
+
+        #[cfg(target_pointer_width = "32")]
+        {
+            push((extrainfo.sigignmask >> 32) as usize)?;
+            push(AT_REDOX_INHERITED_SIGIGNMASK_HI)?;
+        }
+        push(extrainfo.sigignmask as usize)?;
+        push(AT_REDOX_INHERITED_SIGIGNMASK)?;
+        #[cfg(target_pointer_width = "32")]
+        {
+            push((extrainfo.sigprocmask >> 32) as usize)?;
+            push(AT_REDOX_INHERITED_SIGPROCMASK_HI)?;
+        }
+        push(extrainfo.sigprocmask as usize)?;
+        push(AT_REDOX_INHERITED_SIGPROCMASK)?;
+
+        push(extrainfo.umask as usize)?;
+        push(AT_REDOX_UMASK);
+
+        push(0)?;
+
+        for env in envs {
+            push(append(env.as_ref())?)?;
+        }
+
+        push(0)?;
+
+        for arg in args {
+            push(append(arg.as_ref())?)?;
+            argc += 1;
+        }
+    }
+
+    push(argc)?;
+
+    if let Ok(sighandler_fd) = syscall::dup(*open_via_dup, b"sighandler").map(FdGuard::new) {
+        let _ = syscall::write(
+            *sighandler_fd,
+            &SetSighandlerData {
+                user_handler: 0,
+                excp_handler: 0,
+                thread_control_addr: 0,
+                proc_control_addr: 0,
+            },
+        );
+    }
+
+    unsafe {
+        deactivate_tcb(*open_via_dup)?;
+    }
+
+    // TODO: Restore old name if exec failed?
+    if let Ok(name_fd) = syscall::dup(*open_via_dup, b"name").map(FdGuard::new) {
+        let _ = syscall::write(*name_fd, interp_override.as_ref().map_or(path, |o| &o.name));
+    }
+    if interp_override.is_some() {
+        let mmap_min_fd = FdGuard::new(syscall::dup(*grants_fd, b"mmap-min-addr")?);
+        let last_addr = tree.iter().rev().nth(1).map_or(0, |(off, len)| *off + *len);
+        let aligned_last_addr = last_addr.next_multiple_of(PAGE_SIZE);
+        let _ = syscall::write(*mmap_min_fd, &usize::to_ne_bytes(aligned_last_addr));
+    }
+
+    let addrspace_selection_fd = FdGuard::new(syscall::dup(*open_via_dup, b"current-addrspace")?);
+
+    let _ = syscall::write(
+        *addrspace_selection_fd,
+        &create_set_addr_space_buf(*grants_fd, header.e_entry as usize, sp),
+    );
+
+    Ok(FexecResult::Normal {
+        addrspace_handle: addrspace_selection_fd,
+    })
+}
+fn write_usizes<const N: usize>(fd: &FdGuard, usizes: [usize; N]) -> Result<()> {
+    let _ = syscall::write(**fd, unsafe { plain::as_bytes(&usizes) });
+    Ok(())
+}
+fn allocate_remote(
+    addrspace_fd: &FdGuard,
+    memory_scheme_fd: &FdGuard,
+    dst_addr: usize,
+    len: usize,
+    flags: MapFlags,
+) -> Result<()> {
+    mmap_remote(addrspace_fd, memory_scheme_fd, 0, dst_addr, len, flags)
+}
+pub fn mmap_remote(
+    addrspace_fd: &FdGuard,
+    fd: &FdGuard,
+    offset: usize,
+    dst_addr: usize,
+    len: usize,
+    flags: MapFlags,
+) -> Result<()> {
+    write_usizes(
+        addrspace_fd,
+        [
+            // op
+            syscall::flag::ADDRSPACE_OP_MMAP,
+            // fd
+            **fd,
+            // "offset"
+            offset,
+            // address
+            dst_addr,
+            // size
+            len,
+            // flags
+            (flags | MapFlags::MAP_FIXED_NOREPLACE).bits(),
+        ],
+    )
+}
+pub fn mprotect_remote(
+    addrspace_fd: &FdGuard,
+    addr: usize,
+    len: usize,
+    flags: MapFlags,
+) -> Result<()> {
+    write_usizes(
+        addrspace_fd,
+        [
+            // op
+            syscall::flag::ADDRSPACE_OP_MPROTECT,
+            // address
+            addr,
+            // size
+            len,
+            // flags
+            flags.bits(),
+        ],
+    )
+}
+pub fn munmap_remote(addrspace_fd: &FdGuard, addr: usize, len: usize) -> Result<()> {
+    write_usizes(
+        addrspace_fd,
+        [
+            // op
+            syscall::flag::ADDRSPACE_OP_MUNMAP,
+            // address
+            addr,
+            // size
+            len,
+        ],
+    )
+}
+pub fn munmap_transfer(
+    src: &FdGuard,
+    dst: &FdGuard,
+    src_addr: usize,
+    dst_addr: usize,
+    len: usize,
+    flags: MapFlags,
+) -> Result<()> {
+    write_usizes(
+        dst,
+        [
+            // op
+            syscall::flag::ADDRSPACE_OP_TRANSFER,
+            // fd
+            **src,
+            // "offset" (source address)
+            src_addr,
+            // address
+            dst_addr,
+            // size
+            len,
+            // flags
+            (flags | MapFlags::MAP_FIXED_NOREPLACE).bits(),
+        ],
+    )
+}
+fn read_all(fd: usize, offset: Option<u64>, buf: &mut [u8]) -> Result<()> {
+    if let Some(offset) = offset {
+        syscall::lseek(fd, offset as isize, SEEK_SET)?;
+    }
+
+    let mut total_bytes_read = 0;
+
+    while total_bytes_read < buf.len() {
+        total_bytes_read += match syscall::read(fd, &mut buf[total_bytes_read..])? {
+            0 => return Err(Error::new(ENOEXEC)),
+            bytes_read => bytes_read,
+        }
+    }
+    Ok(())
+}
+
+// TODO: With the introduction of remote mmaps, remove this and let the kernel handle address
+// allocation.
+fn find_free_target_addr(tree: &BTreeMap<usize, usize>, size: usize) -> Option<usize> {
+    let mut iterator = tree.iter().peekable();
+
+    // Ignore the space between zero and the first region, to avoid null pointers.
+    while let Some((cur_address, entry_size)) = iterator.next() {
+        let end = *cur_address + entry_size;
+
+        if let Some((next_address, _)) = iterator.peek() {
+            if **next_address - end > size {
+                return Some(end);
+            }
+        }
+        // No need to check last entry, since the stack will always be put at the highest
+        // possible address.
+    }
+
+    None
+}
+
+pub struct MmapGuard {
+    fd: usize,
+    base: usize,
+    size: usize,
+}
+impl MmapGuard {
+    pub fn map(fd: usize, map: &Map) -> Result<Self> {
+        Ok(Self {
+            fd,
+            size: map.size,
+            base: unsafe { syscall::fmap(fd, map)? },
+        })
+    }
+    pub fn remap(&mut self, offset: usize, mut flags: MapFlags) -> Result<()> {
+        flags.remove(MapFlags::MAP_FIXED_NOREPLACE);
+        flags.insert(MapFlags::MAP_FIXED);
+
+        let _new_base = unsafe {
+            syscall::fmap(
+                self.fd,
+                &Map {
+                    offset,
+                    size: self.size,
+                    flags,
+                    address: self.base,
+                },
+            )?
+        };
+
+        Ok(())
+    }
+    pub unsafe fn map_mut_anywhere<'a>(
+        fd: usize,
+        offset: usize,
+        size: usize,
+    ) -> Result<(Self, &'a mut [u8])> {
+        let mut this = Self::map(
+            fd,
+            &Map {
+                size,
+                offset,
+                address: 0,
+                flags: PROT_WRITE,
+            },
+        )?;
+        let slice = &mut *this.as_mut_ptr_slice();
+
+        Ok((this, slice))
+    }
+    pub fn addr(&self) -> usize {
+        self.base
+    }
+    pub fn len(&self) -> usize {
+        self.size
+    }
+    pub fn as_mut_ptr_slice(&mut self) -> *mut [u8] {
+        core::ptr::slice_from_raw_parts_mut(self.base as *mut u8, self.size)
+    }
+    pub fn take(mut self) {
+        self.size = 0;
+    }
+}
+impl Drop for MmapGuard {
+    fn drop(&mut self) {
+        if self.size != 0 {
+            let _ = unsafe { syscall::funmap(self.base, self.size) };
+        }
+    }
+}
+
+pub struct FdGuard {
+    fd: usize,
+    taken: bool,
+}
+impl FdGuard {
+    pub fn new(fd: usize) -> Self {
+        Self { fd, taken: false }
+    }
+    pub fn take(&mut self) -> usize {
+        self.taken = true;
+        self.fd
+    }
+}
+impl core::ops::Deref for FdGuard {
+    type Target = usize;
+
+    fn deref(&self) -> &Self::Target {
+        &self.fd
+    }
+}
+
+impl Drop for FdGuard {
+    fn drop(&mut self) {
+        if !self.taken {
+            let _ = syscall::close(self.fd);
+        }
+    }
+}
+impl Debug for FdGuard {
+    fn fmt(&self, f: &mut core::fmt::Formatter<'_>) -> core::fmt::Result {
+        write!(f, "[fd {}]", self.fd)
+    }
+}
+pub fn create_set_addr_space_buf(
+    space: usize,
+    ip: usize,
+    sp: usize,
+) -> [u8; size_of::<usize>() * 3] {
+    let mut buf = [0_u8; 3 * size_of::<usize>()];
+    let mut chunks = buf.array_chunks_mut::<{ size_of::<usize>() }>();
+    *chunks.next().unwrap() = usize::to_ne_bytes(space);
+    *chunks.next().unwrap() = usize::to_ne_bytes(sp);
+    *chunks.next().unwrap() = usize::to_ne_bytes(ip);
+    buf
+}
+
+/// Spawns a new context which will not share the same address space as the current one. File
+/// descriptors from other schemes are reobtained with `dup`, and grants referencing such file
+/// descriptors are reobtained through `fmap`. Other mappings are kept but duplicated using CoW.
+pub fn fork_impl() -> Result<usize> {
+    let old_mask = crate::signal::get_sigmask()?;
+    let pid = unsafe { Error::demux(__relibc_internal_fork_wrapper())? };
+
+    if pid == 0 {
+        crate::signal::set_sigmask(Some(old_mask), None)?;
+    }
+    Ok(pid)
+}
+
+pub fn fork_inner(initial_rsp: *mut usize) -> Result<usize> {
+    let (cur_filetable_fd, new_pid_fd, new_pid);
+
+    {
+        let cur_pid_fd = FdGuard::new(syscall::open(
+            "/scheme/thisproc/current/open_via_dup",
+            O_CLOEXEC,
+        )?);
+        (new_pid_fd, new_pid) = new_child_process()?;
+
+        copy_str(*cur_pid_fd, *new_pid_fd, "name")?;
+
+        // Copy existing files into new file table, but do not reuse the same file table (i.e. new
+        // parent FDs will not show up for the child).
+        {
+            cur_filetable_fd = FdGuard::new(syscall::dup(*cur_pid_fd, b"filetable")?);
+
+            // This must be done before the address space is copied.
+            unsafe {
+                initial_rsp.write(*cur_filetable_fd);
+                initial_rsp.add(1).write(*new_pid_fd);
+            }
+        }
+
+        // CoW-duplicate address space.
+        {
+            let new_addr_space_sel_fd =
+                FdGuard::new(syscall::dup(*new_pid_fd, b"current-addrspace")?);
+
+            let cur_addr_space_fd = FdGuard::new(syscall::dup(*cur_pid_fd, b"addrspace")?);
+            let new_addr_space_fd = FdGuard::new(syscall::dup(*cur_addr_space_fd, b"exclusive")?);
+
+            let mut grant_desc_buf = [GrantDesc::default(); 16];
+            loop {
+                let bytes_read = {
+                    let buf = unsafe {
+                        core::slice::from_raw_parts_mut(
+                            grant_desc_buf.as_mut_ptr().cast(),
+                            grant_desc_buf.len() * size_of::<GrantDesc>(),
+                        )
+                    };
+                    syscall::read(*cur_addr_space_fd, buf)?
+                };
+                if bytes_read == 0 {
+                    break;
+                }
+
+                let grants = &grant_desc_buf[..bytes_read / size_of::<GrantDesc>()];
+
+                for grant in grants {
+                    if !grant.flags.contains(GrantFlags::GRANT_SCHEME)
+                        || !grant.flags.contains(GrantFlags::GRANT_SHARED)
+                    {
+                        continue;
+                    }
+
+                    let buf;
+
+                    // TODO: write! using some #![no_std] Cursor type (tracking the length)?
+                    #[cfg(target_pointer_width = "64")]
+                    {
+                        //buf = *b"grant-fd-AAAABBBBCCCCDDDD";
+                        //write!(&mut buf, "grant-fd-{:>016x}", grant.base).unwrap();
+                        buf = alloc::format!("grant-fd-{:>016x}", grant.base).into_bytes();
+                    }
+
+                    #[cfg(target_pointer_width = "32")]
+                    {
+                        //buf = *b"grant-fd-AAAABBBB";
+                        //write!(&mut buf[..], "grant-fd-{:>08x}", grant.base).unwrap();
+                        buf = alloc::format!("grant-fd-{:>08x}", grant.base).into_bytes();
+                    }
+
+                    let grant_fd = FdGuard::new(syscall::dup(*cur_addr_space_fd, &buf)?);
+
+                    let mut flags = MAP_SHARED | MAP_FIXED_NOREPLACE;
+
+                    flags.set(PROT_READ, grant.flags.contains(GrantFlags::GRANT_READ));
+                    flags.set(PROT_WRITE, grant.flags.contains(GrantFlags::GRANT_WRITE));
+                    flags.set(PROT_EXEC, grant.flags.contains(GrantFlags::GRANT_EXEC));
+
+                    mmap_remote(
+                        &new_addr_space_fd,
+                        &grant_fd,
+                        grant.offset as usize,
+                        grant.base,
+                        grant.size,
+                        flags,
+                    )?;
+                }
+            }
+
+            let buf = create_set_addr_space_buf(
+                *new_addr_space_fd,
+                __relibc_internal_fork_ret as usize,
+                initial_rsp as usize,
+            );
+            let _ = syscall::write(*new_addr_space_sel_fd, &buf)?;
+        }
+        {
+            // Reuse the same sigaltstack and signal entry (all memory will be re-mapped CoW later).
+            //
+            // Do this after the address space is cloned, since the kernel will get a shared
+            // reference to the TCB and whatever pages stores the signal proc control struct.
+            {
+                let new_sighandler_fd = FdGuard::new(syscall::dup(*new_pid_fd, b"sighandler")?);
+                let _ = syscall::write(
+                    *new_sighandler_fd,
+                    &crate::signal::current_setsighandler_struct(),
+                )?;
+            }
+        }
+        copy_env_regs(*cur_pid_fd, *new_pid_fd)?;
+    }
+    // Copy the file table. We do this last to ensure that all previously used file descriptors are
+    // closed. The only exception -- the filetable selection fd and the current filetable fd --
+    // will be closed by the child process.
+    {
+        // TODO: Use file descriptor forwarding or something similar to avoid copying the file
+        // table in the kernel.
+        let new_filetable_fd = FdGuard::new(syscall::dup(*cur_filetable_fd, b"copy")?);
+        let new_filetable_sel_fd = FdGuard::new(syscall::dup(*new_pid_fd, b"current-filetable")?);
+        let _ = syscall::write(
+            *new_filetable_sel_fd,
+            &usize::to_ne_bytes(*new_filetable_fd),
+        )?;
+    }
+    let start_fd = FdGuard::new(syscall::dup(*new_pid_fd, b"start")?);
+    let _ = syscall::write(*start_fd, &[0])?;
+
+    Ok(new_pid)
+}
+
+pub fn new_child_process() -> Result<(FdGuard, usize)> {
+    // Create a new context (fields such as uid/gid will be inherited from the current context).
+    let fd = FdGuard::new(syscall::open(
+        "/scheme/thisproc/new/open_via_dup",
+        O_CLOEXEC,
+    )?);
+
+    // Extract pid.
+    let mut buffer = [0_u8; 64];
+    let len = syscall::fpath(*fd, &mut buffer)?;
+    let buffer = buffer.get(..len).ok_or(Error::new(ENAMETOOLONG))?;
+
+    let colon_idx = buffer
+        .iter()
+        .position(|c| *c == b':')
+        .ok_or(Error::new(EINVAL))?;
+    let slash_idx = buffer
+        .iter()
+        .skip(colon_idx)
+        .position(|c| *c == b'/')
+        .ok_or(Error::new(EINVAL))?
+        + colon_idx;
+    let pid_bytes = buffer
+        .get(colon_idx + 1..slash_idx)
+        .ok_or(Error::new(EINVAL))?;
+    let pid_str = core::str::from_utf8(pid_bytes).map_err(|_| Error::new(EINVAL))?;
+    let pid = pid_str.parse::<usize>().map_err(|_| Error::new(EINVAL))?;
+
+    Ok((fd, pid))
+}
+
+pub fn copy_str(cur_pid_fd: usize, new_pid_fd: usize, key: &str) -> Result<()> {
+    let cur_name_fd = FdGuard::new(syscall::dup(cur_pid_fd, key.as_bytes())?);
+    let new_name_fd = FdGuard::new(syscall::dup(new_pid_fd, key.as_bytes())?);
+
+    // TODO: Max path size?
+    let mut buf = [0_u8; 256];
+    let len = syscall::read(*cur_name_fd, &mut buf)?;
+    let buf = buf.get(..len).ok_or(Error::new(ENAMETOOLONG))?;
+
+    syscall::write(*new_name_fd, &buf)?;
+
+    Ok(())
+}

redox-rt/src/signal.rs

+use core::{ffi::c_int, mem::MaybeUninit, ptr::NonNull, sync::atomic::Ordering};
+
+use syscall::{
+    data::AtomicU64, Error, RawAction, Result, RtSigInfo, SenderInfo, SetSighandlerData,
+    SigProcControl, Sigcontrol, SigcontrolFlags, TimeSpec, EAGAIN, EINTR, EINVAL, ENOMEM, EPERM,
+    SIGCHLD, SIGKILL, SIGSTOP, SIGTSTP, SIGTTIN, SIGTTOU, SIGURG, SIGWINCH,
+};
+
+use crate::{arch::*, proc::FdGuard, sync::Mutex, RtTcb, Tcb};
+
+#[cfg(target_arch = "x86_64")]
+static CPUID_EAX1_ECX: core::sync::atomic::AtomicU32 = core::sync::atomic::AtomicU32::new(0);
+
+pub fn sighandler_function() -> usize {
+    // TODO: HWCAP?
+
+    __relibc_internal_sigentry as usize
+}
+
+/// ucontext_t representation
+#[repr(C)]
+pub struct SigStack {
+    #[cfg(any(
+        target_arch = "x86_64",
+        target_arch = "aarch64",
+        target_arch = "riscv64"
+    ))]
+    _pad: [usize; 1], // pad from 7*8 to 64
+
+    #[cfg(target_arch = "x86")]
+    _pad: [usize; 3], // pad from 9*4 to 12*4
+
+    pub link: *mut SigStack,
+
+    pub old_stack: PosixStackt,
+    pub old_mask: u64,
+    pub(crate) sival: usize,
+    pub(crate) sig_code: u32,
+    pub(crate) sig_num: u32,
+
+    // x86_64: 864 bytes
+    // i686: 512 bytes
+    // aarch64: 272 bytes (SIMD TODO)
+    // riscv64: 520 bytes (vector extensions TODO)
+    pub regs: ArchIntRegs,
+}
+#[repr(C)]
+pub struct PosixStackt {
+    pub sp: *mut (),
+    pub flags: i32,
+    pub size: usize,
+}
+
+pub const SS_ONSTACK: usize = 1;
+pub const SS_DISABLE: usize = 2;
+
+impl From<Sigaltstack> for PosixStackt {
+    fn from(value: Sigaltstack) -> Self {
+        match value {
+            Sigaltstack::Disabled => PosixStackt {
+                sp: core::ptr::null_mut(),
+                size: 0,
+                flags: SS_DISABLE.try_into().unwrap(),
+            },
+            Sigaltstack::Enabled {
+                onstack,
+                base,
+                size,
+            } => PosixStackt {
+                sp: base.cast(),
+                size,
+                flags: if onstack {
+                    SS_ONSTACK.try_into().unwrap()
+                } else {
+                    0
+                },
+            },
+        }
+    }
+}
+
+#[repr(C)]
+// TODO: This struct is for practical reasons locked to Linux's ABI, but avoid redefining
+// it here. Alternatively, check at compile time that the structs are equivalent.
+pub struct SiginfoAbi {
+    pub si_signo: i32,
+    pub si_errno: i32,
+    pub si_code: i32,
+    pub si_pid: i32,      // pid_t
+    pub si_uid: i32,      // uid_t
+    pub si_addr: *mut (), // *mut c_void
+    pub si_status: i32,
+    pub si_value: usize, // sigval
+}
+
+#[inline(always)]
+unsafe fn inner(stack: &mut SigStack) {
+    let os = &Tcb::current().unwrap().os_specific;
+
+    let stack_ptr = NonNull::from(&mut *stack);
+    stack.link = core::mem::replace(&mut (*os.arch.get()).last_sigstack, Some(stack_ptr))
+        .map_or_else(core::ptr::null_mut, |x| x.as_ptr());
+
+    let signals_were_disabled = (*os.arch.get()).disable_signals_depth > 0;
+
+    let targeted_thread_not_process = stack.sig_num >= 64;
+    stack.sig_num %= 64;
+
+    // asm counts from 0
+    stack.sig_num += 1;
+
+    let (sender_pid, sender_uid) = {
+        let area = &mut *os.arch.get();
+
+        // Undefined if the signal was not realtime
+        stack.sival = area.tmp_rt_inf.arg;
+
+        stack.old_stack = arch_pre(stack, area);
+
+        if (stack.sig_num - 1) / 32 == 1 && !targeted_thread_not_process {
+            stack.sig_code = area.tmp_rt_inf.code as u32;
+            (area.tmp_rt_inf.pid, area.tmp_rt_inf.uid)
+        } else {
+            stack.sig_code = 0; // TODO: SI_USER constant?
+                                // TODO: Handle SIGCHLD. Maybe that should always be queued though?
+            let inf = SenderInfo::from_raw(area.tmp_id_inf);
+            (inf.pid, inf.ruid)
+        }
+    };
+
+    let sigaction = {
+        let guard = SIGACTIONS_LOCK.lock();
+        let action = convert_old(&PROC_CONTROL_STRUCT.actions[stack.sig_num as usize - 1]);
+        if action.flags.contains(SigactionFlags::RESETHAND) {
+            drop(guard);
+            sigaction(
+                stack.sig_num as u8,
+                Some(&Sigaction {
+                    kind: SigactionKind::Default,
+                    mask: 0,
+                    flags: SigactionFlags::empty(),
+                }),
+                None,
+            );
+        }
+        action
+    };
+    let shall_restart = sigaction.flags.contains(SigactionFlags::RESTART);
+
+    let handler = match sigaction.kind {
+        SigactionKind::Ignore => {
+            panic!("ctl {:x?} signal {}", os.control, stack.sig_num)
+        }
+        SigactionKind::Default => {
+            syscall::exit(stack.sig_num as usize);
+            unreachable!();
+        }
+        SigactionKind::Handled { handler } => handler,
+    };
+
+    // Set sigmask to sa_mask and unmark the signal as pending.
+    let prev_sigallow = get_allowset_raw(&os.control.word);
+
+    let mut sigallow_inside = !sigaction.mask & prev_sigallow;
+    if !sigaction.flags.contains(SigactionFlags::NODEFER) {
+        sigallow_inside &= !sig_bit(stack.sig_num);
+    }
+
+    let _pending_when_sa_mask = set_allowset_raw(&os.control.word, prev_sigallow, sigallow_inside);
+
+    // TODO: If sa_mask caused signals to be unblocked, deliver one or all of those first?
+
+    // Re-enable signals again.
+    let control_flags = &os.control.control_flags;
+    control_flags.store(
+        control_flags.load(Ordering::Relaxed) & !SigcontrolFlags::INHIBIT_DELIVERY.bits(),
+        Ordering::Release,
+    );
+    core::sync::atomic::compiler_fence(Ordering::Acquire);
+
+    stack.old_mask = prev_sigallow;
+
+    // Call handler, either sa_handler or sa_siginfo depending on flag.
+    if sigaction.flags.contains(SigactionFlags::SIGINFO)
+        && let Some(sigaction) = handler.sigaction
+    {
+        let info = SiginfoAbi {
+            si_signo: stack.sig_num as c_int,
+            si_addr: core::ptr::null_mut(),
+            si_code: stack.sig_code as i32,
+            si_errno: 0,
+            si_pid: sender_pid as i32,
+            si_status: 0,
+            si_uid: sender_uid as i32,
+            si_value: stack.sival,
+        };
+        sigaction(
+            stack.sig_num as c_int,
+            core::ptr::addr_of!(info).cast(),
+            stack as *mut SigStack as *mut (),
+        );
+    } else if let Some(handler) = handler.handler {
+        handler(stack.sig_num as c_int);
+    }
+
+    // Disable signals while we modify the sigmask again
+    control_flags.store(
+        control_flags.load(Ordering::Relaxed) | SigcontrolFlags::INHIBIT_DELIVERY.bits(),
+        Ordering::Release,
+    );
+    core::sync::atomic::compiler_fence(Ordering::Acquire);
+
+    // Update allowset again.
+
+    let new_mask = stack.old_mask;
+    let old_mask = get_allowset_raw(&os.control.word);
+
+    let _pending_when_restored_mask = set_allowset_raw(&os.control.word, old_mask, new_mask);
+
+    // TODO: If resetting the sigmask caused signals to be unblocked, then should they be delivered
+    // here? And would it be possible to tail-call-optimize that?
+
+    (*os.arch.get()).last_sig_was_restart = shall_restart;
+
+    // TODO: Support setting uc_link to jump back to a different context?
+    (*os.arch.get()).last_sigstack = NonNull::new(stack.link);
+
+    // TODO: Support restoring uc_stack?
+
+    // And re-enable them again
+    if !signals_were_disabled {
+        core::sync::atomic::compiler_fence(Ordering::Release);
+        control_flags.store(
+            control_flags.load(Ordering::Relaxed) & !SigcontrolFlags::INHIBIT_DELIVERY.bits(),
+            Ordering::Relaxed,
+        );
+    }
+}
+#[cfg(not(target_arch = "x86"))]
+pub(crate) unsafe extern "C" fn inner_c(stack: usize) {
+    inner(&mut *(stack as *mut SigStack))
+}
+#[cfg(target_arch = "x86")]
+pub(crate) unsafe extern "fastcall" fn inner_fastcall(stack: usize) {
+    inner(&mut *(stack as *mut SigStack))
+}
+
+pub fn get_sigmask() -> Result<u64> {
+    let mut mask = 0;
+    modify_sigmask(Some(&mut mask), Option::<fn(u64) -> u64>::None)?;
+    Ok(mask)
+}
+pub fn set_sigmask(new: Option<u64>, old: Option<&mut u64>) -> Result<()> {
+    modify_sigmask(old, new.map(move |newmask| move |_| newmask))
+}
+pub fn or_sigmask(new: Option<u64>, old: Option<&mut u64>) -> Result<()> {
+    // Parsing nightmare... :)
+    modify_sigmask(
+        old,
+        new.map(move |newmask| move |oldmask| oldmask | newmask),
+    )
+}
+pub fn andn_sigmask(new: Option<u64>, old: Option<&mut u64>) -> Result<()> {
+    modify_sigmask(
+        old,
+        new.map(move |newmask| move |oldmask| oldmask & !newmask),
+    )
+}
+fn get_allowset_raw(words: &[AtomicU64; 2]) -> u64 {
+    (words[0].load(Ordering::Relaxed) >> 32) | ((words[1].load(Ordering::Relaxed) >> 32) << 32)
+}
+/// Sets mask from old to new, returning what was pending at the time.
+fn set_allowset_raw(words: &[AtomicU64; 2], old: u64, new: u64) -> u64 {
+    // This assumes *only this thread* can change the allowset. If this rule is broken, the use of
+    // fetch_add will corrupt the words entirely. fetch_add is very efficient on x86, being
+    // generated as LOCK XADD which is the fastest RMW instruction AFAIK.
+    let prev_w0 = words[0].fetch_add(
+        ((new & 0xffff_ffff) << 32).wrapping_sub((old & 0xffff_ffff) << 32),
+        Ordering::Relaxed,
+    ) & 0xffff_ffff;
+    let prev_w1 = words[1].fetch_add(
+        ((new >> 32) << 32).wrapping_sub((old >> 32) << 32),
+        Ordering::Relaxed,
+    ) & 0xffff_ffff;
+
+    prev_w0 | (prev_w1 << 32)
+}
+fn modify_sigmask(old: Option<&mut u64>, op: Option<impl FnOnce(u64) -> u64>) -> Result<()> {
+    let _guard = tmp_disable_signals();
+    let ctl = current_sigctl();
+
+    let prev = get_allowset_raw(&ctl.word);
+
+    if let Some(old) = old {
+        *old = !prev;
+    }
+    let Some(op) = op else {
+        return Ok(());
+    };
+
+    let next = !op(!prev);
+
+    let pending = set_allowset_raw(&ctl.word, prev, next);
+
+    // POSIX requires that at least one pending unblocked signal be delivered before
+    // pthread_sigmask returns, if there is one.
+    if pending != 0 {
+        unsafe {
+            manually_enter_trampoline();
+        }
+    }
+
+    Ok(())
+}
+
+#[derive(Clone, Copy, Default)]
+pub enum SigactionKind {
+    #[default]
+    Default,
+
+    Ignore,
+    Handled {
+        handler: SignalHandler,
+    },
+}
+
+#[derive(Clone, Copy, Default)]
+pub struct Sigaction {
+    pub kind: SigactionKind,
+    pub mask: u64,
+    pub flags: SigactionFlags,
+}
+
+impl Sigaction {
+    fn ip(&self) -> usize {
+        unsafe {
+            match self.kind {
+                SigactionKind::Handled { handler } => {
+                    if self.flags.contains(SigactionFlags::SIGINFO) {
+                        handler.sigaction.map_or(0, |a| a as usize)
+                    } else {
+                        handler.handler.map_or(0, |a| a as usize)
+                    }
+                }
+                _ => 0,
+            }
+        }
+    }
+}
+
+const MASK_DONTCARE: u64 = !0;
+
+fn convert_old(action: &RawAction) -> Sigaction {
+    let old_first = action.first.load(Ordering::Relaxed);
+    let old_mask = action.user_data.load(Ordering::Relaxed);
+
+    let handler = (old_first & !(u64::from(STORED_FLAGS) << 32)) as usize;
+    let flags = SigactionFlags::from_bits_retain(((old_first >> 32) as u32) & STORED_FLAGS);
+
+    let kind = if handler == default_handler as usize {
+        SigactionKind::Default
+    } else if flags.contains(SigactionFlags::IGNORED) {
+        SigactionKind::Ignore
+    } else {
+        SigactionKind::Handled {
+            handler: unsafe { core::mem::transmute(handler as usize) },
+        }
+    };
+
+    Sigaction {
+        mask: old_mask,
+        flags,
+        kind,
+    }
+}
+
+pub fn sigaction(signal: u8, new: Option<&Sigaction>, old: Option<&mut Sigaction>) -> Result<()> {
+    if matches!(usize::from(signal), 0 | 32 | SIGKILL | SIGSTOP | 65..) {
+        return Err(Error::new(EINVAL));
+    }
+
+    let _sigguard = tmp_disable_signals();
+    let ctl = current_sigctl();
+
+    let _guard = SIGACTIONS_LOCK.lock();
+
+    let action = &PROC_CONTROL_STRUCT.actions[usize::from(signal) - 1];
+
+    if let Some(old) = old {
+        *old = convert_old(action);
+    }
+
+    let Some(new) = new else {
+        return Ok(());
+    };
+
+    let explicit_handler = new.ip();
+
+    let (mask, flags, handler) = match (usize::from(signal), new.kind) {
+        (_, SigactionKind::Ignore) | (SIGURG | SIGWINCH, SigactionKind::Default) => {
+            // TODO: POSIX specifies that pending signals shall be discarded if set to SIG_IGN by
+            // sigaction.
+            // TODO: handle tmp_disable_signals
+            (
+                MASK_DONTCARE,
+                SigactionFlags::IGNORED,
+                if matches!(new.kind, SigactionKind::Default) {
+                    default_handler as usize
+                } else {
+                    0
+                },
+            )
+        }
+        // TODO: Handle pending signals before these flags are set.
+        (SIGTSTP | SIGTTOU | SIGTTIN, SigactionKind::Default) => (
+            MASK_DONTCARE,
+            SigactionFlags::SIG_SPECIFIC,
+            default_handler as usize,
+        ),
+        (SIGCHLD, SigactionKind::Default) => {
+            let nocldstop_bit = new.flags & SigactionFlags::SIG_SPECIFIC;
+            (
+                MASK_DONTCARE,
+                SigactionFlags::IGNORED | nocldstop_bit,
+                default_handler as usize,
+            )
+        }
+
+        (_, SigactionKind::Default) => (new.mask, new.flags, default_handler as usize),
+        (_, SigactionKind::Handled { .. }) => (new.mask, new.flags, explicit_handler),
+    };
+    let new_first = (handler as u64) | (u64::from(flags.bits() & STORED_FLAGS) << 32);
+    action.first.store(new_first, Ordering::Relaxed);
+    action.user_data.store(mask, Ordering::Relaxed);
+
+    Ok(())
+}
+
+fn current_sigctl() -> &'static Sigcontrol {
+    &unsafe { Tcb::current() }.unwrap().os_specific.control
+}
+
+pub struct TmpDisableSignalsGuard {
+    _inner: (),
+}
+
+pub fn tmp_disable_signals() -> TmpDisableSignalsGuard {
+    unsafe {
+        let ctl = &current_sigctl().control_flags;
+        ctl.store(
+            ctl.load(Ordering::Relaxed) | syscall::flag::INHIBIT_DELIVERY.bits(),
+            Ordering::Release,
+        );
+        core::sync::atomic::compiler_fence(Ordering::Acquire);
+
+        // TODO: fence?
+        (*Tcb::current().unwrap().os_specific.arch.get()).disable_signals_depth += 1;
+    }
+
+    TmpDisableSignalsGuard { _inner: () }
+}
+impl Drop for TmpDisableSignalsGuard {
+    fn drop(&mut self) {
+        unsafe {
+            let depth =
+                &mut (*Tcb::current().unwrap().os_specific.arch.get()).disable_signals_depth;
+            *depth -= 1;
+
+            if *depth == 0 {
+                let ctl = &current_sigctl().control_flags;
+                ctl.store(
+                    ctl.load(Ordering::Relaxed) & !syscall::flag::INHIBIT_DELIVERY.bits(),
+                    Ordering::Release,
+                );
+                core::sync::atomic::compiler_fence(Ordering::Acquire);
+            }
+        }
+    }
+}
+
+bitflags::bitflags! {
+    // Some flags are ignored by the rt, but they match relibc's 1:1 to simplify conversion.
+    #[derive(Clone, Copy, Default)]
+    pub struct SigactionFlags: u32 {
+        const NOCLDWAIT = 2;
+        const RESTORER = 4;
+        const SIGINFO = 0x0200_0000;
+        const ONSTACK = 0x0400_0000;
+        const RESTART = 0x0800_0000;
+        const NODEFER = 0x1000_0000;
+        const RESETHAND = 0x2000_0000;
+        const SIG_SPECIFIC = 0x4000_0000;
+        const IGNORED = 0x8000_0000;
+    }
+}
+
+const STORED_FLAGS: u32 = 0xfe00_0000;
+
+fn default_handler(sig: c_int) {
+    syscall::exit(sig as usize);
+}
+
+#[derive(Clone, Copy)]
+pub union SignalHandler {
+    pub handler: Option<extern "C" fn(c_int)>,
+    pub sigaction: Option<unsafe extern "C" fn(c_int, *const (), *mut ())>,
+}
+
+static SIGACTIONS_LOCK: Mutex<()> = Mutex::new(());
+
+pub(crate) static PROC_CONTROL_STRUCT: SigProcControl = SigProcControl {
+    pending: AtomicU64::new(0),
+    actions: [const {
+        RawAction {
+            first: AtomicU64::new(0),
+            user_data: AtomicU64::new(0),
+        }
+    }; 64],
+    sender_infos: [const { AtomicU64::new(0) }; 32],
+};
+
+fn combine_allowset([lo, hi]: [u64; 2]) -> u64 {
+    (lo >> 32) | ((hi >> 32) << 32)
+}
+
+const fn sig_bit(sig: u32) -> u64 {
+    //assert_ne!(sig, 32);
+    //assert_ne!(sig, 0);
+    1 << (sig - 1)
+}
+
+pub fn setup_sighandler(tcb: &RtTcb) {
+    {
+        let _guard = SIGACTIONS_LOCK.lock();
+        for (sig_idx, action) in PROC_CONTROL_STRUCT.actions.iter().enumerate() {
+            let sig = sig_idx + 1;
+            let bits = if matches!(sig, SIGTSTP | SIGTTIN | SIGTTOU) {
+                SigactionFlags::SIG_SPECIFIC
+            } else if matches!(sig, SIGCHLD | SIGURG | SIGWINCH) {
+                SigactionFlags::IGNORED
+            } else {
+                SigactionFlags::empty()
+            };
+            action.first.store(
+                (u64::from(bits.bits()) << 32) | default_handler as u64,
+                Ordering::Relaxed,
+            );
+        }
+    }
+    let arch = unsafe { &mut *tcb.arch.get() };
+    {
+        // The asm decides whether to use the altstack, based on whether the saved stack pointer
+        // was already on that stack. Thus, setting the altstack to the entire address space, is
+        // equivalent to not using any altstack at all (the default).
+        arch.altstack_top = usize::MAX;
+        arch.altstack_bottom = 0;
+        // TODO
+        #[cfg(any(target_arch = "x86", target_arch = "aarch64", target_arch = "riscv64"))]
+        {
+            arch.pctl = core::ptr::addr_of!(PROC_CONTROL_STRUCT) as usize;
+        }
+    }
+
+    #[cfg(target_arch = "x86_64")]
+    {
+        let cpuid_eax1_ecx = unsafe { core::arch::x86_64::__cpuid(1) }.ecx;
+        CPUID_EAX1_ECX.store(cpuid_eax1_ecx, core::sync::atomic::Ordering::Relaxed);
+        SUPPORTS_AVX.store(u8::from(cpuid_eax1_ecx & 1 << 28 != 0), Ordering::Relaxed);
+    }
+
+    let data = current_setsighandler_struct();
+
+    let fd = FdGuard::new(
+        syscall::dup(**tcb.thread_fd(), b"sighandler").expect("failed to open sighandler fd"),
+    );
+    let _ = syscall::write(*fd, &data).expect("failed to write to sighandler fd");
+
+    // TODO: Inherited set of ignored signals
+    set_sigmask(Some(0), None);
+}
+pub type RtSigarea = RtTcb; // TODO
+pub fn current_setsighandler_struct() -> SetSighandlerData {
+    SetSighandlerData {
+        user_handler: sighandler_function(),
+        excp_handler: 0, // TODO
+        thread_control_addr: core::ptr::addr_of!(
+            unsafe { Tcb::current() }.unwrap().os_specific.control
+        ) as usize,
+        proc_control_addr: &PROC_CONTROL_STRUCT as *const SigProcControl as usize,
+    }
+}
+
+#[derive(Clone, Copy, Default, PartialEq)]
+pub enum Sigaltstack {
+    #[default]
+    Disabled,
+
+    Enabled {
+        onstack: bool,
+        base: *mut (),
+        size: usize,
+    },
+}
+
+pub(crate) fn get_sigaltstack(tcb: &SigArea, sp: usize) -> Sigaltstack {
+    if tcb.altstack_bottom == 0 && tcb.altstack_top == usize::MAX {
+        Sigaltstack::Disabled
+    } else {
+        Sigaltstack::Enabled {
+            base: tcb.altstack_bottom as *mut (),
+            size: tcb.altstack_top - tcb.altstack_bottom,
+            onstack: (tcb.altstack_bottom..tcb.altstack_top).contains(&sp),
+        }
+    }
+}
+
+pub unsafe fn sigaltstack(
+    new: Option<&Sigaltstack>,
+    old_out: Option<&mut Sigaltstack>,
+) -> Result<()> {
+    let _g = tmp_disable_signals();
+    let tcb = &mut *Tcb::current().unwrap().os_specific.arch.get();
+
+    let old = get_sigaltstack(tcb, crate::arch::current_sp());
+
+    if matches!(old, Sigaltstack::Enabled { onstack: true, .. }) && new != Some(&old) {
+        return Err(Error::new(EPERM));
+    }
+
+    if let Some(old_out) = old_out {
+        *old_out = old;
+    }
+    if let Some(new) = new {
+        match *new {
+            Sigaltstack::Disabled => {
+                tcb.altstack_bottom = 0;
+                tcb.altstack_top = usize::MAX;
+            }
+            Sigaltstack::Enabled { onstack: true, .. } => return Err(Error::new(EINVAL)),
+            Sigaltstack::Enabled {
+                base,
+                size,
+                onstack: false,
+            } => {
+                if size < MIN_SIGALTSTACK_SIZE {
+                    return Err(Error::new(ENOMEM));
+                }
+
+                tcb.altstack_bottom = base as usize;
+                tcb.altstack_top = base as usize + size;
+            }
+        }
+    }
+    Ok(())
+}
+
+pub const MIN_SIGALTSTACK_SIZE: usize = 2048;
+
+pub fn currently_pending_blocked() -> u64 {
+    let control = &unsafe { Tcb::current().unwrap() }.os_specific.control;
+    let w0 = control.word[0].load(Ordering::Relaxed);
+    let w1 = control.word[1].load(Ordering::Relaxed);
+    let allow = (w0 >> 32) | ((w1 >> 32) << 32);
+    let thread_pending = (w0 & 0xffff_ffff) | ((w1 >> 32) & 0xffff_ffff);
+    let proc_pending = PROC_CONTROL_STRUCT.pending.load(Ordering::Relaxed);
+
+    core::sync::atomic::fence(Ordering::Acquire); // TODO: Correct ordering?
+
+    (thread_pending | proc_pending) & !allow
+}
+pub enum Unreachable {}
+
+pub fn await_signal_async(inner_allowset: u64) -> Result<Unreachable> {
+    let _guard = tmp_disable_signals();
+    let control = &unsafe { Tcb::current().unwrap() }.os_specific.control;
+
+    let old_allowset = get_allowset_raw(&control.word);
+    set_allowset_raw(&control.word, old_allowset, inner_allowset);
+
+    let res = syscall::nanosleep(
+        &TimeSpec {
+            tv_sec: i64::MAX,
+            tv_nsec: 0,
+        },
+        &mut TimeSpec::default(),
+    );
+    set_allowset_raw(&control.word, inner_allowset, old_allowset);
+
+    if res == Err(Error::new(EINTR)) {
+        unsafe {
+            manually_enter_trampoline();
+        }
+    }
+
+    res?;
+    unreachable!()
+}
+// TODO: deadline-based API
+pub fn await_signal_sync(inner_allowset: u64, timeout: Option<&TimeSpec>) -> Result<SiginfoAbi> {
+    let _guard = tmp_disable_signals();
+    let control = &unsafe { Tcb::current().unwrap() }.os_specific.control;
+
+    let old_allowset = get_allowset_raw(&control.word);
+    let proc_pending = PROC_CONTROL_STRUCT.pending.load(Ordering::Acquire);
+    let thread_pending = set_allowset_raw(&control.word, old_allowset, inner_allowset);
+
+    // Check if there are already signals matching the requested set, before waiting.
+    if let Some(info) = try_claim_multiple(proc_pending, thread_pending, inner_allowset, control) {
+        // TODO: RAII
+        set_allowset_raw(&control.word, inner_allowset, old_allowset);
+        return Ok(info);
+    }
+
+    let res = match timeout {
+        Some(t) => syscall::nanosleep(&t, &mut TimeSpec::default()),
+        None => syscall::nanosleep(
+            &TimeSpec {
+                tv_sec: i64::MAX,
+                tv_nsec: 0,
+            },
+            &mut TimeSpec::default(),
+        ),
+    };
+
+    let thread_pending = set_allowset_raw(&control.word, inner_allowset, old_allowset);
+    let proc_pending = PROC_CONTROL_STRUCT.pending.load(Ordering::Acquire);
+
+    if let Err(error) = res
+        && error.errno != EINTR
+    {
+        return Err(error);
+    }
+
+    // Then check if there were any signals left after waiting.
+    try_claim_multiple(proc_pending, thread_pending, inner_allowset, control)
+        // Normally ETIMEDOUT but not for sigtimedwait.
+        .ok_or(Error::new(EAGAIN))
+}
+fn try_claim_multiple(
+    mut proc_pending: u64,
+    mut thread_pending: u64,
+    allowset: u64,
+    control: &Sigcontrol,
+) -> Option<SiginfoAbi> {
+    while (proc_pending | thread_pending) & allowset != 0 {
+        let sig_idx = ((proc_pending | thread_pending) & allowset).trailing_zeros();
+        if thread_pending & allowset & (1 << sig_idx) != 0
+            && let Some(res) = try_claim_single(sig_idx, Some(control))
+        {
+            return Some(res);
+        }
+        thread_pending &= !(1 << sig_idx);
+        if proc_pending & allowset & (1 << sig_idx) != 0
+            && let Some(res) = try_claim_single(sig_idx, None)
+        {
+            return Some(res);
+        }
+        proc_pending &= !(1 << sig_idx);
+    }
+    None
+}
+fn try_claim_single(sig_idx: u32, thread_control: Option<&Sigcontrol>) -> Option<SiginfoAbi> {
+    let sig_group = sig_idx / 32;
+
+    if sig_group == 1 && thread_control.is_none() {
+        // Queued (realtime) signal
+        let mut ret = MaybeUninit::<RtSigInfo>::uninit();
+        let rt_inf = unsafe {
+            syscall::syscall2(
+                syscall::SYS_SIGDEQUEUE,
+                ret.as_mut_ptr() as usize,
+                sig_idx as usize - 32,
+            )
+            .ok()?;
+            ret.assume_init()
+        };
+        Some(SiginfoAbi {
+            si_signo: sig_idx as i32 + 1,
+            si_errno: 0,
+            si_code: rt_inf.code,
+            si_pid: rt_inf.pid as i32,
+            si_uid: rt_inf.uid as i32,
+            si_status: 0,
+            si_value: rt_inf.arg,
+            si_addr: core::ptr::null_mut(),
+        })
+    } else {
+        // Idempotent (standard or thread realtime) signal
+        let info = SenderInfo::from_raw(match thread_control {
+            Some(ctl) => {
+                // Only this thread can clear pending bits, so this will always succeed.
+                let info = ctl.sender_infos[sig_idx as usize].load(Ordering::Acquire);
+                // TODO: Ordering
+                ctl.word[sig_group as usize].fetch_and(!(1 << (sig_idx % 32)), Ordering::Release);
+                info
+            }
+            None => {
+                let info =
+                    PROC_CONTROL_STRUCT.sender_infos[sig_idx as usize].load(Ordering::Acquire);
+                if PROC_CONTROL_STRUCT
+                    .pending
+                    .fetch_and(!(1 << sig_idx), Ordering::Release)
+                    & (1 << sig_idx)
+                    == 0
+                {
+                    // already claimed
+                    return None;
+                }
+                info
+            }
+        });
+        Some(SiginfoAbi {
+            si_signo: sig_idx as i32 + 1,
+            si_errno: 0,
+            si_code: 0, // TODO: SI_USER const?
+            si_pid: info.pid as i32,
+            si_uid: info.ruid as i32,
+            si_status: 0,
+            si_value: 0, // undefined
+            si_addr: core::ptr::null_mut(),
+        })
+    }
+}

redox-rt/src/sync.rs

+// TODO: Share code for simple futex-based mutex between relibc's Mutex<()> and this.
+
+use core::{
+    cell::UnsafeCell,
+    ops::{Deref, DerefMut},
+    sync::atomic::{AtomicU32, Ordering},
+};
+
+pub struct Mutex<T> {
+    pub lockword: AtomicU32,
+    pub inner: UnsafeCell<T>,
+}
+
+const UNLOCKED: u32 = 0;
+const LOCKED: u32 = 1;
+const WAITING: u32 = 2;
+
+unsafe impl<T: Send> Send for Mutex<T> {}
+unsafe impl<T: Send> Sync for Mutex<T> {}
+
+impl<T> Mutex<T> {
+    pub const fn new(t: T) -> Self {
+        Self {
+            lockword: AtomicU32::new(0),
+            inner: UnsafeCell::new(t),
+        }
+    }
+    pub fn lock(&self) -> MutexGuard<'_, T> {
+        while self
+            .lockword
+            .compare_exchange(UNLOCKED, LOCKED, Ordering::Acquire, Ordering::Relaxed)
+            .is_err()
+        {
+            core::hint::spin_loop();
+        }
+        MutexGuard { lock: self }
+    }
+}
+pub struct MutexGuard<'l, T> {
+    lock: &'l Mutex<T>,
+}
+impl<T> Deref for MutexGuard<'_, T> {
+    type Target = T;
+
+    fn deref(&self) -> &T {
+        unsafe { &*self.lock.inner.get() }
+    }
+}
+impl<T> DerefMut for MutexGuard<'_, T> {
+    fn deref_mut(&mut self) -> &mut T {
+        unsafe { &mut *self.lock.inner.get() }
+    }
+}
+impl<T> Drop for MutexGuard<'_, T> {
+    fn drop(&mut self) {
+        self.lock.lockword.store(UNLOCKED, Ordering::Release);
+    }
+}

redox-rt/src/sys.rs

+use core::{
+    ptr::addr_of,
+    sync::atomic::{AtomicU32, Ordering},
+};
+
+use syscall::{
+    error::{Error, Result, EINTR},
+    RtSigInfo, TimeSpec,
+};
+
+use crate::{arch::manually_enter_trampoline, proc::FdGuard, signal::tmp_disable_signals, Tcb};
+
+#[inline]
+fn wrapper<T>(restart: bool, mut f: impl FnMut() -> Result<T>) -> Result<T> {
+    loop {
+        let _guard = tmp_disable_signals();
+        let rt_sigarea = unsafe { &Tcb::current().unwrap().os_specific };
+        let res = f();
+
+        if let Err(err) = res
+            && err == Error::new(EINTR)
+        {
+            unsafe {
+                manually_enter_trampoline();
+            }
+            if restart && unsafe { (*rt_sigarea.arch.get()).last_sig_was_restart } {
+                continue;
+            }
+        }
+
+        return res;
+    }
+}
+
+// TODO: uninitialized memory?
+#[inline]
+pub fn posix_read(fd: usize, buf: &mut [u8]) -> Result<usize> {
+    wrapper(true, || syscall::read(fd, buf))
+}
+#[inline]
+pub fn posix_write(fd: usize, buf: &[u8]) -> Result<usize> {
+    wrapper(true, || syscall::write(fd, buf))
+}
+#[inline]
+pub fn posix_kill(pid: usize, sig: usize) -> Result<()> {
+    match wrapper(false, || syscall::kill(pid, sig)) {
+        Ok(_) | Err(Error { errno: EINTR }) => Ok(()),
+        Err(error) => Err(error),
+    }
+}
+#[inline]
+pub fn posix_sigqueue(pid: usize, sig: usize, arg: usize) -> Result<()> {
+    let siginf = RtSigInfo {
+        arg,
+        code: -1, // TODO: SI_QUEUE constant
+        uid: 0,   // TODO
+        pid: posix_getpid(),
+    };
+    match wrapper(false, || unsafe {
+        syscall::syscall3(syscall::SYS_SIGENQUEUE, pid, sig, addr_of!(siginf) as usize)
+    }) {
+        Ok(_) | Err(Error { errno: EINTR }) => Ok(()),
+        Err(error) => Err(error),
+    }
+}
+#[inline]
+pub fn posix_getpid() -> u32 {
+    // SAFETY: read-only except during program/fork child initialization
+    unsafe { crate::THIS_PID.get().read() }
+}
+#[inline]
+pub fn posix_killpg(pgrp: usize, sig: usize) -> Result<()> {
+    match wrapper(false, || syscall::kill(usize::wrapping_neg(pgrp), sig)) {
+        Ok(_) | Err(Error { errno: EINTR }) => Ok(()),
+        Err(error) => Err(error),
+    }
+}
+#[inline]
+pub unsafe fn sys_futex_wait(addr: *mut u32, val: u32, deadline: Option<&TimeSpec>) -> Result<()> {
+    wrapper(true, || {
+        syscall::syscall5(
+            syscall::SYS_FUTEX,
+            addr as usize,
+            syscall::FUTEX_WAIT,
+            val as usize,
+            deadline.map_or(0, |d| d as *const _ as usize),
+            0,
+        )
+        .map(|_| ())
+    })
+}
+#[inline]
+pub unsafe fn sys_futex_wake(addr: *mut u32, num: u32) -> Result<u32> {
+    syscall::syscall5(
+        syscall::SYS_FUTEX,
+        addr as usize,
+        syscall::FUTEX_WAKE,
+        num as usize,
+        0,
+        0,
+    )
+    .map(|awoken| awoken as u32)
+}
+pub fn sys_waitpid(pid: usize, status: &mut usize, flags: usize) -> Result<usize> {
+    wrapper(true, || {
+        syscall::waitpid(
+            pid,
+            status,
+            syscall::WaitFlags::from_bits(flags).expect("waitpid: invalid bit pattern"),
+        )
+    })
+}
+pub fn posix_kill_thread(thread_fd: usize, signal: u32) -> Result<()> {
+    let killfd = FdGuard::new(syscall::dup(thread_fd, b"signal")?);
+    match wrapper(false, || syscall::write(*killfd, &signal.to_ne_bytes())) {
+        Ok(_) | Err(Error { errno: EINTR }) => Ok(()),
+        Err(error) => Err(error),
+    }
+}
+
+static UMASK: AtomicU32 = AtomicU32::new(0o022);
+
+/// Controls the set of bits removed from the `mode` mask when new file descriptors are created.
+///
+/// Must be validated by the caller
+//
+// TODO: validate here?
+#[inline]
+pub fn swap_umask(mask: u32) -> u32 {
+    UMASK.swap(mask, Ordering::AcqRel)
+}
+
+#[inline]
+pub fn get_umask() -> u32 {
+    UMASK.load(Ordering::Acquire)
+}

redox-rt/src/thread.rs

+use core::mem::size_of;
+
+use syscall::{Result, O_CLOEXEC};
+
+use crate::{arch::*, proc::*, signal::tmp_disable_signals, RtTcb};
+
+/// Spawns a new context sharing the same address space as the current one (i.e. a new thread).
+pub unsafe fn rlct_clone_impl(stack: *mut usize) -> Result<FdGuard> {
+    let cur_thr_fd = RtTcb::current().thread_fd();
+    let new_thr_fd = FdGuard::new(syscall::open(
+        "/scheme/thisproc/new-thread/open_via_dup",
+        O_CLOEXEC,
+    )?);
+
+    copy_str(**cur_thr_fd, *new_thr_fd, "name")?;
+
+    // Inherit existing address space
+    {
+        let cur_addr_space_fd = FdGuard::new(syscall::dup(**cur_thr_fd, b"addrspace")?);
+        let new_addr_space_sel_fd = FdGuard::new(syscall::dup(*new_thr_fd, b"current-addrspace")?);
+
+        let buf = create_set_addr_space_buf(
+            *cur_addr_space_fd,
+            __relibc_internal_rlct_clone_ret as usize,
+            stack as usize,
+        );
+        let _ = syscall::write(*new_addr_space_sel_fd, &buf)?;
+    }
+
+    // Inherit reference to file table
+    {
+        let cur_filetable_fd = FdGuard::new(syscall::dup(**cur_thr_fd, b"filetable")?);
+        let new_filetable_sel_fd = FdGuard::new(syscall::dup(*new_thr_fd, b"current-filetable")?);
+
+        let _ = syscall::write(
+            *new_filetable_sel_fd,
+            &usize::to_ne_bytes(*cur_filetable_fd),
+        )?;
+    }
+
+    // Since the signal handler is not yet initialized, signals specifically targeting the thread
+    // (relibc is only required to implement thread-specific signals that already originate from
+    // the same process) will be discarded. Process-specific signals will ignore this new thread,
+    // until it has initialized its own signal handler.
+
+    // Unblock context.
+    let start_fd = FdGuard::new(syscall::dup(*new_thr_fd, b"start")?);
+    let _ = syscall::write(*start_fd, &[0])?;
+
+    Ok(new_thr_fd)
+}
+
+pub unsafe fn exit_this_thread(stack_base: *mut (), stack_size: usize) -> ! {
+    let _guard = tmp_disable_signals();
+
+    let tcb = RtTcb::current();
+    let thread_fd = tcb.thread_fd();
+
+    let _ = syscall::funmap(tcb as *const RtTcb as usize, syscall::PAGE_SIZE);
+
+    // TODO: modify interface so it writes directly to the thread fd?
+    let status_fd = syscall::dup(**thread_fd, b"status").unwrap();
+    let mut buf = [0; size_of::<usize>() * 3];
+    plain::slice_from_mut_bytes(&mut buf)
+        .unwrap()
+        .copy_from_slice(&[usize::MAX, stack_base as usize, stack_size]);
+    syscall::write(status_fd, &buf).unwrap();
+    unreachable!()
+}

redoxer.sh

+#!/usr/bin/env bash
+
+set -e
+
+if ! which redoxer
+then
+    cargo install redoxer
+fi
+
+if [ ! -d "$HOME/.redoxer/toolchain" ]
+then
+    redoxer toolchain
+fi
+
+export CARGOFLAGS=""
+export CARGO_TEST="redoxer"
+export TEST_RUNNER="redoxer exec --folder . --"
+redoxer env make "$@"

renamesyms.sh

+#!/bin/bash
+
+set -e
+
+target=$1
+deps_dir=$2
+
+if [ -z "$target" ] || [ -z "$deps_dir" ]; then
+    echo "Usage:\n\t./renamesyms.sh TARGET DEPS_DIR"
+    exit 1
+fi
+
+if [ ! -f "$target" ]; then
+    echo "Target file '$target' does not exist"
+    exit 1
+fi
+if [ ! -d "$deps_dir" ] ; then
+    echo "Deps dir '$deps_dir' does not exist or not a directory"
+    exit 1
+fi
+
+symbols_file=`mktemp`
+special_syms=(
+    __rdl_oom
+    __rg_alloc
+    __rg_alloc_zeroed
+    __rg_dealloc
+    __rg_oom
+    __rg_realloc
+    __rust_alloc
+    __rust_alloc_error_handler
+    __rust_alloc_error_handler_should_panic
+    __rust_alloc_zeroed
+    __rust_dealloc
+    __rust_no_alloc_shim_is_unstable
+    __rust_realloc
+)
+
+for dep in `find $deps_dir -type f -name "*.rlib"`; do
+    "${NM}" --format=posix -g "$dep" 2>/dev/null | sed 's/.*:.*//g' | awk '{if ($2 == "T") print $1}' | sed 's/^\(.*\)$/\1 __relibc_\1/g' >> $symbols_file
+done
+
+for special_sym in "${special_syms[@]}"; do
+    echo "$special_sym __relibc_$special_sym" >> $symbols_file
+done
+
+sorted_file=`mktemp`
+sort -u "$symbols_file" > "$sorted_file"
+rm -f "$symbols_file"
+
+"${OBJCOPY}" --redefine-syms="$sorted_file" "$target"
+
+rm -f "$sorted_file"

rust-toolchain

-nightly-2018-06-19

rust-toolchain.toml

+[toolchain]
+channel = "nightly-2025-01-12"
+components = ["rust-src"]

rustfmt.toml

-max_width = 100
-hard_tabs = false
-tab_spaces = 4
-newline_style = "Unix"
-indent_style = "Block"
-format_strings = false
+blank_lines_lower_bound = 0
+blank_lines_upper_bound = 1
+brace_style = "SameLineWhere"
+disable_all_formatting = false
+edition = "2018"
 empty_item_single_line = true
 fn_single_line = false
-where_single_line = false
-imports_indent = "Visual"
-imports_layout = "Mixed"
-reorder_extern_crates = true
-reorder_extern_crates_in_group = true
-reorder_imports = false
-reorder_imported_names = true
-spaces_within_parens_and_brackets = false
-remove_blank_lines_at_start_or_end_of_block = true
-fn_args_density = "Tall"
-brace_style = "SameLineWhere"
-trailing_comma = "Vertical"
-blank_lines_upper_bound = 1
-blank_lines_lower_bound = 0
 force_explicit_abi = true
-write_mode = "Overwrite"
-disable_all_formatting = false
-skip_children = false
+format_strings = false
+hard_tabs = false
 hide_parse_errors = false
-report_todo = "Never"
-report_fixme = "Never"
+imports_granularity = "Crate"
+imports_indent = "Block"
+imports_layout = "Mixed"
+indent_style = "Block"
+max_width = 100
+newline_style = "Unix"
+skip_children = false
+tab_spaces = 4
+trailing_comma = "Vertical"
+where_single_line = false

src/c/fcntl.c

-#include <stdarg.h>
-#include <sys/types.h>
-
-int sys_open(const char* filename, int flags, mode_t mode);
-
-int open(const char* filename, int flags, ...) {
-    mode_t mode = 0;
-    va_list ap;
-    va_start(ap, flags);
-    mode = va_arg(ap, mode_t);
-    va_end(ap);
-    return sys_open(filename, flags, mode);
-}
-
-int sys_fcntl(int fildes, int cmd, int args);
-
-int fcntl(int fildes, int cmd, ...) {
-    int args = 0;
-    va_list ap;
-    va_start(ap, cmd);
-    args = va_arg(ap, int);
-    va_end(ap);
-    return sys_fcntl(fildes, cmd, args);
-}

src/c/stdio.c

-#include <stdarg.h>
-#include <stddef.h>
-
-typedef struct FILE FILE;
-
-int vfprintf(FILE * stream, const char * fmt, va_list ap);
-
-int fprintf(FILE * stream, const char * fmt, ...) {
-    int ret;
-    va_list ap;
-    va_start(ap, fmt);
-    ret = vfprintf(stream, fmt, ap);
-    va_end(ap);
-    return ret;
-}
-
-int vprintf(const char * fmt, va_list ap);
-
-int printf(const char * fmt, ...) {
-    int ret;
-    va_list ap;
-    va_start(ap, fmt);
-    ret = vprintf(fmt, ap);
-    va_end(ap);
-    return ret;
-}
-
-int vsnprintf(char * s, size_t n, const char * fmt, va_list ap);
-
-int snprintf(char * s, size_t n, const char * fmt, ...) {
-    int ret;
-    va_list ap;
-    va_start(ap, fmt);
-    ret = vsnprintf(s, n, fmt, ap);
-    va_end(ap);
-    return ret;
-}
-
-int vsprintf(char * s, const char * fmt, va_list ap);
-
-int sprintf(char *s, const char * fmt, ...) {
-    int ret;
-    va_list ap;
-    va_start(ap, fmt);
-    ret = vsprintf(s, fmt, ap);
-    va_end(ap);
-    return ret;
-}
-
-int vfscanf(FILE * stream, const char * fmt, va_list ap);
-
-int fscanf(FILE * stream, const char * fmt, ...) {
-    int ret;
-    va_list ap;
-    va_start(ap, fmt);
-    ret = vfscanf(stream, fmt, ap);
-    va_end(ap);
-    return ret;
-}
-
-int vscanf(const char * fmt, va_list ap);
-
-int scanf(const char * fmt, ...) {
-    int ret;
-    va_list ap;
-    va_start(ap, fmt);
-    ret = vscanf(fmt, ap);
-    va_end(ap);
-    return ret;
-}
-
-int vsscanf(const char * input, const char * fmt, va_list ap);
-
-int sscanf(const char * input, const char * fmt, ...) {
-    int ret;
-    va_list ap;
-    va_start(ap, fmt);
-    ret = vsscanf(input, fmt, ap);
-    va_end(ap);
-    return ret;
-}