spawn

    fn spawn(config: SpawnConfig, comptime f: anytype, args: anytype) !Impl {
        const page_size = std.heap.pageSize();
        const Args = @TypeOf(args);
        const Instance = struct {
            fn_args: Args,
            thread: ThreadCompletion,

            fn entryFn(raw_arg: usize) callconv(.c) u8 {
                const self = @as(*@This(), @ptrFromInt(raw_arg));
                defer switch (self.thread.completion.swap(.completed, .seq_cst)) {
                    .running => {},
                    .completed => unreachable,
                    .detached => self.thread.freeAndExit(),
                };
                return callFn(f, self.fn_args);
            }
        };

        var guard_offset: usize = undefined;
        var stack_offset: usize = undefined;
        var tls_offset: usize = undefined;
        var instance_offset: usize = undefined;

        const map_bytes = blk: {
            var bytes: usize = page_size;
            guard_offset = bytes;

            bytes += @max(page_size, config.stack_size);
            bytes = std.mem.alignForward(usize, bytes, page_size);
            stack_offset = bytes;

            bytes = std.mem.alignForward(usize, bytes, linux.tls.area_desc.alignment);
            tls_offset = bytes;
            bytes += linux.tls.area_desc.size;

            bytes = std.mem.alignForward(usize, bytes, @alignOf(Instance));
            instance_offset = bytes;
            bytes += @sizeOf(Instance);

            bytes = std.mem.alignForward(usize, bytes, page_size);
            break :blk bytes;
        };

        // map all memory needed without read/write permissions
        // to avoid committing the whole region right away
        // anonymous mapping ensures file descriptor limits are not exceeded
        const mapped = posix.mmap(
            null,
            map_bytes,
            .{},
            .{ .TYPE = .PRIVATE, .ANONYMOUS = true },
            -1,
            0,
        ) catch |err| switch (err) {
            error.MemoryMappingNotSupported => unreachable,
            error.AccessDenied => unreachable,
            error.PermissionDenied => unreachable,
            error.ProcessFdQuotaExceeded => unreachable,
            error.SystemFdQuotaExceeded => unreachable,
            error.MappingAlreadyExists => unreachable,
            else => |e| return e,
        };
        assert(mapped.len >= map_bytes);
        errdefer posix.munmap(mapped);

        // Map everything but the guard page as read/write.
        const guarded: []align(std.heap.page_size_min) u8 = @alignCast(mapped[guard_offset..]);
        const protection: posix.PROT = .{ .READ = true, .WRITE = true };
        switch (posix.errno(posix.system.mprotect(guarded.ptr, guarded.len, protection))) {
            .SUCCESS => {},
            .NOMEM => return error.OutOfMemory,
            else => |err| return posix.unexpectedErrno(err),
        }

        // Prepare the TLS segment and prepare a user_desc struct when needed on x86
        var tls_ptr = linux.tls.prepareArea(mapped[tls_offset..][0..linux.tls.area_desc.size]);
        var user_desc: if (target.cpu.arch == .x86) linux.user_desc else void = undefined;
        if (target.cpu.arch == .x86) {
            defer tls_ptr = @intFromPtr(&user_desc);
            user_desc = .{
                .entry_number = linux.tls.area_desc.gdt_entry_number,
                .base_addr = tls_ptr,
                .limit = 0xfffff,
                .flags = .{
                    .seg_32bit = 1,
                    .contents = 0, // Data
                    .read_exec_only = 0,
                    .limit_in_pages = 1,
                    .seg_not_present = 0,
                    .useable = 1,
                },
            };
        }

        const instance: *Instance = @ptrCast(@alignCast(&mapped[instance_offset]));
        instance.* = .{
            .fn_args = args,
            .thread = .{ .mapped = mapped },
        };

        const flags: u32 = linux.CLONE.THREAD | linux.CLONE.DETACHED |
            linux.CLONE.VM | linux.CLONE.FS | linux.CLONE.FILES |
            linux.CLONE.PARENT_SETTID | linux.CLONE.CHILD_CLEARTID |
            linux.CLONE.SIGHAND | linux.CLONE.SYSVSEM | linux.CLONE.SETTLS;

        switch (linux.errno(linux.clone(
            Instance.entryFn,
            @intFromPtr(&mapped[stack_offset]),
            flags,
            @intFromPtr(instance),
            &instance.thread.parent_tid,
            tls_ptr,
            &instance.thread.child_tid.raw,
        ))) {
            .SUCCESS => return Impl{ .thread = &instance.thread },
            .AGAIN => return error.ThreadQuotaExceeded,
            .INVAL => unreachable,
            .NOMEM => return error.SystemResources,
            .NOSPC => unreachable,
            .PERM => unreachable,
            .USERS => unreachable,
            else => |err| return posix.unexpectedErrno(err),
        }
    }