Linux passes an initial stack layout to applications, that contains argc, argv, envp, and the auxiliary vector
right above the stack pointer. The libc of a Linux program parses this structure in its _start-symbol ("crt0") and
passes the right pointers as arguments to main afterwards. This crate helps to construct and parse this data structure
in no_std environments and for different address spaces.
Keywords: crt0, stack layout, AT values, AT pairs, auxvec, auxiliary vector
This crate has been tested successfully by myself in a custom runtime system for a Microkernel, that is able to load and start unmodified Linux binaries. The Linux binary (the libc) could find all arguments, environment variables, and the data from the auxiliary vector and print it to stdout.
How does this differ from https://crates.io/crates/crt0stack and https://crates.io/crates/auxv?
This crate supports no_std-contexts plus allows construction the data structure for a different address
space, i.e. the address space of a user application.
When I started creating this crate, I only knew about the latter. It doesn't support no_std. Because
the first one supports no_std but not different address spaces, I still had to create this one.
The typical use case for me is to create the data structure for a different address space, like Linux does.
Last but not least, my crate supports more/all of Linux's AT variables.
✅ build data structure for current address space
✅ build data structure for different address space
✅ parse data structure for current address space + output referenced data/pointers
✅ parse data structure for different address space + prevent memory error / no dereferencing of pointers
The auxiliary vector contains pairs of type (usize, usize). Hence, each entry takes 8 bytes on 32-bit systems
and 16 byte on 64-bit systems. Currently, this crate produces the auxiliary vector for the architecture it is
compiled with. If necessary, create an issue or a PR and this will be a runtime setting. I never tested it
on a 32-bit system, but I am confident it will work.
Right now, this crate can only build and serialize the whole initial stack layout but not the auxiliary vector standalone.
There are multiple code examples in the repository!
cargo run --example linux_parse_print_layout shows you a
real world example.
use linux_libc_auxv::{AuxVar, InitialLinuxLibcStackLayout, InitialLinuxLibcStackLayoutBuilder};
/// Minimal example that builds the initial linux libc stack layout and parses it again.
fn main() {
let builder = InitialLinuxLibcStackLayoutBuilder::new()
// can contain terminating zero; not mandatory in the builder
.add_arg_v("./first_arg\0")
.add_arg_v("./second_arg")
.add_env_v("FOO=BAR\0")
.add_env_v("PATH=/bin")
.add_aux_v(AuxVar::Clktck(100))
.add_aux_v(AuxVar::Random([
1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16,
]))
.add_aux_v(AuxVar::ExecFn("/usr/bin/foo"))
.add_aux_v(AuxVar::Platform("x86_64"));
// memory where we serialize the data structure into
let mut buf = vec![0; builder.total_size()];
// assume user stack is at 0x7fff0000
let user_base_addr = 0x7fff0000;
unsafe {
builder.serialize_into_buf(buf.as_mut_slice(), user_base_addr);
}
// So far, this is memory safe, as long as the slice is valid memory. No pointers are
// dereferenced yet.
let parsed = InitialLinuxLibcStackLayout::from(buf.as_slice());
println!("There are {} arguments.", parsed.argc());
println!(
"There are {} environment variables.",
parsed.envv_ptr_iter().count()
);
println!(
"There are {} auxiliary vector entries/AT variables.",
parsed.aux_serialized_iter().count()
);
println!(" argv");
// ptr iter is safe for other address spaces; the other only because here user_addr == write_addr
for (i, arg) in parsed.argv_ptr_iter().enumerate() {
println!(" [{}] @ {:?}", i, arg);
}
println!(" envp");
// ptr iter is safe for other address spaces; the other only because here user_addr == write_addr
for (i, env) in parsed.envv_ptr_iter().enumerate() {
println!(" [{}] @ {:?}", i, env);
}
println!(" aux");
// ptr iter is safe for other address spaces; the other only because here user_addr == write_addr
for aux in parsed.aux_serialized_iter() {
if aux.key().value_in_data_area() {
println!(" {:?} => @ {:?}", aux.key(), aux.val() as *const u8);
} else {
println!(" {:?} => {:?}", aux.key(), aux.val() as *const u8);
}
}
}There are 2 arguments.
There are 2 environment variables.
There are 5 auxiliary vector entries/AT variables.
argv
[0] @ 0x7fff00b0
[1] @ 0x7fff00bc
envp
[0] @ 0x7fff00c9
[1] @ 0x7fff00d1
aux
Platform => @ 0x7fff0090
Clktck => 0x64
Random => @ 0x7fff0097
ExecFn => @ 0x7fff00db
Null => 0x0
The whole data structure is called InitialLinuxLibcStackLayout by me. There is no official name. It contains
the arguments (argc and argv), the environment variables (envp or envv), and the auxiliary vector
(AT-variables, auxv, aux-pairs, aux entries).
The argv-array will reference data in the argv data area, the envv-array will reference data in the
envv data area, and some of the auxv-values might reference data in the auxv data area.
Sometimes (in some articles), the auxiliary vector even describes the whole data structure.
null [HIGH ADDRESS]
filename (c string)
<env data area>
<args data area>
// round up to 16 byte
<aux vec data area>
// round up to 16 byte alignment
AT_VAR_3 = <points to aux vec data area>
AT_VAR_2 = integer
AT_VAR_1 = integer
// round up to 16 byte alignment
envv[2] = null
envv[1] = <points to env data area>
envv[0] = <points to env data area>
argv[2] = null
argv[1] = <points to args data area>
argv[0] = <points to args data area>
argc = integer <libc entry stack top> [LOW ADDRESS]
1.56.1 stable / Rust edition 2021
- https://lwn.net/Articles/631631/ (good overview with ASCII graphics)
- https://lwn.net/Articles/519085/
- https://elixir.bootlin.com/linux/v5.15.5/source/fs/binfmt_elf.c#L257 (code in Linux that constructs
auxv) - https://man7.org/linux/man-pages/man3/getauxval.3.html
- https://refspecs.linuxfoundation.org/ELF/zSeries/lzsabi0_zSeries/x895.html