Writing Helios drivers in the Mercury driver environment April 8, 2023 on Drew DeVault's blog

Helios is a microkernel written in the Hare programming language and is part of the larger Ares operating system. You can watch my FOSDEM 2023 talk introducing Helios on PeerTube.

Let’s take a look at the new Mercury driver development environment for Helios.

As you may remember from my FOSDEM talk, the Ares operating system is built out of several layers which provide progressively higher-level environments for an operating system. At the bottom is the Helios microkernel, and today we’re going to talk about the second layer: the Mercury environment, which is used for writing and running device drivers in userspace. Let’s take a look at a serial driver written against Mercury and introduce some of the primitives used by driver authors in the Mercury environment.

Drivers for Mercury are written as normal ELF executables with an extra section called .manifest, which includes a file similar to the following (the provided example is for the serial driver we’ll be examining today):

[driver]
name=pcserial
desc=Serial driver for x86_64 PCs

[capabilities]
0:ioport = min=3F8, max=400
1:ioport = min=2E8, max=2F0
2:note = 
3:irq = irq=3, note=2
4:irq = irq=4, note=2
_:cspace = self
_:vspace = self
_:memory = pages=32

[services]
devregistry=

Helios uses a capability-based design, in which access to system resources (such as I/O ports, IRQs, or memory) is governed by capability objects. Each process has a capability space, which is a table of capabilities assigned to that process, and when performing operations (such as writing to an I/O port) the user provides the index of the desired capability in a register when invoking the appropriate syscall.

The manifest first specifies a list of capabilities required to operate the serial port. It requests, assigned static capability addresses, capabilities for the required I/O ports and IRQs, as well as a notification object which the IRQs will be delivered to. Some capability types, such as I/O ports, have configuration parameters, in this case the minimum and maximum port numbers which are relevant. The IRQ capabilities require a reference to a notification as well.

Limiting access to these capabilities provides very strong isolation between device drivers. On a monolithic kernel like Linux, a bug in the serial driver could compromise the entire system, but a vulnerability in our driver could, at worst, write garbage to your serial port. This model also provides better security than something like OpenBSD’s pledge by declaratively specifying what we need and nothing else.

Following the statically allocated capabilities, we request our own capability space and virtual address space, the former so we can copy and destroy our capabilities, and the latter so that we can map shared memory to perform reads and writes for clients. We also request 32 pages of memory, which we use to allocate page tables to perform those mappings; this will be changed later. These capabilities do not require any specific address for the driver to work, so we use “_” to indicate that any slot will suit our needs.

Mercury uses some vendor extensions over the System-V ABI to communicate information about these capabilities to the runtime. Notes about each of the _’d capabilities are provided by the auxiliary vector, and picked up by the Mercury runtime – for instance, the presence of a memory capability is detected on startup and is used to set up the allocator; the presence of a vspace capability is automatically wired up to the mmap implementation.

Each of these capabilities is implemented by the kernel, but additional services are available in userspace via endpoint capabilities. Each of these endpoints implements a particular API, as defined by a protocol definition file. This driver requires access to the device registry, so that it can create devices for its serial ports and expose them to clients.

These protocol definitions are written in a domain-specific language and parsed by ipcgen to generate client and server implementations of each. Here’s a simple protocol to start us off:

namespace io;

# The location with respect to which a seek operation is performed.
enum whence {
	# From the start of the file
	SET,
	# From the current offset
	CUR,
	# From the end of the file
	END,
};

# An object with file-like semantics.
interface file {
	# Reads up to amt bytes of data from a file.
	call read{pages: page...}(buf: uintptr, amt: size) size;

	# Writes up to amt bytes of data to a file.
	call write{pages: page...}(buf: uintptr, amt: size) size;

	# Seeks a file to a given offset, returning the new offset.
	call seek(offs: i64, w: whence) size;
};

Each interface includes a list of methods, each of which can take a number of capabilities and parameters, and return a value. The “read” call here, when implemented by a file-like object, accepts a list of memory pages to perform the read or write with (shared memory), as well as a pointer to the buffer address and size. Error handling is still a to-do.

ipcgen consumes these files and writes client or server code as appropriate. These are generated as part of the Mercury build process and end up in *_gen.ha files. The generated client code is filed away into the relevant modules (this protocol ends up at io/file_gen.ha), alongside various hand-written files which provide additional functionality and often wrap the IPC calls in a higher-level interface. The server implementations end up in the “serv” module, e.g. serv/io/file_gen.ha.

Let’s look at some of the generated client code for io::file objects:

// This file was generated by ipcgen; do not modify by hand
use helios;
use rt;

// ID for the file IPC interface.
export def FILE_ID: u32 = 0x9A533BB3;

// Labels for operations against file objects.
export type file_label = enum u64 {
	READ = FILE_ID << 16u64 | 1,
	WRITE = FILE_ID << 16u64 | 2,
	SEEK = FILE_ID << 16u64 | 3,
};

export fn file_read(
	ep: helios::cap,
	pages: []helios::cap,
	buf: uintptr,
	amt: size,
) size = {
	// ...
};

Each interface has a unique ID (generated from the FNV-1a hash of its fully qualified name), which is bitwise-OR’d with a list of operations to form call labels. The interface ID is used elsewhere; we’ll refer to it again later. Then each method generates an implementation which arranges the IPC details as necessary and invokes the “call” syscall against the endpoint capability.

The generated server code is a bit more involved. Some of the details are similar – FILE_ID is generated again, for instance – but there are some additional details as well. First is the generation of a vtable defining the functions implementing each operation:

// Implementation of a [[file]] object.
export type file_iface = struct {
	read: *fn_file_read,
	write: *fn_file_write,
	seek: *fn_file_seek,
};

We also define a file object which is subtyped by the implementation to store implementation details, and which provides to the generated code the required bits of state.

// Instance of an file object. Users may subtype this object to add
// instance-specific state.
export type file = struct {
	_iface: *file_iface,
	_endpoint: helios::cap,
};

Here’s an example of a subtype of file used by the initramfs to store additional state:

// An open file in the bootstrap filesystem
type bfs_file = struct {
	serv::io::file,
	fs: *bfs,
	ent: tar::entry,
	cur: io::off,
	padding: size,
};

The embedded serv::io::file structure here is populated with an implementation of file_iface, here simplified for illustrative purposes:

const bfs_file_impl = serv_io::file_iface {
	read = &bfs_file_read,
	write = &bfs_file_write,
	seek = &bfs_file_seek,
};

fn bfs_file_read(
	obj: *serv_io::file,
	pages: []helios::cap,
	buf: uintptr,
	amt: size,
) size = {
	let file = obj: *bfs_file;
	const fs = file.fs;
	const offs = (buf & rt::PAGEMASK): size;
	defer helios::destroy(pages...)!;

	assert(offs + amt <= len(pages) * rt::PAGESIZE);
	const buf = helios::map(rt::vspace, 0, map_flags::W, pages...)!: *[*]u8;

	let buf = buf[offs..offs+amt];
	// Not shown: reading the file data into this buffer
};

The implementation can prepare a file object and call dispatch on it to process client requests: this function blocks until a request arrives, decodes it, and invokes the appropriate function. Often this is incorporated into an event loop with poll to service many objects at once.

// Prepare a file object
const ep = helios::newendpoint()!;
append(fs.files, bfs_file {
	_iface = &bfs_file_impl,
	_endpoint = ep,
	fs = fs,
	ent = ent,
	cur = io::tell(fs.buf)!,
	padding = fs.rd.padding,
});

// ...

// Process requests associated with this file
serv::io::file_dispatch(file);

Okay, enough background: back to the serial driver. It needs to implement the following protocol:

namespace dev;
use io;

# TODO: Add busy error and narrow semantics

# Note: TWO is interpreted as 1.5 for some char lengths (5)
enum stop_bits {
	ONE,
	TWO,
};

enum parity {
	NONE,
	ODD,
	EVEN,
	MARK,
	SPACE,
};

# A serial device, which implements the file interface for reading from and
# writing to a serial port. Typical implementations may only support one read
# in-flight at a time, returning errors::busy otherwise.
interface serial :: io::file {
	# Returns the baud rate in Hz.
	call get_baud() uint;

	# Returns the configured number of bits per character.
	call get_charlen() uint;

	# Returns the configured number of stop bits.
	call get_stopbits() stop_bits;

	# Returns the configured parity setting.
	call get_parity() parity;

	# Sets the baud rate in Hz.
	call set_baud(hz: uint) void;

	# Sets the number of bits per character. Must be 5, 6, 7, or 8.
	call set_charlen(bits: uint) void;

	# Configures the number of stop bits to use.
	call set_stopbits(bits: stop_bits) void;

	# Configures the desired parity.
	call set_parity(parity: parity) void;
};

This protocol inherits the io::file interface, so the serial port is usable like any other file for reads and writes. It additionally defines serial-specific methods, such as configuring the baud rate or parity. The generated interface we’ll have to implement looks something like this, embedding the io::file_iface struct:

export type serial_iface = struct {
	io::file_iface,
	get_baud: *fn_serial_get_baud,
	get_charlen: *fn_serial_get_charlen,
	get_stopbits: *fn_serial_get_stopbits,
	get_parity: *fn_serial_get_parity,
	set_baud: *fn_serial_set_baud,
	set_charlen: *fn_serial_set_charlen,
	set_stopbits: *fn_serial_set_stopbits,
	set_parity: *fn_serial_set_parity,
}

Time to dive into the implementation. Recall the driver manifest, which provides the serial driver with a suitable environment:

[driver]
name=pcserial
desc=Serial driver for x86_64 PCs

[capabilities]
0:ioport = min=3F8, max=400
1:ioport = min=2E8, max=2F0
2:note = 
3:irq = irq=3, note=2
4:irq = irq=4, note=2
_:cspace = self
_:vspace = self
_:memory = pages=32

[services]
devregistry=

I/O ports for reading and writing to the serial devices, IRQs for receiving serial-related interrupts, a device registry to add our serial devices to the system, and a few extra things for implementation needs. Some of these are statically allocated, some of them are provided via the auxiliary vector. Our serial driver opens by defining constants for the statically allocated capabilities:

def IOPORT_A: helios::cap = 0;
def IOPORT_B: helios::cap = 1;
def IRQ: helios::cap = 2;
def IRQ3: helios::cap = 3;
def IRQ4: helios::cap = 4;

The first thing we do on startup is create a serial device.

export fn main() void = {
	let serial0: helios::cap = 0;
	const registry = helios::service(sys::DEVREGISTRY_ID);
	sys::devregistry_new(registry, dev::SERIAL_ID, &serial0);
	helios::destroy(registry)!;
	// ...

The device registry is provided via the aux vector, and we can use helios::service to look it up by its interface ID. Then we use the devregistry::new operation to create a serial device:

# Device driver registry.
interface devregistry {
	# Creates a new device implementing the given interface ID using the
	# provided endpoint capability and returns its assigned serial number.
	call new{; out}(iface: u64) uint;
};

After this we can destroy the registry – we won’t need it again and it’s best to get rid of it so that we can work with the minimum possible privileges at runtime. After this we initialize the serial port, acknowledge any interrupts that might have been pending before we got started, an enter the main loop.

com_init(&ports[0], serial0);

helios::irq_ack(IRQ3)!;
helios::irq_ack(IRQ4)!;

let poll: [_]pollcap = [
	pollcap { cap = IRQ, events = pollflags::RECV, ... },
	pollcap { cap = serial0, events = pollflags::RECV, ... },
];
for (true) {
	helios::poll(poll)!;
	if (poll[0].revents & pollflags::RECV != 0) {
		dispatch_irq();
	};
	if (poll[1].revents & pollflags::RECV != 0) {
		dispatch_serial(&ports[0]);
	};
};

The dispatch_serial function is of interest, as this provides the implementation of the serial object we just created with the device registry.

type comport = struct {
	dev::serial,
	port: u16,
	rbuf: [4096]u8,
	wbuf: [4096]u8,
	rpending: []u8,
	wpending: []u8,
};

fn dispatch_serial(dev: *comport) void = {
	dev::serial_dispatch(dev);
};

const serial_impl = dev::serial_iface {
	read = &serial_read,
	write = &serial_write,
	seek = &serial_seek,
        get_baud = &serial_get_baud,
        get_charlen = &serial_get_charlen,
        get_stopbits = &serial_get_stopbits,
        get_parity = &serial_get_parity,
        set_baud = &serial_set_baud,
        set_charlen = &serial_set_charlen,
        set_stopbits = &serial_set_stopbits,
        set_parity = &serial_set_parity,
};

fn serial_read(
	obj: *io::file,
	pages: []helios::cap,
	buf: uintptr,
	amt: size,
) size = {
	const port = obj: *comport;
	const offs = (buf & rt::PAGEMASK): size;
	const buf = helios::map(rt::vspace, 0, map_flags::W, pages...)!: *[*]u8;
	const buf = buf[offs..offs+amt];

	if (len(port.rpending) != 0) {
		defer helios::destroy(pages...)!;
		return rconsume(port, buf);
	};

	pages_static[..len(pages)] = pages[..];
	pending_read = read {
		reply = helios::store_reply(helios::CADDR_UNDEF)!,
		pages = pages_static[..len(pages)],
		buf = buf,
	};
	return 0;
};

// (other functions omitted)

We’ll skip much of the implementation details for this specific driver, but I’ll show you how read works at least. It’s relatively straightforward: first we mmap the buffer provided by the caller. If there’s already readable data pending from the serial port (stored in that rpending slice in the comport struct, which is a slice of the statically-allocated rbuf field), we copy it into the buffer and return the number of bytes we had ready. Otherwise, we stash details about the caller, storing the special reply capability in our cspace (this is one of the reasons we need cspace = self in our manifest) so we can reply to this call once data is available. Then we return to the main loop.

The main loop also wakes up on an interrupt, and we have an interrupt unmasked on the serial device to wake us whenever there’s data ready to be read. Eventually this gets us here, which finishes the call we saved earlier:

// Reads data from the serial port's RX FIFO.
fn com_read(com: *comport) size = {
	let n: size = 0;
	for (comin(com.port, LSR) & RBF == RBF; n += 1) {
		const ch = comin(com.port, RBR);
		if (len(com.rpending) < len(com.rbuf)) {
			// If the buffer is full we just drop chars
			static append(com.rpending, ch);
		};
	};

	if (pending_read.reply != 0) {
		const n = rconsume(com, pending_read.buf);
		helios::send(pending_read.reply, 0, n)!;
		pending_read.reply = 0;
		helios::destroy(pending_read.pages...)!;
	};

	return n;
};

I hope that gives you a general idea of how drivers work in this environment! I encourage you to read the full implementation if you’re curious to know more about the serial driver in particular – it’s just 370 lines of code.

The last thing I want to show you is how the driver gets executed in the first place. When Helios boots up, it starts /sbin/sysinit, which is provided by Mercury and offers various low-level userspace runtime services, such as the device registry and bootstrap filesystem we saw earlier. After setting up its services, sysinit executes /sbin/usrinit, which is provided by the next layer up (Gaia, eventually) and sets up the rest of the system according to user policy, mounting filesystems and starting up drivers and such. At the moment, usrinit is fairly simple, and just runs a little demo. Here it is in full:

use dev;
use fs;
use helios;
use io;
use log;
use rt;
use sys;

export fn main() void = {
	const fs = helios::service(fs::FS_ID);
	const procmgr = helios::service(sys::PROCMGR_ID);
	const devmgr = helios::service(sys::DEVMGR_ID);
	const devload = helios::service(sys::DEVLOADER_ID);

	log::printfln("[usrinit] Running /sbin/drv/serial");
	let proc: helios::cap = 0;
	const image = fs::open(fs, "/sbin/drv/serial")!;
	sys::procmgr_new(procmgr, &proc);
	sys::devloader_load(devload, proc, image);
	sys::process_start(proc);

	let serial: helios::cap = 0;
	log::printfln("[usrinit] open device serial0");
	sys::devmgr_open(devmgr, dev::SERIAL_ID, 0, &serial);

	let buf: [rt::PAGESIZE]u8 = [0...];
	for (true) {
		const n = match (io::read(serial, buf)!) {
		case let n: size =>
			yield n;
		case io::EOF =>
			break;
		};

		// CR => LF
		for (let i = 0z; i < n; i += 1) {
			if (buf[i] == '\r') {
				buf[i] = '\n';
			};
		};

		// echo
		io::write(serial, buf[..n])!;
	};
};

Each of the services shown at the start are automatically provided in usrinit’s aux vector by sysinit, and includes all of the services required to bootstrap the system. This includes a filesystem (the initramfs), a process manager (to start up new processes), the device manager, and the driver loader service.

usrinit starts by opening up /sbin/drv/serial (the serial driver, of course) from the provided initramfs using fs::open, which is a convenience wrapper around the filesystem protocol. Then we create a new process with the process manager, which by default has an empty address space – we could load a normal process into it with sys::process_load, but we want to load a driver, so we use the devloader interface instead. Then we start the process and boom: the serial driver is online.

The serial driver registers itself with the device registry, which means that we can use the device manager to open the 0th device which implements the serial interface. Since this is compatible with the io::file interface, it can simply be used normally with io::read and io::write to utilize the serial port. The main loop simply echos data read from the serial port back out. Simple!


That’s a quick introduction to the driver environment provided by Mercury. I intend to write a few more drivers soon myself – PC keyboard, framebuffer, etc – and set up a simple shell. We have seen a few sample drivers written pre-Mercury which would be nice to bring into this environment, such as virtio networking and block devices. It will be nice to see them re-introduced in an environment where they can provide useful services to the rest of userspace.

If you’re interested in learning more about Helios or Mercury, consult ares-os.org for documentation – though beware of the many stub pages. If you have any questions or want to get involved in writing some drivers yourself, jump into our IRC channel: #helios on Libera Chat.

Articles from blogs I read Generated by openring

Status update, April 2024

Hi! The X.Org Foundation results are in, and I’m now officially part of the Board of Directors. I hope I can be of use to the community on more organizational issues! Speaking of which, I’ve spent quite a bit of time dealing with Code of Conduct matters latel…

via emersion April 16, 2024

M2dir: treating mails as files without going crazy

Sometime recently in the past I complained about Maildir. You can go read the post, but the executive summary is that I think Maildir uses an actively user-hostile directory structure and extremely convoluted filenames that do not convey any meaning at all. …

via blogfehler! April 15, 2024

Go Developer Survey 2024 H1 Results

What we learned from our 2024 H1 developer survey

via The Go Blog April 9, 2024