Files
danos/test/system/services/protocol-conformance-test/protocol-conformance-test.zig
T
Daniel Samson 2719b93530 library: the last three protocols speak the envelope
These were the awkward ones. Each began with an operation packed into a
single byte — two of them with a version wedged in beside it — so there was
no wrapping them: the layouts had to be rebuilt. The device manager's own
enumerate and subscribe become the reserved verbs that mean the same thing
everywhere, its replies lose three status structs the envelope already
carries, and a device id becomes the packet's target. Power drops the
version it repeated on every request, because describe is the handshake,
and stops claiming a 64-byte ceiling it never needed for calls. USB moves a
control transfer's data to the packet tail in both directions, which makes
the status length the transferred length and retires a field that had been
saying the same thing twice.

The danger in this one was not the protocols but their readers. Init
recognised a power button by two bytes at the head of a message, the ACPI
service dispatched on the first byte, the xHCI driver read its operation
with a raw integer load, and the HID drivers reinterpreted a report
wholesale — none of which would have failed to compile once the layouts
moved. They would simply have stopped: no shutdown on the power button, no
reports from the keyboard. Every one of them now reads through the
generated types, and the shutdown gate that answers only a subscriber is
the same code it was.

Two sizes were decided by measuring rather than assuming. The child-added
message is both a request and the event broadcast to subscribers, and
alignment rounds it to 48 bytes, which puts its packet exactly on the
64-byte push floor — a test pins that, because a field added carelessly
would now overflow it. The interrupt report gives up eight bytes of inline
room to make space for the header; the two drivers that produce reports
send eight and four.

Suite 110/110.
2026-08-01 07:20:37 +01:00

349 lines
17 KiB
Zig

//! protocol-conformance-test — P4a's evidence that `envelope.Define` gives every
//! provider the reserved verbs, uniformly and without the provider writing a line
//! for them (docs/security-track-plan.md P4a;
//! docs/os-development/protocol-namespace.md). One binary, one role, driven by
//! the `protocol-conformance` kernel case:
//!
//! - `protocol-conformance-test run` — for each contract it can reach:
//! 1. `describe` — the reserved verb 0 — is answered, and the answer names
//! *that* protocol: the name it was opened under, the version its module
//! declares, and the number of verbs its module declares. No provider in
//! the system implements `describe`; the generated dispatch answers it out
//! of the specification, which is exactly the claim being checked;
//! 2. a verb number no protocol in the system defines answers `-ENOSYS`, and
//! carries no capability. That is the other half of the same generated
//! dispatch: a provider does not have to reject strangers, it gets the
//! rejection for free and every provider gives the same one;
//! 3. `describe` again, after the refusal — a refused verb is an *answer*,
//! not a wedged service, so the channel is still good afterwards.
//!
//! **The set it checks is read, never hardcoded.** The fixture asks `/protocol`
//! for its own listing (`readdir`, which the namespace publishes on purpose) and
//! walks what it finds, so the case cannot drift from what this boot actually
//! bound. What it opens is bounded by P3: the manifest names this binary against
//! exactly the two contracts its scenario boots, and an ungranted name is absent
//! for it like any other client's.
//!
//! **What it covers, and what it cannot — the honest list.** The scenario boots
//! the registry, the input service, and the compositor, so `input` and `display`
//! are checked end to end over real IPC. The other six contracts in the table
//! are not asked here, and the reason is the provider, not the protocol:
//!
//! - `vfs` — the FAT server, which needs a mounted volume behind the whole USB
//! storage chain (the `fat-mount` scenario);
//! - `block` — the usb-storage driver, which the device manager spawns after
//! enumerating an xHCI bus (the `usb-storage` scenario);
//! - `scanout` — the virtio-gpu driver, which needs an emulated virtio-gpu the
//! default harness does not attach (the `virtio-gpu` scenario);
//! - `device-manager`, `power` and `usb-transfer` — the three P4b rebased. All
//! three come with the device manager: it *is* the first, it spawns the
//! discovery service that binds the second, and the xHCI driver it spawns
//! binds the third. So booting a provider for any one of them means booting
//! the whole driver tree here.
//!
//! That is the reason this scenario stays at two providers rather than five or
//! eight. It is not only the cost of booting half the system to send two
//! packets: this fixture takes **one snapshot** of `/protocol` and checks what
//! is in it, so a scenario whose bound set depends on how far a driver tree got
//! by that instant would make the case's own summary a boot race. What proves
//! the six instead is the scenarios that already drive them end to end —
//! `fat-mount`, `usb-storage`, `virtio-gpu`, and for the P4b three the
//! `device-list`, `driver-restart`, `pci-scan`, `usb-*`, `power-button` and
//! `orderly-shutdown` cases, every one of which is a live conversation over
//! these wires.
//!
//! All six sit in the table below regardless, so a scenario that binds one gets
//! it conformance-checked without this file being edited — and every run prints,
//! by name, the ones it found no provider for.
//!
//! The registry itself — PID 1 serving `/protocol` — is the one vfs backend
//! deliberately NOT dispatched through the generated table (it reads a
//! stranger's packet by hand, `system/services/init/init.zig`), so `describe` is
//! not asked of it and nothing here claims it.
//!
//! Prints `protocol-conformance: ok` on success, or a `protocol-conformance:
//! FAIL` line naming the step. Spawned bare (the initial-ramdisk sweep starts
//! every bundled binary), it exits silently so it cannot derange other tests.
const std = @import("std");
const channel = @import("channel");
const envelope = @import("envelope");
const file_system = @import("file-system");
const ipc = @import("ipc");
const logging = @import("logging");
const process = @import("process");
const time = @import("time");
const block_protocol = @import("block-protocol");
const device_manager_protocol = @import("device-manager-protocol");
const display_protocol = @import("display-protocol");
const input_protocol = @import("input-protocol");
const power_protocol = @import("power-protocol");
const scanout_protocol = @import("scanout-protocol");
const usb_transfer_protocol = @import("usb-transfer-protocol");
const vfs_protocol = @import("vfs-protocol");
// --- what conformance means, per contract -----------------------------------
/// One contract this fixture knows how to check, and what the answer must say.
/// Every field is read off the protocol module itself, so the expectation is the
/// contract's own definition rather than a number copied beside it — a version
/// bump or a new verb updates this table by recompiling.
const Contract = struct {
name: []const u8,
version: u32,
/// How many verbs the module declares — `describe` reports it, so it is
/// checked. The reserved verbs are not counted: they are the envelope's.
operations: u32,
/// Whether this scenario boots a provider for it. A required contract that
/// is missing, unreachable or non-conforming fails the case; the rest are
/// checked when some other scenario happens to bind them.
required: bool,
};
fn contractOf(comptime Protocol: type, required: bool) Contract {
return .{
.name = Protocol.protocol_name,
.version = Protocol.version,
.operations = @typeInfo(Protocol.Operation).@"enum".fields.len,
.required = required,
};
}
/// The protocols built on `envelope.Define`. A name listed by `/protocol` that
/// is absent from here is reported and left alone rather than probed: a
/// hand-numbered provider would read operation 0 as one of its own verbs, so
/// asking it for `describe` would *do* something. Only `ps2-bus` is still in
/// that state today.
const contracts = [_]Contract{
contractOf(input_protocol.Protocol, true), // the input fan-out service
contractOf(display_protocol.Protocol, true), // the compositor
contractOf(vfs_protocol.Protocol, false), // the FAT server — needs a volume
contractOf(block_protocol.Protocol, false), // usb-storage — needs the xHCI chain
contractOf(scanout_protocol.Protocol, false), // virtio-gpu — needs the device
// The three P4b rebased. Each needs the device manager (and, for the last
// two, what the device manager starts), which is more than this scenario
// boots — see the header.
contractOf(device_manager_protocol.Protocol, false),
contractOf(power_protocol.Protocol, false), // the discovery service
contractOf(usb_transfer_protocol.Protocol, false), // the xHCI bus driver
};
/// A verb number no protocol in the system defines, and none plausibly will: far
/// above the reserved range, so it is unambiguously a protocol verb, and far
/// above any protocol's verb count, so the generated dispatch has nothing to
/// match it against. The answer must be `-ENOSYS` at every provider.
const stranger_operation: u32 = envelope.first_protocol_operation + 4096;
fn fail(step: []const u8) noreturn {
_ = logging.write("protocol-conformance: FAIL ");
_ = logging.write(step);
_ = logging.write("\n");
process.exit(1);
}
fn report(comptime format: []const u8, arguments: anytype) void {
var line: [192]u8 = undefined;
_ = logging.write(std.fmt.bufPrint(&line, format, arguments) catch return);
}
// --- reading the namespace ---------------------------------------------------
/// The cadence every client in the tree spends finding a service.
const resolve_attempts: u32 = 200;
const resolve_retry_ms: u64 = 20;
/// The registry's endpoint, obtained the way every process obtains it: resolve
/// `/protocol`. The handle is the kernel's, shared with every other user of the
/// mount, so it is never ours to close. Patient, because the harness starts the
/// registrar and this fixture together and a first resolve can land before init
/// has mounted `/protocol` at all.
fn registryEndpoint() ?ipc.Handle {
var attempt: u32 = 0;
while (attempt < resolve_attempts) : (attempt += 1) {
var relative: [channel.path_maximum]u8 = undefined;
if (file_system.fsResolve(channel.root, 0, &relative)) |route| switch (route) {
.kernel => return null, // a kernel route means something other than the registry owns the name
.backend => |backend| return backend.handle,
};
time.sleepMillis(resolve_retry_ms);
}
return null;
}
/// One `readdir(cursor)` at the registry, into `into`. Null at end of directory
/// or on any failure — the caller is walking a listing, and both mean "stop".
///
/// The listing is what makes this test un-driftable: `/protocol` publishes what
/// is bound (protocol-namespace.md — the tree stays diagnosable), so the set
/// under test is the set this boot actually produced.
fn entryAt(registry: ipc.Handle, cursor: u64, into: []u8) ?[]u8 {
var packet: [vfs_protocol.message_maximum]u8 = undefined;
const framed = vfs_protocol.Protocol.encodeRequest(.readdir, 0, .{ .cursor = cursor }, &.{}, &packet) orelse return null;
var reply: [vfs_protocol.message_maximum]u8 = undefined;
const got = ipc.callCap(registry, framed, &reply, null) catch return null;
// A readdir owes no capability; one that arrived anyway is a handle slot.
if (got.cap) |handle| _ = ipc.close(handle);
const answer = reply[0..got.len];
const status = envelope.statusOf(answer) orelse return null;
if (status.status != 0) return null;
const entry = vfs_protocol.Protocol.decodeReply(.readdir, answer) orelse return null;
if (entry.name_len == 0) return null; // end of directory
const text = vfs_protocol.Protocol.replyTail(.readdir, answer);
const length = @min(@as(usize, entry.name_len), @min(text.len, into.len));
@memcpy(into[0..length], text[0..length]);
return into[0..length];
}
/// The listing, taken once so every later question is asked of one observation
/// rather than of a namespace that may have moved underneath it.
const maximum_listed: usize = 32;
var listed_names: [maximum_listed][channel.name_maximum]u8 = undefined;
var listed_lengths: [maximum_listed]usize = undefined;
var listed_count: usize = 0;
fn listedName(index: usize) []const u8 {
return listed_names[index][0..listed_lengths[index]];
}
fn takeListing(registry: ipc.Handle) void {
listed_count = 0;
var cursor: u64 = 0;
while (cursor < maximum_listed) : (cursor += 1) {
const name = entryAt(registry, cursor, &listed_names[listed_count]) orelse return;
listed_lengths[listed_count] = name.len;
listed_count += 1;
}
}
/// Whether `/protocol` currently lists `name`.
fn listed(registry: ipc.Handle, name: []const u8) bool {
var cursor: u64 = 0;
while (cursor < maximum_listed) : (cursor += 1) {
var scratch: [channel.name_maximum]u8 = undefined;
const entry = entryAt(registry, cursor, &scratch) orelse return false;
if (std.mem.eql(u8, entry, name)) return true;
}
return false;
}
/// Wait until `/protocol` lists `name` — the providers this case needs come up
/// alongside the fixture, and racing them would make the listing a boot race
/// rather than an observation.
fn awaitListed(registry: ipc.Handle, name: []const u8) void {
var attempts: u32 = 0;
while (attempts < 400) : (attempts += 1) {
if (listed(registry, name)) return;
time.sleepMillis(20);
}
report("protocol-conformance: FAIL /protocol never listed {s}\n", .{name});
process.exit(1);
}
// --- the assertions ----------------------------------------------------------
/// The three checks, against one open channel. Every failure is fatal: the point
/// of the case is that these hold at *every* provider, so one that does not is
/// not a degraded result, it is the regression.
fn conform(link: channel.Channel, contract: Contract) void {
var buffer: [envelope.packet_maximum]u8 = undefined;
// 1. The reserved verb no provider implements. `describe` is answered from
// the specification by the generated dispatch, so what comes back is the
// contract's own identity — checked field by field against the module
// this fixture compiled against.
const described = link.describe(&buffer) orelse fail("describe was not answered");
if (!std.mem.eql(u8, described.name, contract.name)) fail("describe named a different protocol");
if (described.description.version != contract.version) fail("describe answered the wrong version");
if (described.description.operation_count != contract.operations) fail("describe counted the wrong number of verbs");
// 2. A number no protocol wears. Nothing in the provider looks at it; the
// dispatch table finds no handler and refuses, identically everywhere.
var into: [envelope.packet_maximum]u8 = undefined;
const answered = link.call(.{ .operation = stranger_operation }, &.{}, &into) orelse
fail("a stranger verb was not answered at all");
if (answered.status.status != -envelope.ENOSYS) fail("a stranger verb did not answer -ENOSYS");
if (answered.status.len != 0) fail("a refused verb promised a payload");
if (answered.capability) |handle| {
_ = ipc.close(handle);
fail("a refused verb handed back a capability");
}
// 3. A refusal is an answer, not a wedge — so the same channel still works.
const again = link.describe(&buffer) orelse fail("the provider stopped answering after a refused verb");
if (!std.mem.eql(u8, again.name, contract.name)) fail("describe changed its answer after a refused verb");
report("protocol-conformance: {s} v{d} describes itself ({d} verbs), verb {d} -> -ENOSYS\n", .{
contract.name,
contract.version,
contract.operations,
stranger_operation,
});
}
fn contractIndex(name: []const u8) ?usize {
for (contracts, 0..) |contract, index| {
if (std.mem.eql(u8, contract.name, name)) return index;
}
return null;
}
fn run() void {
const registry = registryEndpoint() orelse fail("resolve /protocol");
// Every contract this scenario is supposed to be able to check must be bound
// before the listing is taken, or the case would assert nothing on a slow
// boot instead of failing on a broken one.
for (contracts) |contract| {
if (contract.required) awaitListed(registry, contract.name);
}
takeListing(registry);
if (listed_count == 0) fail("/protocol listed nothing at all");
report("protocol-conformance: /protocol lists {d} contract(s)\n", .{listed_count});
var checked = [_]bool{false} ** contracts.len;
for (0..listed_count) |index| {
const name = listedName(index);
const found = contractIndex(name) orelse {
// Not a lie of omission: named on serial, with the reason.
report("protocol-conformance: {s} skipped — not built on envelope.Define yet\n", .{name});
continue;
};
const contract = contracts[found];
const link = channel.Channel.connect(name) orelse {
// P3 is in force: an ungranted name is absent for this binary, and
// that is a manifest fact, not a failure — unless the scenario is
// supposed to have granted it.
if (contract.required) fail("a contract this fixture is granted would not open");
report("protocol-conformance: {s} skipped — not granted to this fixture\n", .{name});
continue;
};
conform(link, contract);
link.close();
checked[found] = true;
}
// The vacuity guard, and the honest tail: a required contract that went
// unchecked fails the case, and every other one this fixture knows how to
// check but found no provider for is named, so the coverage is legible on
// serial rather than inferred from what is absent.
var count: usize = 0;
for (contracts, 0..) |contract, index| {
if (checked[index]) {
count += 1;
continue;
}
if (contract.required) fail("a contract this scenario boots was never conformance-checked");
report("protocol-conformance: {s} not bound in this scenario — no provider to ask\n", .{contract.name});
}
report("protocol-conformance: {d} provider(s) answered the reserved verbs identically\n", .{count});
_ = logging.write("protocol-conformance: ok\n");
}
pub fn main(startup: process.Init) void {
const role = startup.arguments.get(1) orelse return; // bare (ramdisk sweep): stay silent
if (std.mem.eql(u8, role, "run")) run();
}