Compare commits

...
Author SHA1 Message Date
jschoubben 05fb7fb5eb Merge pull request 'The mesh makes the bus's certificate itself' (#145) from fix/the-mesh-makes-its-own-bus-certificate into main 2026-09-29 14:06:31 +00:00
jschoubben 2c1733de8d The mesh makes the bus's certificate itself
novox/hq 04-ISSUES/146. The foundation made it by running openssl inside the
broker's image, which worked while the broker was one that carried it and
stopped the day the bus changed: the new one has a shell and no openssl, so
the step exited 127 and no mesh could be raised. No other image the bundle
names has it either, so there was nothing to substitute.

broker certificate --into <dir> writes the pair, --check is the step's verify.
Self-signed on purpose — a host pins this server's exact certificate (ADR
0004) and at genesis there is no authority to ask — and made once, because a
second certificate is one every host that pinned the first no longer believes.
The key is written before the certificate, so an interruption never leaves
something that looks finished.
2026-09-29 15:42:51 +02:00
jschoubben e51c94dcb5 The trust module renders the authority it was bound to
novox/hq ADR 0147. ca-trust carries a script and a unit; the one thing
neither can state is where the authority is, because that is a fact about
the mesh. This checks the rendering — the script fetches from the bound
address and is executable, and the unit runs it both ways, install and
remove. The verification itself is the lab's.
2026-09-29 15:07:33 +02:00
mesh-admin 07c07902ff Merge pull request 'Anything on this machine may call anything on this machine' (#143) from fix/local-calls-are-not-filtered into main 2026-09-29 11:30:34 +00:00
jschoubben 864cdea4c6 Anything on this machine may call anything on this machine
Local is not a boundary this mesh draws. A service here is callable by everything
else here, whatever form either takes — a package with a unit, a binary, a
container. Whether a caller sits in a container was never meant to change the
answer, and the only reason it did was that this chain asked about addresses: a
caller on the machine carries the machine's address, a caller in one of its
containers carries a bridge address, and a rule naming the former silently refused
the latter.

One rule for every service here, replacing the line-per-port added an hour ago,
which only ever covered the ports somebody remembered to think about. The three
reaches are now three lines: on this machine, over the private network, from
anywhere.

The tests assert per chain body, because the forward chain carries the same line in
the same words and an assertion on the whole file passed with the input chain's copy
deleted — which is what ADR 0137's own tests say to do and this file was not doing.
2026-09-29 13:30:32 +02:00
mesh-admin 6e810907b2 Merge pull request 'This machine's own guests are on the private network' (#142) from fix/this-machines-own-guests-are-on-the-private-network into main 2026-09-29 10:30:32 +00:00
jschoubben 2b20a12c4a This machine's own guests are on the private network
A port declared from the mesh admitted the machines' own addresses on the private
network. A container reaching a port on the machine it runs on comes from a bridge,
matching none of them — and where the container runtime routes directly, that packet
is delivered to this machine rather than forwarded, so the forward chain's allowance
never saw it either.

ADR 0100 requires this to work: the store is reachable 'from a container on the node
itself'. It was, through a rule the predecessor left, which allowed the private
ranges wholesale. Converging the machine replaced that with the four overlay
addresses and closed it.

Measured, and it was an outage: every module reaching another by its machine's own
name timed out for eleven hours while the mesh reported the machine healthy. A web
application logged 'connection to server at novox.internal (10.10.0.1), port 6852
failed: timeout expired' throughout.

Asked for by the link it arrives on, for the reason the forward chain no longer
names an address: a range describes one machine and goes stale in silence. A port
open to everything needs no such line.
2026-09-29 12:30:15 +02:00
mesh-admin 9c83dacfce Merge pull request 'A route that names an endpoint still carries that endpoint's port' (#141) from fix/an-endpoint-named-by-a-route-still-carries-its-port into main 2026-09-29 09:55:45 +00:00
jschoubben 64ba053f3b A route that names an endpoint still carries that endpoint's port
Everything downstream reads the port: the provider is told where to reach the
consumer, and the redirection that turns a declared port into the number the
machine published is keyed on it. A route naming only its endpoint left the proxy
with no port at all, and a proxy with no port has nothing to dial.

Caught after the catalogue had already been changed to name endpoints and before
the mesh picked those manifests up, which is the only reason nothing broke: every
module's manifest is behind its source right now, so the plan still renders from
the old shape.

The declared port, not the machine one — the redirection happens later and is keyed
on the declared number, so filling in the machine port here would be redirected
twice or not at all. A route that repeats a port keeps it.
2026-09-29 11:55:28 +02:00
mesh-admin 96416bd8a7 Merge pull request 'Run the real catalogue through the real manifest gate' (#140) from feat/an-assignment-configures-an-endpoint into main 2026-09-29 09:49:29 +00:00
jschoubben 4d2003d77b Run the real catalogue through the real manifest gate
A test that reads every manifest in a catalogue checkout and parses it with the
control plane's own parser, rather than asserting against a fixture: whether the
manifests as written are accepted is the question, and a copy of one proves nothing
about the other seventy-one.

Skipped unless MESH_CATALOGUE names a checkout, so it costs nothing in ordinary
runs and is there when the catalogue changes shape. It also refuses to pass if no
endpoint is named, because a run that validated nothing would otherwise read as
success.
2026-09-29 11:49:22 +02:00
mesh-admin aaad02fd38 Merge pull request 'An assignment configures an endpoint as one thing' (#139) from feat/an-assignment-configures-an-endpoint into main 2026-09-29 09:25:55 +00:00
jschoubben c68d3a7432 An assignment configures an endpoint as one thing
novox/hq ADR 0138, completing it. One block per endpoint instead of three keys
joined by a port number:

  {"endpoints": {"web":    {"port": 20009, "label": "cinema", "reach": "both"},
                 "stream": {"reach": "internal"}}}

Which machine port it lands on, the subdomain a proxy serves it under, and how far
it reaches are the three things an operator says when a module is assigned, and they
were said in ports, in the route's label and in reach — each keyed by the port. A
module with two endpoints of different shapes could only be configured by a reader
who knew which number was which.

Every field is optional; a block that says only a reach leaves the port to the mesh
and the label to the module, which is the ordinary case. A name the module does not
declare is refused, and the refusal lists what it does declare. A port or a reach
said both here and through the older key is refused rather than merged — two places
saying one thing is what this key exists to end, and merging would follow whichever
was read last.

Eight tests. The reach assertion deliberately narrows what the manifest says, because
a reach that agrees with the manifest proves nothing about whether the block was read
— which I found by writing the weaker version first and watching a revert not fail.
2026-09-29 11:25:40 +02:00
mesh-admin a5209bd849 Merge pull request 'A module names its endpoints, and a route names the one it serves' (#138) from feat/a-module-names-its-endpoints into main 2026-09-29 07:28:41 +00:00
jschoubben bdf965dab6 A module names its endpoints, and a route names the one it serves
novox/hq ADR 0138's remaining half, and the words ship one release before any
manifest uses them.

A port number is not a name. Three facts have to be said about an endpoint when a
module is assigned — which machine port it lands on, the subdomain a proxy serves
it under, and how far it reaches — and they were said in three places keyed by the
port. A module with two endpoints of different shapes cannot be configured that way
without a reader joining numbers by hand: a web surface behind the proxy, whose
port only the proxy need reach, and a protocol port clients dial directly because
the client expects that number.

So a listen carries a name, lowercase and unique within the module, and a route
names the endpoint it serves instead of repeating its port. Two endpoints with one
name are refused, because an assignment configuring one would silently configure
whichever the mesh read last. A route naming an endpoint the module does not declare
is refused where it is written rather than resolving to no port and serving nothing.

An unnamed endpoint stays valid and a route repeating a port still resolves, which
is every module in the catalogue today.
2026-09-29 09:28:24 +02:00
mesh-admin b4da20ecc0 Merge pull request 'Reach asks for names on a routed endpoint' (#137) from fix/reach-names-a-route-not-a-port into main 2026-09-29 00:53:28 +00:00
jschoubben 4b33b72160 Reach asks for names on a routed endpoint, and its port stays the manifest's
A routed endpoint's port is how the proxy reaches it and nothing else (ADR 0045):
a public service listens from the mesh, only the proxy reaches it, and it is
exposed by name. So reach on a routed endpoint asks for names, and the port keeps
what the manifest said; on an unrouted one — git over ssh, a mail port, the bus —
it governs the port, because there is no name and the port is the only way in.

Found by trying to express a real module rather than by review: routed name public
because browsers post to it, machine-side port private because it serves a
dashboard in cleartext. Under one value for both there was no way to say it, and
'public' would have reopened a port narrowed an hour earlier.

novox/hq ADR 0138, corrected in place the same day.
2026-09-29 02:50:41 +02:00
mesh-admin d5505fe3d4 Merge pull request 'An assignment says how far an endpoint reaches' (#136) from feat/an-assignment-says-how-far-an-endpoint-reaches into main 2026-09-29 00:47:27 +00:00
jschoubben 264c9e41e9 An assignment says how far an endpoint reaches, and three things read it
novox/hq ADR 0138. Reachability was settled three times over: the filter read a
listen's source with expose able to override it; the proxy composed a public name
and an internal name for every route it was given, because it could; and the
certificate authority followed from which names existed. Each was defensible and
the combination was unstated, so "this endpoint must not be public" could not be
written and was enforced by nothing — while a public certificate for that name was
obtained anyway. Measured on the control node: an identity provider holding a
90-day public certificate and a 24-hour internal one, neither asked for.

`reach` is one value per endpoint, per node — machine, internal, public or both —
and the filter's source and the composed names both follow it. The authority needs
no work: the proxy already asks the public authority for a route's own name and its
internal authority for the internal one, so controlling the names controls the
authority.

Joined by the port, which a route already names: 35 of the catalogue's 36 route
entries name a port the same module declares a listen on, and the one that does not
is a path-level refusal — a rule about a name rather than an endpoint, left alone.

Nothing said composes both names and follows the manifest's `from`, so every mesh
already running is unchanged until an assignment speaks. A port that says both
reach and expose is refused: they say the same thing in different words, and the
filter would follow one while the names followed the other.
2026-09-29 02:47:06 +02:00
mesh-admin 76ac3c99bd Merge pull request 'The filter constrains what arrives from outside, and names no network' (#135) from feat/filter-what-arrives-from-outside into main 2026-09-28 23:01:53 +00:00
jschoubben fe5988c536 The filter constrains what arrives from outside, and names no network
The forward chain blocked everything passing through the machine and then allowed
the machine's own containers back by naming their address ranges: 172.16.0.0/12 and
192.168.128.0/17 fixed here, the rest recorded per machine by 0043. Every way of
keeping that list correct fails — a constant describes one machine, a recorded range
goes stale in silence and cannot tell a network the mesh made from one a predecessor
left behind, and generating it from the modules would put half the rule set on the
machine.

The mesh has no position on a container reaching outward: that is not a port opened
to anybody. So both chains are written around the links traffic arrives on. What did
not arrive from outside is accepted in one line; what did meets the declared rules.
The tunnel is named beside the outward links rather than treated as inside, or a port
nothing declares would be reachable from every machine in the mesh.

A machine that has not reported an outward link is sent no filter and keeps the one
it has, refused where a person reads it rather than as a rule set that will not load.

Removes the two constants, `node networks`, and the column behind it. novox/hq ADR
0140, superseding 0137 and 0139.
2026-09-29 01:01:29 +02:00
mesh-admin ed5d467d90 Merge pull request 'An artifact says which operating system it is built for' (#134) from feat/an-artifact-says-which-system-it-is-built-for into main 2026-09-28 22:54:42 +00:00
jschoubben 228d0226dd An artifact says which operating system it is built for
First of the steps in novox/hq ADR 0142, and it ships alone: a new manifest word
reaches the builder and the controller one release before any manifest uses it.

A toolchain deliberately accepts nothing from the module — anything a module
could override there it would be writing a Dockerfile to override — and yet a
compiled binary is per operating system, pinned at link time so a host refuses to
touch a machine it was not built for (ADR 0005). The way out is that the target
belongs to the artifact: one artifact per system, one build each, recipe still the
mesh's.

A bundle in a language that compiles to a binary must name a system, or it would
be built for whatever the build machine happened to be — which reads as portable
and is not. A bundle in a language that runs anywhere may not name one, because a
system that decides nothing reads as though it did. The list is the host's own
names, not a compiler's: the difference between two of them is a C library rather
than a kernel.

Nothing declares a system yet, and no toolchain compiles to a binary yet, so this
changes no build.
2026-09-29 00:54:27 +02:00
mesh-admin 6215ff0760 Merge pull request 'A machine says which networks it routes, and its filter forwards them' (#133) from feat/a-machine-says-which-networks-it-routes into main 2026-09-28 19:31:06 +00:00
jschoubben 54812306be A machine says which networks it routes, and its filter forwards them
The derived filter denies forwarding by default and then allows the container
runtime's two default pools, named in this code with a comment saying a machine
configured otherwise needs to say so -- and no way to say it. So the filter was
right on a machine using the defaults and silently wrong on any other.

Measured today: flipping a workstation to the derived filter cut egress for five
of its container networks and for every network its test beds create, because
those come from ranges the defaults do not cover. Nothing reported a fault; the
guests just could not reach anything, while the machine reported it had applied
what it was told.

A node-level fact beside the public domain, because the machine routes them and
the module that loads the filter may be replaced. Added to the defaults, never
replacing them. Their guests also keep address and name service, without which a
network does not work at all, and the converge preview now says what a machine
routes instead of leaving it to a sentence about what it cannot preview.
2026-09-28 21:29:10 +02:00
mesh-admin ce9e20fbbc Merge pull request 'A push raises what a module hears, not only a start' (#132) from fix/a-push-raises-what-a-module-hears into main 2026-09-28 14:46:54 +00:00
27 changed files with 2206 additions and 64 deletions
+23 -1
View File
@@ -309,7 +309,7 @@ func converge(ctx context.Context, open *stores, node string, yes bool, digest s
return "", err
}
derived := derivedFilter{rules: rules, foundation: with.Foundation, mesh: with.Mesh,
outward: plan.PublicDomain != ""}
outward: plan.PublicDomain != "", outwardLinks: with.OutwardLinks}
preview, saw := previewOf(node, reported, derived, plan, taken, filter, runs[filter])
preview += "\n\n preview " + saw
if !yes {
@@ -414,6 +414,18 @@ func previewOf(node string, reported inventory.Adoption, derived derivedFilter,
b.WriteString(" not previewed: traffic the machine routes that is not a published port " +
"(a tunnel, NAT in the found firewall) — the derived filter drops it unless a module " +
"declares it\n")
// Which links the filter constrains, said rather than left to the sentence above (novox/hq ADR
// 0140). Everything arriving anywhere else is this machine's own guest and keeps working — which
// is what a reader most wants to know, because the previous shape of this filter cut a machine's
// guests off at the flip without saying so, and that is how this was found.
if len(derived.outwardLinks) > 0 {
b.WriteString(fmt.Sprintf(" it filters what arrives on: %s, and on the private network "+
"— everything its own guests send keeps working\n",
strings.Join(derived.outwardLinks, ", ")))
} else {
b.WriteString(" it has reported no link facing outside, so no filter can be composed " +
"for it — the flip is refused until it reports one\n")
}
isTaken := map[string]bool{}
for _, m := range taken {
@@ -473,6 +485,10 @@ type derivedFilter struct {
// mesh is every address on the private network; outward says the machine faces outside.
mesh []string
outward bool
// outwardLinks is the links this machine reported as facing outside it (novox/hq ADR 0140).
// The filter constrains what arrives on them; everything arriving elsewhere is this machine's
// own guest and is not filtered.
outwardLinks []string
}
// closesOutside is what a narrowing from everywhere to the private network is called: it closes.
@@ -498,6 +514,12 @@ func (d derivedFilter) fate(r inventory.Reach) string {
return "stays open — the mesh's own, from anywhere"
}
}
// This machine's own guests ask it for an address and for names, and those two arrive here
// (novox/hq ADR 0140). Admitted by the link they arrive on, so a listener bound anywhere but an
// outward link keeps answering them.
if (r.Protocol == "udp" && (r.Port == 53 || r.Port == 67)) || (r.Protocol == "tcp" && r.Port == 53) {
return "stays open — this machine's own guests asking it for an address and for names"
}
for _, rule := range d.rules {
if rule.Port != r.Port || rule.Protocol != r.Protocol {
continue
+171
View File
@@ -0,0 +1,171 @@
package main
import (
"crypto/rand"
"crypto/rsa"
"crypto/x509"
"crypto/x509/pkix"
"encoding/pem"
"errors"
"fmt"
"math/big"
"net"
"os"
"path/filepath"
"time"
)
// The bus's own certificate, made by the mesh rather than borrowed from an image.
//
// **The foundation asked a third-party image for a tool it never said must be there** (novox/hq
// 04-ISSUES/146). The bootstrap made this certificate by running `openssl` inside the broker's
// image, which worked while the broker was one that happened to carry it and stopped the day the
// bus changed: the new one has a shell and no openssl, so the step exited 127 and no mesh could be
// raised. Substituting another image the bundle names does not help — none of them carry it
// either.
//
// So the program that needs a certificate makes one. It is the mesh's own binary, already on the
// machine at this point in the bootstrap (the schema step ran it), and it needs nothing from the
// image it writes into but a mounted directory.
//
// **Self-signed, and that is the design** — a host pins this server's exact certificate and
// authenticates with a password (novox/hq ADR 0004). There is no authority above it to ask, and at
// this moment in a bootstrap there is no mesh to ask one of.
//
// Idempotent, because the step is applied again on every reconcile and a second certificate would
// be one the hosts that pinned the first no longer believe.
// busCertificateNames is what the bus is reached by: the container name on a mesh network, and the
// loopback address the machine's own foundation dials.
var busCertificateNames = []string{"mesh-broker"}
const busCertificateLife = 10 * 365 * 24 * time.Hour
// busCertificate makes the bus's certificate in a directory, or says whether one is there.
//
// broker certificate --into /tls make it if it is not there
// broker certificate --check --into /tls exit non-zero unless a usable pair is
func busCertificate(args []string) error {
into, check := "", false
for i := 0; i < len(args); i++ {
switch args[i] {
case "--check":
check = true
case "--into":
if i+1 >= len(args) {
return errors.New("--into needs a directory")
}
into = args[i+1]
i++
default:
return fmt.Errorf("broker certificate [--check] --into <directory>: %q", args[i])
}
}
if into == "" {
return errors.New("broker certificate [--check] --into <directory>")
}
crt, key := filepath.Join(into, "tls.crt"), filepath.Join(into, "tls.key")
if usable, err := busCertificateUsable(crt, key); err != nil {
return err
} else if usable {
fmt.Printf("the bus already has a certificate at %s, and it was left alone\n", crt)
return nil
}
if check {
// Said as a failure, because that is what the caller asked: a bootstrap's verify runs
// this and a false answer is what makes the step run.
return fmt.Errorf("no usable certificate and key at %s", into)
}
return writeBusCertificate(crt, key)
}
// busCertificateUsable says whether a certificate and its key are both there and parse.
//
// Both, and parsed rather than stat'ed: a half-written pair is the state a bootstrap interrupted
// between the two files leaves behind, and a step that treated it as done would hand the server a
// certificate with no key and report success.
func busCertificateUsable(crt, key string) (bool, error) {
certPEM, err := os.ReadFile(crt)
if errors.Is(err, os.ErrNotExist) {
return false, nil
}
if err != nil {
return false, err
}
keyPEM, err := os.ReadFile(key)
if errors.Is(err, os.ErrNotExist) {
return false, nil
}
if err != nil {
return false, err
}
if _, err := tlsPairParses(certPEM, keyPEM); err != nil {
return false, nil
}
return true, nil
}
func tlsPairParses(certPEM, keyPEM []byte) (*x509.Certificate, error) {
block, _ := pem.Decode(certPEM)
if block == nil || block.Type != "CERTIFICATE" {
return nil, errors.New("not a certificate")
}
certificate, err := x509.ParseCertificate(block.Bytes)
if err != nil {
return nil, err
}
keyBlock, _ := pem.Decode(keyPEM)
if keyBlock == nil {
return nil, errors.New("not a key")
}
if _, err := x509.ParsePKCS8PrivateKey(keyBlock.Bytes); err != nil {
if _, err := x509.ParsePKCS1PrivateKey(keyBlock.Bytes); err != nil {
return nil, err
}
}
return certificate, nil
}
func writeBusCertificate(crt, key string) error {
private, err := rsa.GenerateKey(rand.Reader, 2048)
if err != nil {
return err
}
serial, err := rand.Int(rand.Reader, new(big.Int).Lsh(big.NewInt(1), 128))
if err != nil {
return err
}
template := &x509.Certificate{
SerialNumber: serial,
Subject: pkix.Name{CommonName: busCertificateNames[0]},
DNSNames: busCertificateNames,
IPAddresses: []net.IP{net.ParseIP("127.0.0.1")},
NotBefore: time.Now().Add(-time.Hour),
NotAfter: time.Now().Add(busCertificateLife),
KeyUsage: x509.KeyUsageDigitalSignature | x509.KeyUsageKeyEncipherment,
ExtKeyUsage: []x509.ExtKeyUsage{x509.ExtKeyUsageServerAuth},
BasicConstraintsValid: true,
}
der, err := x509.CreateCertificate(rand.Reader, template, template, &private.PublicKey, private)
if err != nil {
return err
}
pkcs8, err := x509.MarshalPKCS8PrivateKey(private)
if err != nil {
return err
}
// **The key first, and only then the certificate**, so the pair a reader finds is never a
// certificate whose key has not been written yet — the one order in which an interruption
// leaves something that looks finished (novox/hq 04-ISSUES/014, a key present and unusable).
if err := os.WriteFile(key, pem.EncodeToMemory(&pem.Block{Type: "PRIVATE KEY", Bytes: pkcs8}), 0o600); err != nil {
return err
}
if err := os.WriteFile(crt, pem.EncodeToMemory(&pem.Block{Type: "CERTIFICATE", Bytes: der}), 0o644); err != nil {
return err
}
fmt.Printf("made the bus a certificate for %v, valid until %s\n %s\n %s\n",
busCertificateNames, template.NotAfter.Format(time.RFC3339), crt, key)
return nil
}
+121
View File
@@ -0,0 +1,121 @@
package main
import (
"crypto/tls"
"crypto/x509"
"os"
"path/filepath"
"strings"
"testing"
)
// novox/hq 04-ISSUES/146. The bootstrap could not make the bus a certificate: it asked an image for
// `openssl` and the image it asks has none. What replaces it is this command, so what is checked is
// what the bootstrap needs from it — a pair a TLS server can actually load, made once and only once.
func TestTheBusCertificateLoadsAsAServersWould(t *testing.T) {
into := t.TempDir()
if err := busCertificate([]string{"--into", into}); err != nil {
t.Fatalf("the bus could not be given a certificate: %v", err)
}
// The check a cheaper test would not make. The key was present and valid and the server could
// not start, once, because nothing loaded the pair the way a server loads it
// (novox/hq 04-ISSUES/014).
pair, err := tls.LoadX509KeyPair(filepath.Join(into, "tls.crt"), filepath.Join(into, "tls.key"))
if err != nil {
t.Fatalf("a TLS server cannot load what was written: %v", err)
}
leaf := pair.Leaf
if leaf == nil {
if leaf, err = x509.ParseCertificate(pair.Certificate[0]); err != nil {
t.Fatal(err)
}
}
if err := leaf.VerifyHostname("mesh-broker"); err != nil {
t.Errorf("the certificate is not for the name the bus is reached by: %v", err)
}
if len(leaf.IPAddresses) == 0 || leaf.IPAddresses[0].String() != "127.0.0.1" {
t.Errorf("the certificate does not cover the loopback address the foundation dials: %v", leaf.IPAddresses)
}
// The key is not readable by anything else on the machine; the certificate is public and is.
key, err := os.Stat(filepath.Join(into, "tls.key"))
if err != nil {
t.Fatal(err)
}
if key.Mode().Perm() != 0o600 {
t.Errorf("the key is %v", key.Mode().Perm())
}
crt, err := os.Stat(filepath.Join(into, "tls.crt"))
if err != nil {
t.Fatal(err)
}
if crt.Mode().Perm() != 0o644 {
t.Errorf("the certificate is %v, which the server runs as another user cannot read", crt.Mode().Perm())
}
}
// **Made once.** The step is applied again on every reconcile, and a second certificate is one the
// hosts that pinned the first no longer believe (novox/hq ADR 0004).
func TestTheBusCertificateIsMadeOnce(t *testing.T) {
into := t.TempDir()
if err := busCertificate([]string{"--into", into}); err != nil {
t.Fatal(err)
}
first, err := os.ReadFile(filepath.Join(into, "tls.crt"))
if err != nil {
t.Fatal(err)
}
if err := busCertificate([]string{"--into", into}); err != nil {
t.Fatal(err)
}
again, err := os.ReadFile(filepath.Join(into, "tls.crt"))
if err != nil {
t.Fatal(err)
}
if string(first) != string(again) {
t.Fatal("running it twice replaced the certificate every host had pinned")
}
}
// The verify half: false before, true after, which is what makes the bootstrap run the step at all.
func TestTheCheckIsFalseUntilThereIsAPair(t *testing.T) {
into := t.TempDir()
if err := busCertificate([]string{"--check", "--into", into}); err == nil {
t.Fatal("an empty directory reported a usable certificate")
}
if err := busCertificate([]string{"--into", into}); err != nil {
t.Fatal(err)
}
if err := busCertificate([]string{"--check", "--into", into}); err != nil {
t.Fatalf("the certificate it just made does not satisfy its own check: %v", err)
}
}
// A half-written pair is not a pair. An interrupted bootstrap leaves exactly this, and a step that
// called it done would hand the server a certificate with no key and report success.
func TestACertificateWithoutItsKeyIsNotUsable(t *testing.T) {
into := t.TempDir()
if err := busCertificate([]string{"--into", into}); err != nil {
t.Fatal(err)
}
if err := os.Remove(filepath.Join(into, "tls.key")); err != nil {
t.Fatal(err)
}
if err := busCertificate([]string{"--check", "--into", into}); err == nil {
t.Fatal("a certificate with no key passed the check")
}
if err := busCertificate([]string{"--into", into}); err != nil {
t.Fatal(err)
}
if _, err := tls.LoadX509KeyPair(filepath.Join(into, "tls.crt"), filepath.Join(into, "tls.key")); err != nil {
t.Fatalf("it did not replace the unusable pair: %v", err)
}
}
func TestWhereToWriteIsRequired(t *testing.T) {
if err := busCertificate(nil); err == nil || !strings.Contains(err.Error(), "--into") {
t.Fatalf("it did not ask where to write: %v", err)
}
}
+3
View File
@@ -148,6 +148,9 @@ func usage() {
node public-domain <name> the domain it composes its routed names under
node public-domain <name> <d> ...set it to d
node public-domain <name> --clear ...it faces the outside no longer
node networks <name> the networks it routes for what it hosts
node networks <name> <cidr>... ...set them; its filter forwards these too
node networks <name> --clear ...only the container runtime's own
token issue --node <name> a one-time right to join, for an existing record
token issue --new <name> create the record and issue for it
token issue ... --adopted ...for a machine in use, which joins adopted
+16 -1
View File
@@ -66,6 +66,18 @@ func nodeCommand(ctx context.Context, args []string) error {
// because the damage is already done by the time it prints.
return publicDomain(ctx, inv, args[1:])
case "networks":
// Removed by novox/hq ADR 0140, which superseded the record that added it. The filter no
// longer names any network: it constrains what arrives from outside the machine and says
// nothing about what did not, so there is no list to keep. Answered rather than met with
// "unknown command", because this was the documented way to stop a flip cutting a machine's
// containers off and somebody will reasonably still type it.
return errors.New("`node networks` is gone (novox/hq ADR 0140). The filter constrains what " +
"arrives from outside this machine and says nothing about traffic that did not, so no " +
"network is named anywhere and nothing needs to be said to keep a machine's own " +
"containers reaching outward. The machine reports which of its links face outside; see " +
"`node show <name>`")
case "account":
// The operator's login on this machine (novox/hq to-be 29): what a home-scoped file is
// owned by and which account `ssh <node>` uses. Reports with no argument; sets with one;
@@ -352,8 +364,11 @@ func identityCommand(ctx context.Context, args []string) error {
}
func brokerCommand(args []string) error {
if len(args) > 0 && args[0] == "certificate" {
return busCertificate(args[1:])
}
if len(args) == 0 || args[0] != "show" {
return errors.New("broker show")
return errors.New("broker show | broker certificate [--check] --into <directory>")
}
known, err := broker.FromEnvironment()
if errors.Is(err, broker.ErrNotConfigured) {
+8 -2
View File
@@ -640,13 +640,19 @@ func renderingFor(ctx context.Context, open *stores, node string,
if err != nil {
return catalogue.Rendering{}, inventory.Node{}, err
}
// Which of this machine's links face outside, which is what the derived filter is written
// around (novox/hq ADR 0140). Reported by the machine, never set.
outwardLinks, err := inv.OutwardLinksOf(ctx, node)
if err != nil {
return catalogue.Rendering{}, inventory.Node{}, err
}
return catalogue.Rendering{
BusMembership: memberships[node],
Settings: settings, Generators: gens, Grants: grants, Needed: needed, Ports: ports,
Certificate: certificate, Authority: authority, Mesh: private, Names: names,
Machines: machines,
Suffix: overlay.Suffix(), MeshRange: meshRange, Accounts: accounts, Foundation: foundation,
Kept: kept, Adopted: record.Adopted,
Suffix: overlay.Suffix(), MeshRange: meshRange, TunnelInterface: overlay.Interface, Accounts: accounts, Foundation: foundation,
Kept: kept, Adopted: record.Adopted, OutwardLinks: outwardLinks,
Given: given, Taken: taken, Seats: seats, ArtifactStore: artifactStore, Built: built,
BusUsers: busUsers,
}, record, nil
+17 -9
View File
@@ -59,7 +59,11 @@ func anchorRendering(adopted bool) Rendering {
Values: map[string]any{ExposeSetting: map[string]any{"5000": FromEverywhere}}}}},
Mesh: []string{"10.42.0.1"},
Foundation: []int{5671},
Adopted: adopted,
// What the machine reported faces outside, which every rule in the filter is written
// around (novox/hq ADR 0140).
OutwardLinks: []string{"eth0"},
TunnelInterface: "mesh0",
Adopted: adopted,
// Genesis takes the foundation's modules.
Taken: map[string]bool{"postgres": true, "lavinmq": true},
}
@@ -575,11 +579,13 @@ func TestAGivenMachineSideReachesTheFilterTheOpeningAndTheConsumer(t *testing.T)
}
r := Resolution{Node: "anchor", Modules: []Manifest{forge}}
with := Rendering{
Ports: map[string]map[int]int{"forge": portsAsThePlanWould(forge, given)},
Given: map[string]map[int]int{"forge": given},
Mesh: []string{"10.77.0.1"},
Adopted: true,
Taken: map[string]bool{"forge": true},
Ports: map[string]map[int]int{"forge": portsAsThePlanWould(forge, given)},
Given: map[string]map[int]int{"forge": given},
Mesh: []string{"10.77.0.1"},
Adopted: true,
OutwardLinks: []string{"eth0"},
TunnelInterface: "mesh0",
Taken: map[string]bool{"forge": true},
}
// What the runtime is handed: the machine's own port on the outside, the container's within.
@@ -660,9 +666,11 @@ func TestALongFormPortIsOpenedWhereTheManifestPublishesIt(t *testing.T) {
forge := aForge()
r := Resolution{Node: "anchor", Modules: []Manifest{forge}}
composed, err := r.Compose(Rendering{
Ports: map[string]map[int]int{"forge": portsAsThePlanWould(forge, nil)},
Mesh: []string{"10.77.0.1"},
Adopted: true,
Ports: map[string]map[int]int{"forge": portsAsThePlanWould(forge, nil)},
Mesh: []string{"10.77.0.1"},
Adopted: true,
OutwardLinks: []string{"eth0"},
TunnelInterface: "mesh0",
})
if err != nil {
t.Fatal(err)
+59
View File
@@ -153,6 +153,26 @@ func (b *Build) problems(module string) []string {
"%s: %q is a bundle and says no language, so nothing can choose a compiler "+
"for it", module, a.Name))
}
// **A system, for a language that compiles to a binary** (novox/hq ADR 0142). A binary
// is pinned to one operating system at link time so a host refuses to touch a machine
// it was not built for (novox/hq ADR 0005); an artifact that says nothing would be
// compiled for whatever the build machine happened to be, which reads as portable and
// is not.
if compiled := compilesToABinary(a.Language); compiled && strings.TrimSpace(a.System) == "" {
problems = append(problems, fmt.Sprintf(
"%s: %q is compiled to a binary and says no system, so it would be built for "+
"whatever the build machine happens to be. Declare one artifact per "+
"system: %s", module, a.Name, spokenSystems()))
} else if !compiled && strings.TrimSpace(a.System) != "" {
problems = append(problems, fmt.Sprintf(
"%s: %q names the system %q and is written in %q, which compiles to code that "+
"runs anywhere — a system that decides nothing reads as though it did",
module, a.Name, a.System, a.Language))
} else if compiled && !knownSystem(a.System) {
problems = append(problems, fmt.Sprintf(
"%s: %q is built for %q, and a system is %s",
module, a.Name, a.System, spokenSystems()))
}
} else {
if a.From == "" {
problems = append(problems, fmt.Sprintf(
@@ -193,3 +213,42 @@ func oneOrOther(n int) string {
}
return "them"
}
// Systems the mesh builds binaries for, which is the set a host may be pinned to (novox/hq ADR 0005).
//
// **A closed list, and the host's own, not the compiler's.** These are not the values a Go toolchain
// would call an operating system — the difference between two of them is a C library, not a kernel.
// They are what a machine reports itself to be and what a host is linked to refuse, so the list that
// matters is the one the host understands.
var systems = []string{"alpine", "android", "arch"}
// knownSystem is whether the mesh builds for it.
func knownSystem(system string) bool {
want := strings.ToLower(strings.TrimSpace(system))
for _, s := range systems {
if s == want {
return true
}
}
return false
}
// spokenSystems is the list as a refusal says it, so a reader is one edit from right.
func spokenSystems() string {
return strings.Join(systems, ", ")
}
// compilesToABinary is whether this language's bundle is a binary for one operating system rather
// than code that runs wherever its interpreter does.
//
// **Asked of the language, not of the artifact.** A module says what it is written in; what that
// implies is the mesh's to know, exactly as the compiler is (novox/hq ADR 0142). Asking the artifact
// would let two artifacts in one language disagree about whether they are portable.
func compilesToABinary(language string) bool {
switch strings.ToLower(strings.TrimSpace(language)) {
case "go":
return true
default:
return false
}
}
+82
View File
@@ -0,0 +1,82 @@
package catalogue
import (
"strings"
"testing"
)
// bundleFor is a manifest whose one artifact is a bundle in the given language and system.
func bundleFor(language, system string) Manifest {
return Manifest{Module: "a-component", Build: &Build{Artifacts: []Artifact{
{Name: "binary", Kind: ArtifactBundle, Language: language, System: system},
}}}
}
func problemsOf(t *testing.T, m Manifest) string {
t.Helper()
return strings.Join(m.Build.problems(m.Module), "\n")
}
// **A language that compiles to a binary must say which system.**
//
// A binary is pinned to one operating system at link time, so a host refuses to touch a machine it
// was not built for. An artifact that says nothing would be compiled for whatever the build machine
// happened to be — which reads as portable and is not, and is the fault this check exists for.
func TestABinaryMustSayWhichSystemItIsFor(t *testing.T) {
got := problemsOf(t, bundleFor("go", ""))
if !strings.Contains(got, "says no system") {
t.Fatalf("a compiled bundle with no system was accepted:\n%s", got)
}
// And the refusal names what it could have said, so a reader is one edit from right.
for _, system := range []string{"alpine", "android", "arch"} {
if !strings.Contains(got, system) {
t.Fatalf("the refusal does not name %q as a choice:\n%s", system, got)
}
}
}
func TestABinaryThatNamesASystemIsAccepted(t *testing.T) {
if got := problemsOf(t, bundleFor("go", "arch")); got != "" {
t.Fatalf("a compiled bundle naming a system was refused:\n%s", got)
}
}
// A system the mesh does not build for is refused where it is written. These are the host's own
// names, not a compiler's: the difference between two of them is a C library rather than a kernel,
// so a value that looks like an operating system to a toolchain is still wrong here.
func TestASystemTheMeshDoesNotBuildForIsRefused(t *testing.T) {
for _, wrong := range []string{"linux", "debian", "darwin"} {
got := problemsOf(t, bundleFor("go", wrong))
if !strings.Contains(got, "and a system is") {
t.Fatalf("%q was accepted as a system:\n%s", wrong, got)
}
}
}
// **And a language that runs anywhere must not name one.** A system that decides nothing reads as
// though it did, which is the same fault as a restriction that restricts nothing (novox/hq ADR 0045).
func TestAPortableBundleMayNotNameASystem(t *testing.T) {
got := problemsOf(t, bundleFor("typescript", "arch"))
if !strings.Contains(got, "runs anywhere") {
t.Fatalf("a portable bundle was allowed to name a system:\n%s", got)
}
}
func TestAPortableBundleNamingNoSystemIsAccepted(t *testing.T) {
if got := problemsOf(t, bundleFor("typescript", "")); got != "" {
t.Fatalf("an ordinary bundle was refused:\n%s", got)
}
}
// One component, one artifact per system: the shape the mesh's own binaries are declared in, and the
// reason the target is the artifact's rather than the recipe's.
func TestOneArtifactPerSystemIsAccepted(t *testing.T) {
m := Manifest{Module: "the-host", Build: &Build{Artifacts: []Artifact{
{Name: "arch", Kind: ArtifactBundle, Language: "go", System: "arch"},
{Name: "alpine", Kind: ArtifactBundle, Language: "go", System: "alpine"},
{Name: "android", Kind: ArtifactBundle, Language: "go", System: "android"},
}}}
if got := problemsOf(t, m); got != "" {
t.Fatalf("one artifact per system was refused:\n%s", got)
}
}
@@ -0,0 +1,47 @@
package catalogue
import (
"os"
"path/filepath"
"testing"
)
// TestEveryCatalogueManifestParses runs the real catalogue through the real gate.
//
// Not a fixture: the point is whether the manifests as written are accepted by the control plane that
// will read them, and a copy of one manifest proves nothing about the other seventy-one.
func TestEveryCatalogueManifestParses(t *testing.T) {
root := os.Getenv("MESH_CATALOGUE")
if root == "" {
t.Skip("set MESH_CATALOGUE to a catalogue checkout to run this")
}
found, err := filepath.Glob(filepath.Join(root, "modules", "*", "module.json"))
if err != nil || len(found) == 0 {
t.Fatalf("no manifests under %s: %v", root, err)
}
named, routed := 0, 0
for _, p := range found {
raw, err := os.ReadFile(p)
if err != nil {
t.Fatalf("%s: %v", p, err)
}
m, err := ParseManifest(raw)
if err != nil {
t.Errorf("%s: %v", filepath.Base(filepath.Dir(p)), err)
continue
}
for _, l := range m.Listens {
if l.Name != "" {
named++
}
}
for port := range RoutedPorts(m) {
_ = port
routed++
}
}
t.Logf("%d manifests, %d named endpoints, %d routed endpoints resolved", len(found), named, routed)
if named == 0 {
t.Fatal("no endpoint in the catalogue is named, so this proved nothing")
}
}
@@ -0,0 +1,71 @@
package catalogue
import (
"os"
"strings"
"testing"
)
// **A machine trusts the mesh's authority because a module put its root there** (novox/hq ADR
// 0147, issue 129). The module carries a shell script and a unit, and both are worthless unless
// the mesh fills in where the authority is — which is the one thing about it the manifest cannot
// state, because the authority's address is a fact about the mesh and not about the module.
//
// So what is checked here is the rendering, not the parsing: the script the machine will run
// names the authority it was bound to, and the unit runs that script both ways. The verification
// itself — a plain client trusting an internal name on a machine holding this, and failing on one
// that does not — is the lab's, and cannot be had here.
func TestCaTrustRendersTheAuthorityItWasBoundTo(t *testing.T) {
raw, err := os.ReadFile("../../../mesh-catalog/modules/ca-trust/module.json")
if err != nil {
t.Skipf("the catalogue is not beside this checkout: %v", err)
}
m, err := ParseManifest(raw)
if err != nil {
t.Fatalf("the trust module does not parse:\n%v", err)
}
r := Resolution{
Node: "workstation",
Modules: []Manifest{m},
Needs: []Needed{{
Name: "internal-acme-ca", From: "anchor", At: "anchor.internal", For: "ca-trust",
Serves: map[string]any{
"port": float64(9000), "path": "/acme/acme/directory", "roots": "/roots.pem",
},
}},
}
out, err := r.Declaration(Rendering{})
if err != nil {
t.Fatalf("the trust module could not be composed for a machine: %v", err)
}
script := fileNamed(out, "ca-trust.anchor")
if script == nil {
t.Fatalf("nothing writes the script the unit runs: %v", out)
}
body, _ := script["content"].(string)
if !strings.Contains(body, "https://anchor.internal:9000/roots.pem") {
t.Errorf("the script does not fetch from the authority it was bound to:\n%s", body)
}
if script["mode"] != "0755" {
t.Errorf("the script is written %v, which systemd cannot execute", script["mode"])
}
unit := fileNamed(out, "ca-trust.unit")
if unit == nil {
t.Fatalf("no unit: %v", out)
}
text, _ := unit["content"].(string)
// Both halves. A unit that only installs the anchor leaves a machine trusting an authority
// nobody assigned it to any more, which is the half issue 129 asked for by name.
for _, want := range []string{
"ExecStart=" + script["path"].(string) + " install",
"ExecStop=" + script["path"].(string) + " remove",
"RemainAfterExit=yes",
} {
if !strings.Contains(text, want) {
t.Errorf("the unit does not say %q:\n%s", want, text)
}
}
}
+174 -4
View File
@@ -124,6 +124,16 @@ type Rendering struct {
// nothing on this node keeps them, or the mesh has no operator key.
Kept *KeptExport
// OutwardLinks is the links this machine reported as facing outside it, which the filter is
// written around (novox/hq ADR 0140). Empty means the machine has not said, and the mesh
// composes no filter for it rather than writing a rule around a link with no name.
OutwardLinks []string
// TunnelInterface is the interface the mesh's private network runs on, named here rather than
// imported because the overlay package rests on this one. Traffic arriving on it is the mesh's,
// not this machine's own guest, so the filter admits it only by a rule.
TunnelInterface string
// Foundation is the ports the mesh itself needs reachable on every machine, which no module
// declares because the foundation is not a module (novox/hq 04-ISSUES/051 and 052). The broker
// is the one that matters: a machine dials it to enrol, and a firewall derived only from
@@ -345,7 +355,21 @@ func (r Resolution) compose(with Rendering, owner map[string]string) ([]map[stri
if err != nil {
return nil, err
}
filtering := AsNftables(rules, with.Mesh, r.PublicDomain != "", with.Foundation)
// **A machine that has not said which links face outside is sent no filter** (novox/hq ADR
// 0140). The whole chain is written around those links: with none, the rule that lets this
// machine's own guests keep working would name an empty set, which nftables refuses, and a rule
// set that does not load is a machine filtering nothing while its unit reports success. Refused
// here, where a person reads it, rather than on the machine — and the machine keeps the filter
// it already has.
if filters := r.filtersHere(); filters != "" && len(with.OutwardLinks) == 0 {
return nil, fmt.Errorf(
"%s cannot be sent a filter: it has not reported which of its links face outside, and "+
"every rule in the chain is written around them. It reports that on each apply; "+
"`node show %s` says whether it has. Until then %s is not sent, and the machine "+
"keeps the filter it has", r.Node, r.Node, filters)
}
filtering := AsNftables(rules, with.Mesh, r.PublicDomain != "", with.Foundation,
with.OutwardLinks, with.TunnelInterface)
var out []map[string]any
for _, m := range r.Modules {
@@ -852,6 +876,17 @@ func mapping(written string) (outer, inner int, address string, ok bool) {
return outer, inner, strings.Join(parts[:len(parts)-2], ":"), true
}
// filtersHere is the module on this node that loads the machine's packet filter, or empty when none
// does. Named rather than counted: a refusal that says which module is one step from acted on.
func (r Resolution) filtersHere() string {
for _, m := range r.Modules {
if m.Filtering != nil {
return m.Module
}
}
return ""
}
// Rules is the rule set this node's filter is derived from: every module's listens, what was
// computed for this machine, and each module's per-node exposure. The same answer whether the node
// is adopted or converged — the one loads it as a filter, the other declares it as openings.
@@ -862,6 +897,35 @@ func (r Resolution) Rules(with Rendering) ([]Rule, error) {
if err != nil {
return nil, err
}
// And how far each endpoint reaches, which says the same thing to the filter and more
// besides (novox/hq ADR 0138). Folded in here rather than beside: the filter has one
// question — from where — and a reach answers it, so giving it two inputs would let them
// disagree. Reaches refuses a port that both name, so this cannot silently prefer one.
reaches, err := Reaches(m, with.Settings[m.Module])
if err != nil {
return nil, err
}
// **Only for an endpoint the proxy does not serve.** A routed endpoint's port is how the
// proxy reaches it and nothing else (ADR 0045), so `public` there asks for a public name and
// says nothing about the port — opening it to the world as well would undo the arrangement
// the proxy exists for, and would silently reopen a port an operator had narrowed.
//
// Found by trying to express a real module: one whose routed name must be public and whose
// machine-side port must not be. Under one value for both, there was no way to say it.
routed := RoutedPorts(m)
for port, reach := range reaches {
if routed[port] {
continue
}
source, ok := FilterSource(reach)
if !ok {
return nil, fmt.Errorf("%s: %q is not a reach the filter can read", m.Module, reach)
}
if e == nil {
e = map[int]string{}
}
e[port] = source
}
if e != nil {
exposure[m.Module] = e
}
@@ -1030,7 +1094,16 @@ func (r Resolution) contributions(settings SettingsBy, grants []Grant,
if err != nil {
return nil, fmt.Errorf("%s contributing to %s: %w", m.Module, to, err)
}
composeName(values, r.PublicDomain, r.At)
reaches, err := Reaches(m, settings[m.Module])
if err != nil {
return nil, fmt.Errorf("%s contributing to %s: %w", m.Module, to, err)
}
blocks, err := Endpoints(m, settings[m.Module])
if err != nil {
return nil, fmt.Errorf("%s contributing to %s: %w", m.Module, to, err)
}
portOfEndpoint(values, endpointPorts(m))
composeName(values, r.PublicDomain, r.At, reaches, endpointPorts(m), blocks)
out[to] = append(out[to], Contribution{From: m.Module, Values: values})
}
// Several contributions to one requirement (ADR 0094's sibling for `contributes`): an
@@ -1044,7 +1117,16 @@ func (r Resolution) contributions(settings SettingsBy, grants []Grant,
if err != nil {
return nil, fmt.Errorf("%s contributing %s to %s: %w", m.Module, local, to, err)
}
composeName(values, r.PublicDomain, r.At)
reaches, err := Reaches(m, settings[m.Module])
if err != nil {
return nil, fmt.Errorf("%s contributing %s to %s: %w", m.Module, local, to, err)
}
blocks, err := Endpoints(m, settings[m.Module])
if err != nil {
return nil, fmt.Errorf("%s contributing %s to %s: %w", m.Module, local, to, err)
}
portOfEndpoint(values, endpointPorts(m))
composeName(values, r.PublicDomain, r.At, reaches, endpointPorts(m), blocks)
out[to] = append(out[to], Contribution{From: m.Module, Values: values})
}
}
@@ -1075,10 +1157,44 @@ func (r Resolution) contributions(settings SettingsBy, grants []Grant,
// the running mesh keeps serving the full names it has. And a labelled contribution on a node with
// no public domain composes nothing — there is nothing to join it to — which reads downstream as a
// route that named no host, the same as it would have before this existed.
func composeName(values map[string]any, publicDomain, internalDomain string) {
func composeName(values map[string]any, publicDomain, internalDomain string, reaches map[int]string,
ports map[string]int, blocks map[string]Endpoint) {
if values == nil {
return
}
// **The subdomain an assignment gave this endpoint**, before the name is joined (novox/hq ADR
// 0138). The module contributes a label because it names its own parts; an assignment may say a
// different one, because where a thing lives under a domain is the operator's to choose and used
// to require editing the module to change.
if name, ok := values[RouteEndpoint].(string); ok {
if ep, said := blocks[strings.TrimSpace(name)]; said && ep.Label != "" {
values["label"] = ep.Label
}
}
// **How far the endpoint this route serves reaches decides which names exist** (novox/hq ADR
// 0138). Both were composed whenever the node had both domains, so every routed module got a
// public name and an internal one whether anybody wanted them or not — and a certificate for
// each, because the proxy certifies the names it is given.
//
// Joined by the port: a route entry names the port it serves and the module declares a listen on
// it. An entry with no port is not an endpoint's route but a rule about a name — a path-level
// refusal shadowing another route — and it inherits whatever that route's names turned out to
// be, which is why it is left alone here.
//
// Nothing said is both names, as before. That is what keeps every mesh already running identical
// until an assignment speaks.
wantPublic, wantInternal := true, true
if port, ok := endpointPortOf(values, ports); ok {
if reach, said := reaches[port]; said {
wantPublic, wantInternal = WantsPublicName(reach), WantsInternalName(reach)
}
}
if !wantPublic {
publicDomain = ""
}
if !wantInternal {
internalDomain = ""
}
if _, already := values["name"]; already {
// A full name was given rather than a label. Left as-is: this is the legacy shape, and the
// point of the label is to not have to write the full name — a contribution that wrote both
@@ -1698,3 +1814,57 @@ func prepared(from map[string]any) map[string]any {
delete(step, "reload-on")
return step
}
// endpointPortOf is the port the endpoint a route serves listens on: looked up by the name the route
// gives, or read from the port it repeats (novox/hq ADR 0138).
//
// `ports` maps this module's endpoint names to their ports, computed once per module rather than
// re-scanned per contribution.
func endpointPortOf(values map[string]any, ports map[string]int) (int, bool) {
if name, ok := values[RouteEndpoint].(string); ok {
if port, found := ports[strings.TrimSpace(name)]; found {
return port, true
}
}
return asPort(values["port"])
}
// endpointPorts is a module's endpoint names against the ports they listen on.
func endpointPorts(m Manifest) map[string]int {
out := map[string]int{}
for _, l := range m.Listens {
if name := strings.TrimSpace(l.Name); name != "" {
out[name] = l.Port
}
}
return out
}
// portOfEndpoint fills in the port of the endpoint a contribution names, in place.
//
// **A contribution that names an endpoint must still carry that endpoint's port**, because everything
// downstream reads the port: the provider is told where to reach the consumer, and the machine-side
// redirection that turns a declared port into the number the machine published is keyed on it
// (atMachinePort). A route that named only its endpoint left the proxy with no port at all, and a
// proxy with no port has nothing to dial.
//
// Found before it shipped and after the catalogue had already been changed to name endpoints — the
// manifests were merged and the mesh had not yet picked them up, so nothing was broken yet. The
// declared port, not the machine one: the redirection happens later and is keyed on the declared
// number, so filling in the machine port here would be redirected a second time or not at all.
func portOfEndpoint(values map[string]any, ports map[string]int) {
if values == nil {
return
}
if _, already := values["port"]; already {
// A route that says both is its own answer; the older shape repeated the port and is still read.
return
}
name, ok := values[RouteEndpoint].(string)
if !ok {
return
}
if port, found := ports[strings.TrimSpace(name)]; found {
values["port"] = port
}
}
+205
View File
@@ -0,0 +1,205 @@
package catalogue
import (
"strings"
"testing"
)
// aMediaServer is the shape one port number per key cannot express: two endpoints of different kinds.
// A web surface a proxy serves under a subdomain, and a protocol port clients dial directly because
// the client expects that number.
func aMediaServer() Manifest {
return Manifest{
Module: "media",
Listens: []Listening{
{Name: "web", Port: 80, From: FromMesh, Why: "the app, behind the proxy"},
{Name: "stream", Port: 32400, From: FromEverywhere, Fixed: true,
Why: "the client dials this number; the protocol chose it"},
},
Contributes: map[string]map[string]any{
"route": {"label": "media", RouteEndpoint: "web"},
},
// Both endpoints are published by its container, which is what lets a machine port be given
// for either: the mesh moves a port the module publishes, never one it merely listens on.
Resources: []map[string]any{
{"id": "server", "type": "container", "name": "media",
"ports": []any{"80", "32400"}},
},
}
}
// **A route names the endpoint it serves.** A route and a listen both carried a port and nothing said
// they were the same thing; now one of them says so.
func TestARouteNamesTheEndpointItServes(t *testing.T) {
m := aMediaServer()
if port, ok := EndpointPort(m, "web"); !ok || port != 80 {
t.Fatalf("the web endpoint resolves to %d (%v), want 80", port, ok)
}
if port, ok := EndpointPort(m, "stream"); !ok || port != 32400 {
t.Fatalf("the stream endpoint resolves to %d (%v), want 32400", port, ok)
}
if _, ok := EndpointPort(m, "absent"); ok {
t.Fatal("an endpoint the module does not declare resolved to a port")
}
}
// And the routed set is read through the name, so the endpoint the proxy serves is known without a
// reader joining two numbers.
func TestTheRoutedEndpointIsFoundByName(t *testing.T) {
routed := RoutedPorts(aMediaServer())
if !routed[80] {
t.Fatalf("the routed endpoint was not found by name: %v", routed)
}
// And the directly-dialled one is not routed, which is what lets its reach govern its port.
if routed[32400] {
t.Fatalf("the endpoint clients dial directly reads as routed: %v", routed)
}
}
// **Two endpoints of different shapes, configured as themselves.** The web endpoint's reach asks for
// names and leaves its port to the proxy; the stream endpoint's reach governs its port, because
// clients dial it and there is no name.
func TestTwoEndpointsOfDifferentShapesAreConfiguredSeparately(t *testing.T) {
m := aMediaServer()
settings := SettingsBy{"media": {{From: "node anchor", Values: map[string]any{
ReachSetting: map[string]any{"80": ReachBoth, "32400": ReachPublic},
}}}}
r := Resolution{Node: "anchor", Modules: []Manifest{m},
PublicDomain: "example.test", At: "anchor.internal"}
rules, err := r.Rules(Rendering{Settings: settings})
if err != nil {
t.Fatal(err)
}
for _, rule := range rules {
switch rule.Port {
case 80:
if rule.From != FromMesh {
t.Fatalf("the routed endpoint's port opened to %q; the proxy is how it is reached",
rule.From)
}
case 32400:
if rule.From != FromEverywhere {
t.Fatalf("the directly-dialled endpoint's port is %q, want anywhere", rule.From)
}
}
}
// And the routed one carries both names, asked for by the same statement.
given, err := r.contributions(settings, nil, nil)
if err != nil {
t.Fatal(err)
}
var public, internal string
for _, c := range given["route"] {
public, _ = c.Values["name"].(string)
internal, _ = c.Values["internal-name"].(string)
}
if public != "media.example.test" || internal != "media.anchor.internal" {
t.Fatalf("names are %q and %q, want both", public, internal)
}
}
// A route naming an endpoint the module does not declare reaches nothing, and is refused where it is
// written rather than resolving to no port and serving nothing.
func TestARouteNamingAnEndpointTheModuleLacksIsRefused(t *testing.T) {
m := aMediaServer()
m.Contributes["route"][RouteEndpoint] = "absent"
got := strings.Join(RouteProblems(m), "\n")
if !strings.Contains(got, "does not declare") {
t.Fatalf("a route naming an absent endpoint was accepted:\n%s", got)
}
}
// **Two endpoints called the same would make an assignment configure whichever was read last.** The
// point of a name is that it identifies one thing.
func TestTwoEndpointsWithOneNameAreRefused(t *testing.T) {
m := Manifest{Module: "twice", Listens: []Listening{
{Name: "web", Port: 80, From: FromMesh},
{Name: "web", Port: 8080, From: FromMesh},
}}
got := strings.Join(endpointNameProblems(m), "\n")
if !strings.Contains(got, "could mean either") {
t.Fatalf("two endpoints with one name were accepted:\n%s", got)
}
}
// A name that is not a name is refused where it is written: it ends up in something a person types.
func TestAnEndpointNameIsHeldToItsShape(t *testing.T) {
for _, wrong := range []string{"Web", "web port", "3000", "-web", "web_surface"} {
m := Manifest{Module: "odd", Listens: []Listening{{Name: wrong, Port: 80, From: FromMesh}}}
if got := strings.Join(endpointNameProblems(m), "\n"); !strings.Contains(got, "a name is lowercase") {
t.Fatalf("%q was accepted as an endpoint name:\n%s", wrong, got)
}
}
}
// **Every endpoint in the catalogue is unnamed today, and must stay valid.** The word ships one
// release before anything uses it.
func TestAnUnnamedEndpointIsStillValid(t *testing.T) {
m := Manifest{Module: "ordinary", Listens: []Listening{{Port: 443, From: FromEverywhere}}}
if got := endpointNameProblems(m); len(got) != 0 {
t.Fatalf("an unnamed endpoint was refused: %v", got)
}
if got := RouteProblems(m); len(got) != 0 {
t.Fatalf("a module with no route was refused: %v", got)
}
}
// **A route that names an endpoint still carries that endpoint's port.**
//
// Everything downstream reads the port: the provider is told where to reach the consumer, and the
// redirection that turns a declared port into the number the machine published is keyed on it. A route
// naming only its endpoint left the proxy with no port, and a proxy with no port has nothing to dial.
//
// Caught after the catalogue had already been changed to name endpoints, and before the mesh picked
// those manifests up — which is the only reason nothing broke.
func TestARouteNamingAnEndpointStillCarriesItsPort(t *testing.T) {
m := aMediaServer()
r := Resolution{Node: "anchor", Modules: []Manifest{m},
PublicDomain: "example.test", At: "anchor.internal"}
given, err := r.contributions(nil, nil, nil)
if err != nil {
t.Fatal(err)
}
var saw bool
for _, c := range given["route"] {
saw = true
port, ok := asPort(c.Values["port"])
if !ok {
t.Fatalf("the route carries no port, so the proxy has nothing to dial: %v", c.Values)
}
if port != 80 {
t.Fatalf("the route carries port %d, want the web endpoint's 80", port)
}
}
if !saw {
t.Fatal("the module contributed no route")
}
}
// And the declared port, not the machine one: the redirection to where the machine published it
// happens later and is keyed on the declared number, so filling the machine port in here would be
// redirected twice or not at all.
func TestTheEndpointsDeclaredPortIsFilledInNotTheMachineOne(t *testing.T) {
m := aMediaServer()
values := map[string]any{RouteEndpoint: "web", "label": "media"}
portOfEndpoint(values, endpointPorts(m))
if got, _ := asPort(values["port"]); got != 80 {
t.Fatalf("filled in port %d, want the declared 80", got)
}
// Then the ordinary redirection puts it where the machine published it.
moved := atMachinePort(values, m.Module, map[string]map[int]int{"media": {80: 20009}})
if got, _ := asPort(moved["port"]); got != 20009 {
t.Fatalf("after redirection the port is %d, want the machine's 20009", got)
}
}
// A route that repeats a port keeps it, because that is the older shape and still read.
func TestARouteThatRepeatsItsPortKeepsIt(t *testing.T) {
values := map[string]any{RouteEndpoint: "web", "port": 8080}
portOfEndpoint(values, map[string]int{"web": 80})
if got, _ := asPort(values["port"]); got != 8080 {
t.Fatalf("the port it stated was overwritten with %d", got)
}
}
@@ -0,0 +1,140 @@
package catalogue
import (
"strings"
"testing"
)
func configured(block map[string]any) SettingsBy {
return SettingsBy{"media": {{From: "node anchor",
Values: map[string]any{EndpointsSetting: block}}}}
}
// **One block per endpoint, saying all three things.** The machine port, the subdomain and the reach
// were `ports`, the route's label and `reach`, each keyed by a port number, so configuring a module
// with two endpoints of different shapes meant knowing which number was which.
func TestAnEndpointsBlockSaysPortLabelAndReach(t *testing.T) {
m := aMediaServer()
// stream's reach NARROWS what the manifest says — the manifest has it from anywhere, the
// assignment says internal. Chosen deliberately: a reach that agrees with the manifest proves
// nothing about whether the block was read at all.
settings := configured(map[string]any{
"web": map[string]any{"port": 20009, "label": "cinema", "reach": ReachBoth},
"stream": map[string]any{"reach": ReachInternal},
})
// The machine port, where the mapping is read.
given, err := GivenPorts(m, settings["media"])
if err != nil {
t.Fatal(err)
}
if given[80] != 20009 {
t.Fatalf("the web endpoint is on machine port %d, want 20009: %v", given[80], given)
}
// The reach, where the filter reads it.
r := Resolution{Node: "anchor", Modules: []Manifest{m},
PublicDomain: "example.test", At: "anchor.internal"}
rules, err := r.Rules(Rendering{Settings: settings})
if err != nil {
t.Fatal(err)
}
for _, rule := range rules {
if rule.Port == 32400 && rule.From != FromMesh {
t.Fatalf("the directly-dialled endpoint is %q; the assignment narrowed it to the private "+
"network and the manifest's 'anywhere' should not win", rule.From)
}
if rule.Port == 80 && rule.From != FromMesh {
t.Fatalf("the routed endpoint's port opened to %q; the proxy is how it is reached", rule.From)
}
}
// And the subdomain, where the name is composed — the assignment's, not the module's.
nodes, err := r.contributions(settings, nil, nil)
if err != nil {
t.Fatal(err)
}
for _, c := range nodes["route"] {
if got, _ := c.Values["name"].(string); got != "cinema.example.test" {
t.Fatalf("the public name is %q, want the label the assignment gave", got)
}
if got, _ := c.Values["internal-name"].(string); got != "cinema.anchor.internal" {
t.Fatalf("the internal name is %q, want the label the assignment gave", got)
}
}
}
// A block that says only a reach leaves the port to the mesh and the label to the module, which is the
// ordinary case and must not require writing the other two.
func TestABlockMaySayOnlyAReach(t *testing.T) {
m := aMediaServer()
settings := configured(map[string]any{"web": map[string]any{"reach": ReachInternal}})
r := Resolution{Node: "anchor", Modules: []Manifest{m},
PublicDomain: "example.test", At: "anchor.internal"}
nodes, err := r.contributions(settings, nil, nil)
if err != nil {
t.Fatal(err)
}
for _, c := range nodes["route"] {
if got, _ := c.Values["name"].(string); got != "" {
t.Fatalf("an internal endpoint composed the public name %q", got)
}
// The module's own label, untouched.
if got, _ := c.Values["internal-name"].(string); got != "media.anchor.internal" {
t.Fatalf("the internal name is %q, want the module's own label", got)
}
}
}
// An endpoint the module does not declare reaches nothing, and the refusal says what it does declare.
func TestConfiguringAnEndpointTheModuleLacksIsRefused(t *testing.T) {
_, err := Endpoints(aMediaServer(), configured(map[string]any{
"admin": map[string]any{"reach": ReachInternal}})["media"])
if err == nil || !strings.Contains(err.Error(), "does not declare") {
t.Fatalf("configuring an absent endpoint was accepted: %v", err)
}
if err != nil && !strings.Contains(err.Error(), "stream") {
t.Fatalf("the refusal does not name what the module declares: %v", err)
}
}
func TestAReachInABlockIsHeldToTheFourValues(t *testing.T) {
_, err := Endpoints(aMediaServer(), configured(map[string]any{
"web": map[string]any{"reach": "mesh"}})["media"])
if err == nil || !strings.Contains(err.Error(), "a reach is") {
t.Fatalf("a filter word was accepted as a reach: %v", err)
}
}
// **Two places giving one endpoint a machine port is the confusion this key exists to end.**
func TestAnEndpointGivenAPortTwiceIsRefused(t *testing.T) {
m := aMediaServer()
_, err := GivenPorts(m, []Layer{{From: "node anchor", Values: map[string]any{
EndpointsSetting: map[string]any{"web": map[string]any{"port": 20009}},
PortsSetting: map[string]any{"80": 30000},
}}})
if err == nil || !strings.Contains(err.Error(), "published once") {
t.Fatalf("an endpoint given two machine ports was accepted: %v", err)
}
}
// And the same for its reach, said once here and once through the older key.
func TestAnEndpointWhoseReachIsAlsoExposedIsRefused(t *testing.T) {
_, err := Endpoints(aMediaServer(), []Layer{{From: "node anchor", Values: map[string]any{
EndpointsSetting: map[string]any{"web": map[string]any{"reach": ReachInternal}},
ExposeSetting: map[string]any{"80": FromEverywhere},
}}})
if err == nil || !strings.Contains(err.Error(), "same thing in different words") {
t.Fatalf("a reach said two ways was accepted: %v", err)
}
}
// A module whose endpoints are unnamed cannot be configured this way, and is told so rather than
// having a block silently reach nothing — which is every module in the catalogue today.
func TestAModuleWithNoNamedEndpointsIsToldSo(t *testing.T) {
m := Manifest{Module: "media", Listens: []Listening{{Port: 80, From: FromMesh}}}
_, err := Endpoints(m, configured(map[string]any{"web": map[string]any{"reach": ReachBoth}})["media"])
if err == nil || !strings.Contains(err.Error(), "no endpoints by name") {
t.Fatalf("a module with no named endpoints accepted a block: %v", err)
}
}
+433 -21
View File
@@ -230,7 +230,23 @@ const SSHPort = 22
// It is a floor for the same reason ssh is. A machine nobody can reach is a machine nobody can
// repair; a machine the mesh cannot reach is a machine the mesh cannot manage. Neither is a thing
// any module asks for, and neither may be derived away.
func AsNftables(rules []Rule, mesh []string, outward bool, foundation []int) string {
func AsNftables(rules []Rule, mesh []string, outward bool, foundation []int,
outwardLinks []string, tunnel string) string {
// The links that are not this machine's own: the ones facing outside, and the mesh's tunnel.
// Traffic arriving on any of them is admitted only by a rule below; traffic arriving anywhere
// else is this machine's own guest and is not something the mesh has a position on.
//
// The tunnel is named here deliberately. Treating it as "not outside" would make a port nothing
// declares reachable from every machine in the mesh, which is the derivation abandoned.
quoted := make([]string, 0, len(outwardLinks)+1)
for _, link := range outwardLinks {
quoted = append(quoted, fmt.Sprintf("%q", link))
}
if tunnel != "" {
quoted = append(quoted, fmt.Sprintf("%q", tunnel))
}
inward := strings.Join(quoted, ", ")
var b strings.Builder
b.WriteString("# Computed by the mesh from what is assigned to this node.\n")
b.WriteString("# Edits are lost on the next declaration; change a module's listens instead.\n\n")
@@ -249,9 +265,45 @@ func AsNftables(rules []Rule, mesh []string, outward bool, foundation []int) str
b.WriteString("\t\tct state established,related accept\n")
b.WriteString("\t\tct state invalid drop\n")
b.WriteString("\t\tiif lo accept\n")
// **Anything on this machine may call anything on this machine.**
//
// Local is not a boundary this mesh draws. A service running here is callable by everything else
// running here, whatever form either takes — a package with a unit, a binary, a container. Whether
// a caller sits in a container was never meant to change the answer, and the only reason it did was
// that this chain asked about addresses: a caller on the machine carries the machine's address, a
// caller in one of its containers carries a bridge address, and a rule naming the former silently
// refused the latter.
//
// Measured: a module reaching its database on this machine's own name timed out for eleven hours
// while the machine itself could reach it, and the mesh called the machine healthy throughout
// (novox/hq 04-ISSUES/145).
//
// Asked by the link it arrives on rather than the address it comes from: anything that did not
// arrive from outside this machine, and did not arrive over the private network, is this machine's
// own. One rule for every service here, in place of a line per port that only ever covered the
// ports somebody remembered to think about.
if inward != "" {
b.WriteString(fmt.Sprintf("\t\tiifname != { %s } accept\n", inward))
}
b.WriteString("\t\ticmp type echo-request accept\n")
b.WriteString("\t\ticmpv6 type { echo-request, nd-neighbor-solicit, nd-neighbor-advert, nd-router-advert } accept\n")
// **What this machine's own guests must be able to ask it** (novox/hq ADR 0140). A guest gets
// its address and its names from this machine, over the link it is on, and those two questions
// arrive at the input chain like any other. Denied, the guest never gets an address and never
// resolves a name — which is not "a closed port" but a network that does not work at all, and it
// is this machine's own guest asking.
//
// Asked for by the link it arrives on rather than by the address it comes from, for the reason
// the forward chain below no longer names an address: a range describes one machine and goes
// stale in silence. Anything arriving from outside, or over the tunnel, is not a guest of this
// machine and asks through a port somebody declared, like everything else.
if len(inward) > 0 {
b.WriteString("\t\t# this machine's own guests asking it for an address and for names\n")
b.WriteString(fmt.Sprintf("\t\tiifname != { %s } udp dport { 53, 67 } accept\n", inward))
b.WriteString(fmt.Sprintf("\t\tiifname != { %s } tcp dport 53 accept\n", inward))
}
// **ssh, always, and not because a module asked.**
//
// Every other line in this chain is derived from what is assigned here, which is the whole
@@ -361,19 +413,36 @@ func AsNftables(rules []Rule, mesh []string, outward bool, foundation []int) str
// about the ports most worth protecting. Rehearsed on three machines: loading these rules
// refused a port on the host and left a published container port reachable (novox/hq issue 047).
//
// The way through is the one the system being replaced already used: deny by default here, and
// then explicitly allow the runtime's own networks, so containers keep working while everything
// else has to be asked for.
// **What it constrains is traffic arriving from OUTSIDE this machine, and nothing else**
// (novox/hq ADR 0140).
//
// It used to deny everything here and then allow the machine's own containers back by naming
// the address ranges they sit on — two ranges fixed in this file and the rest recorded per
// machine. Every way of keeping that list correct failed. A constant describes one machine. A
// recorded range goes stale in silence and cannot tell a network the mesh made from one a
// predecessor left behind. Generating it from the modules would have put half this rule set on
// the machine.
//
// The list should not exist, because the mesh has no position on a container reaching outward:
// that is not a port opened to anybody. So traffic that did not arrive from outside is accepted
// in one line, and what did arrive from outside is allowed only where a rule below admits it.
//
// The tunnel is not "not outside". Accepting everything off it would make a port nothing
// declares reachable from any machine in the mesh, which is the derivation abandoned — so it is
// named here beside the outward links, and traffic arriving on it meets the rules below like
// anything else.
b.WriteString("\tchain forward {\n")
b.WriteString("\t\ttype filter hook forward priority filter; policy drop;\n")
b.WriteString("\t\tct state established,related accept\n")
b.WriteString("\t\tct state invalid drop\n")
b.WriteString("\n")
// What the container runtime created. Without these, denying by default stops every container
// on the machine — which is exactly the failure the absent chain was avoiding, avoided properly.
for _, network := range runtimeNetworks {
b.WriteString(fmt.Sprintf("\t\t# %s\n", network.why))
b.WriteString(fmt.Sprintf("\t\tip saddr %s accept\n", network.cidr))
// Only when there is a link to name. An empty set is a line nftables refuses, and a rule set
// that does not load is a machine filtering nothing while its unit reports success — so the
// chain denies rather than renders nonsense. Composing a declaration for a machine that has
// named none is refused upstream, so this is a floor and not a path anything travels.
if inward != "" {
b.WriteString("\t\t# this machine's own guests reaching outward: not a port opened to anybody\n")
b.WriteString(fmt.Sprintf("\t\tiifname != { %s } accept\n", inward))
}
if len(rules) > 0 {
@@ -430,18 +499,6 @@ func AsNftables(rules []Rule, mesh []string, outward bool, foundation []int) str
return b.String()
}
// runtimeNetworks are the container runtime's own networks, which must keep working when the
// forward chain denies by default.
//
// Taken from what the system being replaced allows, which has been carrying this machine's traffic
// for months: the runtime's bridge range and the range its compose files are given. A machine whose
// runtime is configured with something else needs this to say so — which is a thing the mesh cannot
// derive and a reason this list is named here rather than computed.
var runtimeNetworks = []struct{ cidr, why string }{
{"172.16.0.0/12", "the container runtime's bridge networks"},
{"192.168.128.0/17", "the networks its compose files are given"},
}
// byFamily splits addresses into the two nftables understands separately.
//
// `ip saddr` and `ip6 saddr` are different matches, and one set holding both families is a syntax
@@ -499,6 +556,21 @@ const MeshWideLayer = "the mesh"
// two mappings share a number, it is an entry one of them writes over the other's, and the reader
// that finds the survivor disagrees with the reader that recomputes it.
func GivenPorts(m Manifest, layers []Layer) (map[int]int, error) {
// An endpoint's own block may put it on a machine port, which is the same thing `ports` says about
// the number rather than about the endpoint (novox/hq ADR 0138). Collected first and then let the
// older key be read, which refuses a port said twice.
byName, err := Endpoints(m, layers)
if err != nil {
return nil, err
}
named := map[int]int{}
for name, ep := range byName {
if ep.Port == 0 {
continue
}
named[endpointPorts(m)[name]] = ep.Port
}
// Every name a setting may use, and the mapping it names.
names := map[int][]publishing{}
for _, p := range publishedPorts(m) {
@@ -597,6 +669,17 @@ func GivenPorts(m Manifest, layers []Layer) (map[int]int, error) {
out[key], by[key] = at, port
}
}
// And what the endpoints' own blocks put them on. Refused rather than merged where both keys name
// one endpoint: two places giving a port is the confusion this key exists to end.
for wanted, at := range named {
if was, twice := out[wanted]; twice && was != at {
return nil, fmt.Errorf(
"%s puts its port %d on %d through %s and on %d through %s — one endpoint, two "+
"machine ports, and it is published once. Keep the endpoint's own block",
m.Module, wanted, at, EndpointsSetting, was, PortsSetting)
}
out[wanted] = at
}
if len(out) == 0 {
return nil, nil
}
@@ -664,3 +747,332 @@ func sortedPorts(of map[int]int) []int {
sort.Ints(out)
return out
}
// ReachSetting is the settings key that says how far one of a module's endpoints reaches, per node
// (novox/hq ADR 0138):
//
// {"reach": {"3000": "internal"}}
//
// **One value, three readers.** Reachability used to be settled three times over: the filter read a
// listen's source, which `expose` could override; the proxy composed a public name and an internal
// name for every route it was given, because it could; and the certificate authority followed from
// which names existed. Each was defensible and the combination was unstated, so "this endpoint must
// not be public" could not be written and was therefore enforced by nothing — while a public
// certificate for that very name was obtained anyway.
//
// It keys on the port the module declares, the same key `ports` and `expose` use. A route names that
// port too, which is what lets one statement reach the names as well as the filter: of the 36 route
// entries in the catalogue, 35 name a port that the same module declares a listen on, and the one
// that does not is a path-level refusal — a rule about a name rather than an endpoint.
const ReachSetting = "reach"
// How far an endpoint reaches. Four values, because they have to cover everything `expose` could say
// as well as the two names.
const (
// ReachMachine is this machine only: not the private network, not the world, and no name.
ReachMachine = "machine"
// ReachInternal is the private network, under the internal name and not the public one.
ReachInternal = "internal"
// ReachPublic is the world, under the public name and not the internal one.
ReachPublic = "public"
// ReachBoth is the world, under both names — each certified by its own authority.
//
// The filter cannot distinguish this from ReachPublic, and should not try: the mesh's addresses
// are a subset of anywhere. What differs is the names, which is the whole reason reach is not
// simply the filter's vocabulary with nicer words.
ReachBoth = "both"
)
// reaches is every value, in the order a refusal lists them.
var reaches = []string{ReachMachine, ReachInternal, ReachPublic, ReachBoth}
// RoutedPorts are the ports a module serves through a proxy, taken from its route contributions.
//
// **A routed endpoint's port is how the proxy reaches it, and nothing else.** That is ADR 0045's
// decision and it is older than reach: a public service listens `from: mesh`, only the proxy reaches
// it, and it is exposed by name. So `public` on a routed endpoint asks for a public *name*; opening
// that port to the world as well would undo the arrangement the proxy exists for.
//
// Measured before this was written, not reasoned: a module's routed name answered from the internet
// over TLS while its machine-side port was refused from the same place. The port is not the path.
func RoutedPorts(m Manifest) map[int]bool {
out := map[int]bool{}
note := func(values map[string]any) {
// **The endpoint it serves, by name where it says one.** A route repeating a port number is
// the older shape and still read: 35 of the catalogue's 36 route entries name a port their
// module declares a listen on (novox/hq ADR 0138).
if name, ok := values[RouteEndpoint].(string); ok {
if port, found := EndpointPort(m, name); found {
out[port] = true
return
}
}
if port, ok := asPort(values["port"]); ok {
out[port] = true
}
}
if values, ok := m.Contributes["route"]; ok {
note(values)
}
for _, values := range m.ContributesMany["route"] {
note(values)
}
return out
}
// FilterSource is the source a reach means to the packet filter.
//
// `public` and `both` are the same here. A reach that opened a port to the mesh and not to the world
// would be `internal`; there is no reach that opens it to the world and *not* to the mesh, because a
// filter cannot express "everyone except these" and nobody has asked for it.
func FilterSource(reach string) (string, bool) {
switch reach {
case ReachMachine:
return FromMachine, true
case ReachInternal:
return FromMesh, true
case ReachPublic, ReachBoth:
return FromEverywhere, true
default:
return "", false
}
}
// WantsPublicName is whether a reach asks for the route's public name to be composed.
func WantsPublicName(reach string) bool { return reach == ReachPublic || reach == ReachBoth }
// WantsInternalName is whether a reach asks for the route's internal name to be composed.
func WantsInternalName(reach string) bool { return reach == ReachInternal || reach == ReachBoth }
// Reaches reads a module's per-node reach settings: declared port → how far it reaches.
//
// It refuses a reach for a port the module does not listen on, or a value that is not one of the
// four — the "reads as a restriction and is none" fault this whole mechanism exists to prevent
// (novox/hq ADR 0043/0045). It also refuses a port that `expose` names as well: the two say the same
// thing in different words, and a module whose reach and exposure disagree would have the filter
// following one and the names following the other, which is the very confusion ADR 0138 removes.
//
// A module with no `reach` setting yields nothing, and everything behaves exactly as before: the
// filter follows the manifest's `from`, and both names are composed. That is what keeps every machine
// already running unchanged until an assignment says otherwise.
func Reaches(m Manifest, layers []Layer) (map[int]string, error) {
listened := make(map[int]bool, len(m.Listens))
for _, l := range m.Listens {
listened[l.Port] = true
}
exposed, err := Exposure(m, layers)
if err != nil {
return nil, err
}
out := map[int]string{}
// What an endpoint's own block says, which is the same statement in the shape that names the
// endpoint rather than its port (novox/hq ADR 0138). Read first so the older key, which says less,
// cannot quietly win over the newer one that says more.
blocks, err := Endpoints(m, layers)
if err != nil {
return nil, err
}
declared := endpointPorts(m)
for name, ep := range blocks {
if ep.Reach == "" {
continue
}
out[declared[name]] = ep.Reach
}
for _, layer := range layers {
raw, ok := layer.Values[ReachSetting]
if !ok {
continue
}
entries, ok := raw.(map[string]any)
if !ok {
return nil, fmt.Errorf("%s: %s is a { port: reach } map, and %q set it to something else",
m.Module, ReachSetting, layer.From)
}
for portText, value := range entries {
port, err := strconv.Atoi(portText)
if err != nil {
return nil, fmt.Errorf("%s says how far %q reaches, which is not a port", m.Module, portText)
}
if !listened[port] {
return nil, fmt.Errorf(
"%s says how far port %d reaches, which it does not listen on — the setting "+
"reaches nothing", m.Module, port)
}
reach, ok := value.(string)
if !ok || !slices.Contains(reaches, reach) {
return nil, fmt.Errorf("%s says port %d reaches %v; a reach is %s",
m.Module, port, value, strings.Join(reaches, ", "))
}
if _, both := exposed[port]; both {
return nil, fmt.Errorf(
"%s sets both %s and %s for port %d. They say the same thing in different "+
"words, and the filter would follow one while its names followed the other "+
"— which is what %s exists to stop. Keep %s",
m.Module, ReachSetting, ExposeSetting, port, ReachSetting, ReachSetting)
}
out[port] = reach
}
}
if len(out) == 0 {
return nil, nil
}
return out, nil
}
// RouteEndpoint is the key a route contribution names the endpoint it serves with, instead of
// repeating that endpoint's port (novox/hq ADR 0138).
//
// **A route and a listen both carried a port, and nothing said they were the same thing.** They
// always were — a route serves one of the module's own endpoints — but a reader had to join two
// numbers, and an assignment configuring "the web endpoint" had to know which number that was. A
// route that names the endpoint says what it means, and the mesh looks the port up.
const RouteEndpoint = "endpoint"
// RouteProblems holds a module's route contributions to naming an endpoint it actually has.
//
// A route naming an endpoint the module does not declare reaches nothing, and is refused where it is
// written rather than resolving to no port and serving nothing — the fault this repository names most
// often, a declaration that reads as though it did something.
func RouteProblems(m Manifest) []string {
var problems []string
check := func(where string, values map[string]any) {
name, ok := values[RouteEndpoint].(string)
if !ok || strings.TrimSpace(name) == "" {
return
}
if _, found := EndpointPort(m, name); !found {
problems = append(problems, fmt.Sprintf(
"%s routes %s to the endpoint %q, which it does not declare", m.Module, where, name))
}
}
if values, ok := m.Contributes["route"]; ok {
check("a name", values)
}
for local, values := range m.ContributesMany["route"] {
check(local, values)
}
return problems
}
// EndpointsSetting is the settings key that configures a module's endpoints by name, per node
// (novox/hq ADR 0138):
//
// {"endpoints": {"web": {"port": 20009, "label": "media", "reach": "both"},
// "stream": {"reach": "public"}}}
//
// **One block per endpoint, instead of three keys joined by a number.** Which machine port it lands
// on, the subdomain a proxy serves it under, and how far it reaches are the three things an operator
// says when a module is assigned, and they were said in `ports`, in the route's label and in `reach`,
// each keyed by the port. A module with two endpoints of different shapes — a web surface behind the
// proxy and a protocol port clients dial directly — could only be configured by a reader who knew
// which number was which.
//
// Every field is optional. A block that says only a reach leaves the port to the mesh and the label to
// the module, which is the ordinary case.
const EndpointsSetting = "endpoints"
// Endpoint is what an assignment says about one of a module's endpoints.
type Endpoint struct {
// Port is the machine-side port it is published on. Zero means the mesh assigns one, which it
// does anyway — a fixed port is the module's claim and is honoured without being said here.
Port int
// Label is the subdomain a proxy serves it under, overriding the one the module contributes.
Label string
// Reach is how far it reaches: machine, internal, public or both.
Reach string
}
// Endpoints reads a module's per-node endpoint configuration, by endpoint name.
//
// It refuses a name the module does not declare — the setting would reach nothing — and a reach that
// is not one of the four. It also refuses an endpoint whose port or reach is said twice, once here and
// once through the older key: two places saying the same thing is what this key exists to end, and
// letting both stand would mean the mesh followed whichever it read last.
func Endpoints(m Manifest, layers []Layer) (map[string]Endpoint, error) {
declared := endpointPorts(m)
exposed, err := Exposure(m, layers)
if err != nil {
return nil, err
}
out := map[string]Endpoint{}
for _, layer := range layers {
raw, ok := layer.Values[EndpointsSetting]
if !ok {
continue
}
blocks, ok := raw.(map[string]any)
if !ok {
return nil, fmt.Errorf("%s: %s is a { endpoint: { … } } map, and %q set it to something else",
m.Module, EndpointsSetting, layer.From)
}
for name, body := range blocks {
port, known := declared[name]
if !known {
return nil, fmt.Errorf(
"%s configures the endpoint %q, which it does not declare — the setting reaches "+
"nothing. It declares %s", m.Module, name, spokenEndpoints(m))
}
values, ok := body.(map[string]any)
if !ok {
return nil, fmt.Errorf("%s: the endpoint %q is configured with something that is not a "+
"block of settings", m.Module, name)
}
ep := out[name]
if reach, said := values["reach"]; said {
text, ok := reach.(string)
if !ok || !slices.Contains(reaches, text) {
return nil, fmt.Errorf("%s says the endpoint %q reaches %v; a reach is %s",
m.Module, name, reach, strings.Join(reaches, ", "))
}
if _, also := exposed[port]; also {
return nil, fmt.Errorf(
"%s says how far %q reaches and also exposes port %d. They say the same thing "+
"in different words; keep the endpoint's own block",
m.Module, name, port)
}
ep.Reach = text
}
if at, said := values["port"]; said {
machine, ok := asPort(at)
if !ok {
return nil, fmt.Errorf("%s puts the endpoint %q on %v, which is not a port",
m.Module, name, at)
}
ep.Port = machine
}
if label, said := values["label"]; said {
text, ok := label.(string)
if !ok || strings.TrimSpace(text) == "" {
return nil, fmt.Errorf("%s gives the endpoint %q a label that is not a name: %v",
m.Module, name, label)
}
ep.Label = strings.TrimSpace(text)
}
out[name] = ep
}
}
if len(out) == 0 {
return nil, nil
}
return out, nil
}
// spokenEndpoints is what a module's endpoints are called, as a refusal lists them — so a reader who
// named one wrongly is one edit from right, and a module that has named none is told so.
func spokenEndpoints(m Manifest) string {
names := make([]string, 0, len(m.Listens))
for _, l := range m.Listens {
if name := strings.TrimSpace(l.Name); name != "" {
names = append(names, name)
}
}
if len(names) == 0 {
return "no endpoints by name"
}
sort.Strings(names)
return strings.Join(names, ", ")
}
@@ -19,7 +19,7 @@ func TestTheBrokersPortIsOpenedThoughNoModuleDeclaresIt(t *testing.T) {
// A machine on the private network, with one ordinary module rule, and nothing that mentions
// the broker — which is every machine.
rules := []Rule{{Port: 8080, From: FromMesh, Because: []string{"some-module"}}}
out := AsNftables(rules, []string{"10.42.0.1"}, false, []int{brokerPort})
out := AsNftables(rules, []string{"10.42.0.1"}, false, []int{brokerPort}, nil, "mesh0")
if !strings.Contains(out, "tcp dport 5671 accept") {
t.Fatalf("the broker's port is not opened, so no machine could enrol:\n%s", out)
@@ -48,7 +48,7 @@ func TestTheBrokersPortIsOpenedThoughNoModuleDeclaresIt(t *testing.T) {
// And a mesh that was never told about a broker still gets a ruleset, rather than an empty one or
// a panic. A control plane in that state cannot issue tokens either, which is where it surfaces.
func TestNoBrokerMeansNoFoundationRuleRatherThanNoRuleset(t *testing.T) {
out := AsNftables(nil, []string{"10.42.0.1"}, false, nil)
out := AsNftables(nil, []string{"10.42.0.1"}, false, nil, nil, "mesh0")
if !strings.Contains(out, "table inet mesh") {
t.Fatalf("no ruleset at all:\n%s", out)
}
+206 -21
View File
@@ -79,7 +79,7 @@ func TestTwoModulesWantingOnePortAreBothNamed(t *testing.T) {
t.Fatalf("a module that wanted this port open is not named: %+v", rules[0])
}
// The consequence, which is the reason this matters: removing web must not read as closing 443.
nft := AsNftables(rules, nil, false, nil)
nft := AsNftables(rules, nil, false, nil, nil, "mesh0")
if !strings.Contains(nft, "web") || !strings.Contains(nft, "board") {
t.Fatalf("the rendered rule set does not name both sources:\n%s", nft)
}
@@ -107,7 +107,7 @@ func TestAPortOpenToEveryoneIsNotAlsoRestrictedToTheMesh(t *testing.T) {
func TestWhatNoModuleDeclaredIsClosed(t *testing.T) {
nft := AsNftables(mustFilter(t, Resolution{Modules: []Manifest{
{Module: "web", Listens: []Listening{{Port: 443, From: FromEverywhere}}},
}}, nil), []string{"198.51.100.2"}, false, nil)
}}, nil), []string{"198.51.100.2"}, false, nil, nil, "mesh0")
// Naming the chain, not just the policy: the forward chain drops too, and an assertion on
// "policy drop" alone passes while the input chain accepts everything. It did, once, here.
if !strings.Contains(nft, "type filter hook input priority filter; policy drop;") {
@@ -134,7 +134,7 @@ func TestWhatNoModuleDeclaredIsClosed(t *testing.T) {
// `flush ruleset` would do the first and not the second: it empties every table on the machine,
// including the ones the container runtime writes for its bridges.
func TestReloadingReplacesOnlyTheMeshsOwnRules(t *testing.T) {
nft := AsNftables(nil, nil, false, nil)
nft := AsNftables(nil, nil, false, nil, nil, "mesh0")
if strings.Contains(nft, "flush ruleset") {
t.Fatalf("loading the rule set empties every table on the machine:\n%s", nft)
}
@@ -160,22 +160,122 @@ func TestReloadingReplacesOnlyTheMeshsOwnRules(t *testing.T) {
// So the chain exists and denies by default, and the runtime's own networks are allowed explicitly
// — which is how the system being replaced has been doing it on these machines for months.
func TestWhatIsForwardedIsGovernedToo(t *testing.T) {
nft := AsNftables(nil, []string{"198.51.100.2"}, false, nil)
nft := AsNftables(nil, []string{"198.51.100.2"}, false, nil, nil, "mesh0")
if !strings.Contains(nft, "hook forward priority filter; policy drop") {
t.Fatalf("forwarded traffic is not governed, so container ports are open:\n%s", nft)
}
}
// And containers keep working, which is the whole reason the chain was left out before.
func TestTheRuntimesOwnNetworksKeepWorking(t *testing.T) {
nft := AsNftables(nil, []string{"198.51.100.2"}, false, nil)
for _, network := range []string{"172.16.0.0/12", "192.168.128.0/17"} {
if !strings.Contains(nft, "ip saddr "+network+" accept") {
t.Fatalf("%s is not allowed, so denying by default stops every container:\n%s", network, nft)
// And this machine's own guests keep working, which is the whole reason the chain was left out
// before — by not being mentioned (novox/hq ADR 0140).
//
// It used to be done by naming the address ranges they sit on: two fixed here and the rest recorded
// per machine. That list broke a workstation's containers at a flip and could not be made correct,
// because a range describes one machine and cannot tell a network the mesh made from one a
// predecessor left behind. What replaced it is a single line about the links traffic arrives on.
func TestThisMachinesOwnGuestsKeepWorkingWithoutBeingNamed(t *testing.T) {
nft := AsNftables(nil, []string{"198.51.100.2"}, false, nil, []string{"eth0"}, "mesh0")
if !strings.Contains(nft, `iifname != { "eth0", "mesh0" } accept`) {
t.Fatalf("what did not arrive from outside is not accepted, so this machine's own guests "+
"reach nothing:\n%s", nft)
}
}
// No address of a machine's own networks appears anywhere in a rendered filter.
//
// This is the assertion that fails against the previous behaviour, and it is why it is written on
// the text rather than on an outcome: the two ranges were a constant in this file, so nothing but
// reading the output catches one creeping back in.
func TestNoNetworkOfTheMachinesOwnIsNamed(t *testing.T) {
nft := AsNftables(nil, []string{"198.51.100.2"}, false, nil, []string{"eth0"}, "mesh0")
for _, gone := range []string{"172.16.0.0/12", "192.168.128.0/17", "saddr 192.168", "saddr 172."} {
if strings.Contains(nft, gone) {
t.Fatalf("%q is named, and a range describes one machine and goes stale in silence:\n%s",
gone, nft)
}
}
}
// **The tunnel is constrained, not treated as inside.**
//
// Accepting everything arriving over the private network would make a port nothing declares
// reachable from every machine in the mesh — the derivation abandoned, and a rule that reads as a
// restriction while restricting nothing. So the tunnel is named beside the outward links, and
// traffic arriving on it meets the declared rules like anything else.
func TestTheTunnelIsConstrainedLikeAnOutwardLink(t *testing.T) {
nft := AsNftables(nil, []string{"198.51.100.2"}, false, nil, []string{"eth0"}, "mesh0")
line := `iifname != { "eth0", "mesh0" } accept`
if !strings.Contains(nft, line) {
t.Fatalf("the tunnel is not constrained, so an undeclared port is reachable from any "+
"machine in the mesh:\n%s", nft)
}
}
// A machine with two links facing outside has both constrained. Asserted on the one line, because a
// rule covering one and not the other would leave a machine filtering half of what reaches it.
func TestEveryOutwardLinkIsConstrained(t *testing.T) {
nft := AsNftables(nil, []string{"198.51.100.2"}, false, nil, []string{"eth0", "wlan0"}, "mesh0")
if !strings.Contains(nft, `iifname != { "eth0", "wlan0", "mesh0" } accept`) {
t.Fatalf("not every outward link is constrained:\n%s", nft)
}
}
// A guest asks its host for an address and for names, and those two arrive at the input chain. Asked
// for by the link they arrive on, so a resolver bound anywhere but an outward link keeps answering.
func TestGuestsMayAskTheirHostForAnAddressAndNames(t *testing.T) {
nft := AsNftables(nil, []string{"198.51.100.2"}, false, nil, []string{"eth0"}, "mesh0")
for _, want := range []string{
`iifname != { "eth0", "mesh0" } udp dport { 53, 67 } accept`,
`iifname != { "eth0", "mesh0" } tcp dport 53 accept`,
} {
if !strings.Contains(nft, want) {
t.Fatalf("a guest cannot ask its host for an address or a name, which is not a closed "+
"port but a network that does not work:\n%s", nft)
}
}
}
// With no link named at all the chain denies rather than rendering an empty set, which nftables
// refuses — and a rule set that does not load is a machine filtering nothing while its unit reports
// success. Composing a declaration for such a machine is refused upstream; this is the floor.
func TestNoLinkNamedRendersNoCatchAllRatherThanAnEmptySet(t *testing.T) {
nft := AsNftables(nil, []string{"198.51.100.2"}, false, nil, nil, "")
if strings.Contains(nft, "{ }") || strings.Contains(nft, "iifname != {}") {
t.Fatalf("an empty set is rendered, which nftables refuses:\n%s", nft)
}
if !strings.Contains(nft, "hook forward priority filter; policy drop") {
t.Fatalf("the forward chain does not deny:\n%s", nft)
}
}
// A machine that has not said which links face outside is sent no filter, and the refusal names the
// module that would have loaded it so the reader knows what is being withheld.
func TestAMachineThatNamedNoOutwardLinkIsSentNoFilter(t *testing.T) {
r := Resolution{Node: "anchor", Modules: []Manifest{
{Module: "nftables", Filtering: &Filtering{Into: "/etc/mesh/filter.nft"}},
{Module: "web", Listens: []Listening{{Port: 443, From: FromEverywhere}}},
}}
_, err := r.Declaration(Rendering{Mesh: []string{"198.51.100.2"}, TunnelInterface: "mesh0"})
if err == nil {
t.Fatal("a machine that named no outward link was sent a filter written around none")
}
for _, want := range []string{"anchor", "nftables", "face outside"} {
if !strings.Contains(err.Error(), want) {
t.Fatalf("the refusal does not say %q: %v", want, err)
}
}
}
// And a machine that names none but loads no filter is not refused: there is nothing to write.
func TestAMachineWithNoFilterModuleIsNotRefused(t *testing.T) {
r := Resolution{Node: "anchor", Modules: []Manifest{
{Module: "web", Listens: []Listening{{Port: 443, From: FromEverywhere}}},
}}
if _, err := r.Declaration(Rendering{Mesh: []string{"198.51.100.2"}}); err != nil {
t.Fatalf("a machine that loads no filter was refused one: %v", err)
}
}
// A published port is matched by what the client asked for, not by where the packet ends up.
//
// The runtime rewrites the destination before this chain sees it, so a rule naming the published
@@ -183,7 +283,7 @@ func TestTheRuntimesOwnNetworksKeepWorking(t *testing.T) {
func TestAPublishedPortIsMatchedByWhatWasAskedFor(t *testing.T) {
nft := AsNftables(mustFilter(t, Resolution{Modules: []Manifest{
{Module: "web", Listens: []Listening{{Port: 8080, From: FromEverywhere}}},
}}, nil), []string{"198.51.100.2"}, false, nil)
}}, nil), []string{"198.51.100.2"}, false, nil, nil, "mesh0")
if !strings.Contains(nft, "ct original proto-dst 8080 accept") {
t.Fatalf("the forwarded rule does not match the port a client asked for:\n%s", nft)
}
@@ -193,7 +293,7 @@ func TestAPublishedPortIsMatchedByWhatWasAskedFor(t *testing.T) {
func TestAMeshScopedPortIsMeshScopedWhenForwarded(t *testing.T) {
nft := AsNftables(mustFilter(t, Resolution{Modules: []Manifest{
{Module: "store", Listens: []Listening{{Port: 5432, From: FromMesh}}},
}}, nil), []string{"198.51.100.2"}, false, nil)
}}, nil), []string{"198.51.100.2"}, false, nil, nil, "mesh0")
if !strings.Contains(nft, "ip saddr { 198.51.100.2 } ct original proto-dst 5432 accept") {
t.Fatalf("a mesh-only port is reachable from anywhere once forwarded:\n%s", nft)
}
@@ -203,7 +303,7 @@ func TestAMeshScopedPortIsMeshScopedWhenForwarded(t *testing.T) {
func TestFromTheMeshIsTheNodesTheMeshKnows(t *testing.T) {
nft := AsNftables(mustFilter(t, Resolution{Modules: []Manifest{
{Module: "store", Listens: []Listening{{Port: 5432, From: FromMesh}}},
}}, nil), []string{"198.51.100.2", "198.51.100.3"}, false, nil)
}}, nil), []string{"198.51.100.2", "198.51.100.3"}, false, nil, nil, "mesh0")
if !strings.Contains(nft, "ip saddr { 198.51.100.2, 198.51.100.3 } tcp dport 5432 accept") {
t.Fatalf("a mesh-scoped port was not restricted to the mesh's addresses:\n%s", nft)
}
@@ -213,7 +313,7 @@ func TestFromTheMeshIsTheNodesTheMeshKnows(t *testing.T) {
func TestAMeshPortOnANodeWithNoMeshIsClosedAndSaysSo(t *testing.T) {
nft := AsNftables(mustFilter(t, Resolution{Modules: []Manifest{
{Module: "store", Listens: []Listening{{Port: 5432, From: FromMesh}}},
}}, nil), nil, false, nil)
}}, nil), nil, false, nil, nil, "mesh0")
if strings.Contains(nft, "dport 5432 accept") {
t.Fatalf("a port meant for the mesh was opened to everything:\n%s", nft)
}
@@ -226,7 +326,7 @@ func TestAMeshPortOnANodeWithNoMeshIsClosedAndSaysSo(t *testing.T) {
func TestAMachineScopedPortIsNotOpened(t *testing.T) {
nft := AsNftables(mustFilter(t, Resolution{Modules: []Manifest{
{Module: "cache", Listens: []Listening{{Port: 6379, From: FromMachine}}},
}}, nil), []string{"198.51.100.2"}, false, nil)
}}, nil), []string{"198.51.100.2"}, false, nil, nil, "mesh0")
if strings.Contains(nft, "dport 6379 accept") {
t.Fatalf("a port for this machine only was opened to the network:\n%s", nft)
}
@@ -238,7 +338,8 @@ func TestTheModuleAskingForTheRuleSetGetsEveryModulesPorts(t *testing.T) {
{Module: "firewall", Filtering: &Filtering{Into: "/etc/mesh/filter.nft"}},
{Module: "web", Listens: []Listening{{Port: 443, From: FromEverywhere}}},
}}
out, err := r.Declaration(Rendering{Mesh: []string{"198.51.100.2"}})
out, err := r.Declaration(Rendering{Mesh: []string{"198.51.100.2"},
OutwardLinks: []string{"eth0"}, TunnelInterface: "mesh0"})
if err != nil {
t.Fatalf("declaration: %v", err)
}
@@ -268,7 +369,7 @@ func TestAskingForTheRuleSetWithNowhereToPutItIsRefused(t *testing.T) {
func TestAMeshOnBothAddressFamiliesRendersBoth(t *testing.T) {
nft := AsNftables(mustFilter(t, Resolution{Modules: []Manifest{
{Module: "store", Listens: []Listening{{Port: 5432, From: FromMesh}}},
}}, nil), []string{"198.51.100.2", "2001:db8::2"}, false, nil)
}}, nil), []string{"198.51.100.2", "2001:db8::2"}, false, nil, nil, "mesh0")
if !strings.Contains(nft, "ip saddr { 198.51.100.2 } tcp dport 5432 accept") {
t.Fatalf("the machines with v4 addresses were dropped:\n%s", nft)
}
@@ -296,7 +397,8 @@ func TestWhatTheMeshComputesIsAppliedBeforeWhatTheModuleDeclared(t *testing.T) {
"restart-on": []any{"filtering"}},
},
}}}
out, err := r.Declaration(Rendering{Mesh: []string{"198.51.100.2"}})
out, err := r.Declaration(Rendering{Mesh: []string{"198.51.100.2"},
OutwardLinks: []string{"eth0"}, TunnelInterface: "mesh0"})
if err != nil {
t.Fatalf("declaration: %v", err)
}
@@ -672,7 +774,7 @@ func TestExposureRefusesAPortNotListenedOnAndABadSource(t *testing.T) {
// loading the rules lives on conntrack until it drops, and then the machine is reached from a
// rescue console (novox/hq issue 047).
func TestSSHIsOpenFromTheMeshEvenWhenNothingIsAssigned(t *testing.T) {
nft := AsNftables(nil, []string{"198.51.100.2", "198.51.100.3"}, false, nil)
nft := AsNftables(nil, []string{"198.51.100.2", "198.51.100.3"}, false, nil, nil, "mesh0")
if !strings.Contains(nft, "ip saddr { 198.51.100.2, 198.51.100.3 } tcp dport 22 accept") {
t.Fatalf("ssh is not open to the mesh, so a machine can lock everyone out:\n%s", nft)
}
@@ -685,7 +787,7 @@ func TestSSHIsOpenFromTheMeshEvenWhenNothingIsAssigned(t *testing.T) {
// And from outside as well, on a machine that faces outward — because that is the way in when the
// private network is the thing that broke.
func TestSSHIsOpenFromOutsideOnAMachineThatFacesIt(t *testing.T) {
nft := AsNftables(nil, []string{"198.51.100.2"}, true, nil)
nft := AsNftables(nil, []string{"198.51.100.2"}, true, nil, nil, "mesh0")
if !strings.Contains(nft, "\t\ttcp dport 22 accept") {
t.Fatalf("a machine reachable from outside does not answer ssh there:\n%s", nft)
}
@@ -697,8 +799,91 @@ func TestSSHIsOpenFromOutsideOnAMachineThatFacesIt(t *testing.T) {
// to narrow the rule to, so narrowing it shuts the port entirely — on the first machine anybody
// adopts, reached over the network, closed by the act of adopting it.
func TestSSHIsNeverLeftWithoutARule(t *testing.T) {
nft := AsNftables(nil, nil, false, nil)
nft := AsNftables(nil, nil, false, nil, nil, "mesh0")
if !strings.Contains(nft, "tcp dport 22 accept") {
t.Fatalf("a machine with no mesh addresses has no ssh rule, so adopting it locks it:\n%s", nft)
}
}
// chainBody is one chain's own lines, so an assertion cannot be satisfied by an identical line in
// another chain.
//
// **Written because that happened.** The rule letting this machine's own callers through appears in the
// input chain and, in the same words, in the forward chain. A test asserting on the whole rendered file
// passed with the input chain's copy deleted — it was reading the forward chain's. ADR 0137's own tests
// say to assert per chain body for exactly this reason, and this file was not doing it.
func chainBody(t *testing.T, nft, chain string) string {
t.Helper()
open := "\tchain " + chain + " {"
i := strings.Index(nft, open)
if i < 0 {
t.Fatalf("no chain %q in:\n%s", chain, nft)
}
rest := nft[i+len(open):]
j := strings.Index(rest, "\n\t}")
if j < 0 {
t.Fatalf("chain %q does not close in:\n%s", chain, nft)
}
return rest[:j]
}
// **Anything on this machine may call anything on this machine.**
//
// Local is not a boundary this mesh draws, and whether a caller sits in a container was never meant to
// change the answer. It did, because the chain asked about addresses: a caller on the machine carries
// the machine's address and a caller in one of its containers carries a bridge address, so a rule
// naming the machines' own addresses silently refused every container on them.
//
// Measured: a module reaching its database on its own machine's name timed out for eleven hours while
// the machine itself could reach it (novox/hq 04-ISSUES/145).
func TestAnythingOnThisMachineMayCallAnythingOnIt(t *testing.T) {
nft := AsNftables(mustFilter(t, Resolution{Modules: []Manifest{
{Module: "store", Listens: []Listening{{Port: 5432, From: FromMesh}}},
{Module: "private", Listens: []Listening{{Port: 9999, From: FromMachine}}},
}}, nil), []string{"10.10.0.1", "10.10.0.2"}, false, nil, []string{"eth0"}, "mesh0")
// In the INPUT chain, which is where a call to a service on this machine arrives. The forward
// chain carries the same line in the same words, so asserting on the whole file proves nothing.
input := chainBody(t, nft, "input")
if !strings.Contains(input, `iifname != { "eth0", "mesh0" } accept`) {
t.Fatalf("a caller on this machine cannot reach a service on it:\n%s", input)
}
// One rule, for every service here — not a line per port that only covers the ports somebody
// remembered to think about.
if strings.Contains(input, `iifname != { "eth0", "mesh0" } tcp dport 5432`) {
t.Fatalf("the local allowance is still written per port:\n%s", input)
}
// And the private network still reaches what is exposed to it, which is a different question.
if !strings.Contains(input, "ip saddr { 10.10.0.1, 10.10.0.2 } tcp dport 5432 accept") {
t.Fatalf("the private network no longer reaches a service exposed to it:\n%s", input)
}
}
// The three reaches, as three lines. This is the whole of what the filter says about who may call what.
func TestTheThreeReachesAreThreeLines(t *testing.T) {
nft := AsNftables(mustFilter(t, Resolution{Modules: []Manifest{
{Module: "internal-only", Listens: []Listening{{Port: 5432, From: FromMesh}}},
{Module: "public", Listens: []Listening{{Port: 443, From: FromEverywhere}}},
}}, nil), []string{"10.10.0.1"}, false, nil, []string{"eth0"}, "mesh0")
input := chainBody(t, nft, "input")
for what, want := range map[string]string{
"on this machine": `iifname != { "eth0", "mesh0" } accept`,
"over the private network": "ip saddr { 10.10.0.1 } tcp dport 5432 accept",
"from anywhere": "tcp dport 443 accept",
} {
if !strings.Contains(input, want) {
t.Fatalf("a caller %s cannot reach what is exposed to it (%q):\n%s", what, want, input)
}
}
}
// A port open to everything needs no such line — it is already open to a guest.
func TestAPublicPortNeedsNoGuestLine(t *testing.T) {
nft := AsNftables(mustFilter(t, Resolution{Modules: []Manifest{
{Module: "web", Listens: []Listening{{Port: 443, From: FromEverywhere}}},
}}, nil), []string{"10.10.0.1"}, false, nil, []string{"eth0"}, "mesh0")
if strings.Count(nft, `iifname != { "eth0", "mesh0" } tcp dport 443`) != 0 {
t.Fatalf("a public port was given a guest line it does not need:\n%s", nft)
}
}
+83 -1
View File
@@ -571,6 +571,21 @@ type Artifact struct {
// image built from this same module's own repository, the same as every other artifact.
Context *ArtifactContext `json:"context,omitempty"`
// System is the operating system this artifact is compiled for, for a bundle whose output is a
// binary rather than portable code (novox/hq ADR 0142).
//
// **Named by the artifact, not by the recipe.** A toolchain deliberately accepts nothing from
// the module — anything a module could override there it would be writing a Dockerfile to
// override — and yet a compiled binary is per operating system, pinned at link time so a host
// refuses to touch a machine it was not built for (novox/hq ADR 0005). The way out is that the
// target is a property of the artifact: one artifact declared per system, one build each, and
// the recipe stays the mesh's.
//
// Empty for a bundle whose output runs anywhere, which is every interpreted language, and for
// every other kind. A bundle in a language that compiles to a binary must say one, because
// "compiled for whatever the build machine happened to be" is the fault this exists to prevent.
System string `json:"system,omitempty"`
// Language is what this module's code is written in, for a bundle.
//
// **Declared, never guessed.** Inferring it from what files happen to be present makes a
@@ -642,8 +657,23 @@ const (
// on is a fact, and it should be written once.
const ArtifactStoreProvision = "artifact-store"
// Listening is one port a module accepts connections on.
// Listening is one endpoint a module serves: a port it accepts connections on, and what may be said
// about that port from outside the module.
type Listening struct {
// Name is what this endpoint is called, so an assignment and a route can refer to it as one thing
// (novox/hq ADR 0138).
//
// **Because a port number is not a name.** Three facts have to be said about an endpoint when a
// module is assigned — which machine port it lands on, the subdomain a proxy serves it under, and
// how far it reaches — and they were said in three places keyed by the port. A module with two
// endpoints of different shapes, a web surface behind a proxy and a protocol port clients dial
// directly, cannot be configured that way without a reader joining numbers by hand.
//
// The module's to choose, like the route's label: it names its own parts. Lowercase, and unique
// within the module, so a reference to it is unambiguous. Empty is allowed and means an endpoint
// nothing refers to by name, which is every endpoint in the catalogue until they are named.
Name string `json:"name,omitempty"`
Port int `json:"port"`
// Protocol is "tcp" or "udp". Absent means tcp, which is what almost everything is — and a
// field that had to be written every time would be written wrongly some of the time.
@@ -1250,6 +1280,8 @@ func ParseManifest(raw []byte) (Manifest, error) {
"%s listens on %d over %q, which is tcp or udp", m.Module, l.Port, p))
}
}
problems = append(problems, endpointNameProblems(m)...)
problems = append(problems, RouteProblems(m)...)
for _, port := range m.Guards {
if port < 1 || port > 65535 {
problems = append(problems, fmt.Sprintf(
@@ -1674,3 +1706,53 @@ func (m Manifest) undeclaredMounts() []string {
}
return problems
}
// endpointName is what an endpoint may be called: lowercase letters, digits and dashes, starting
// with a letter. The same shape a label has, because both end up in something a person types.
var endpointName = regexp.MustCompile(`^[a-z][a-z0-9-]*$`)
// endpointNameProblems holds a module's endpoint names to being usable as references (novox/hq ADR
// 0138).
//
// **Unique, because the point of a name is that it identifies one thing.** Two endpoints called the
// same would make an assignment that configures one silently configure whichever the mesh read last
// — the shape of fault this repository keeps finding, where a declaration appears to say something
// and says something else.
func endpointNameProblems(m Manifest) []string {
var problems []string
seen := map[string]int{}
for _, l := range m.Listens {
name := strings.TrimSpace(l.Name)
if name == "" {
continue
}
if !endpointName.MatchString(name) {
problems = append(problems, fmt.Sprintf(
"%s calls the endpoint on port %d %q; a name is lowercase letters, digits and "+
"dashes, starting with a letter", m.Module, l.Port, l.Name))
continue
}
if before, already := seen[name]; already {
problems = append(problems, fmt.Sprintf(
"%s calls both port %d and port %d %q, so anything naming that endpoint could mean "+
"either", m.Module, before, l.Port, name))
continue
}
seen[name] = l.Port
}
return problems
}
// EndpointPort is the port of the endpoint a module calls this, and whether it has one.
func EndpointPort(m Manifest, name string) (int, bool) {
want := strings.TrimSpace(name)
if want == "" {
return 0, false
}
for _, l := range m.Listens {
if strings.TrimSpace(l.Name) == want {
return l.Port, true
}
}
return 0, false
}
+1 -1
View File
@@ -12,5 +12,5 @@ func TestPrintRehearsalRuleset(t *testing.T) {
rules := mustFilter(t, Resolution{Modules: []Manifest{
{Module: "pub", Listens: []Listening{{Port: 8099, From: FromMesh, Why: "the thing it serves"}}},
}}, nil)
t.Log("\n" + AsNftables(rules, []string{"192.0.2.20"}, true, nil))
t.Log("\n" + AsNftables(rules, []string{"192.0.2.20"}, true, nil, nil, "mesh0"))
}
+206
View File
@@ -0,0 +1,206 @@
package catalogue
import (
"strings"
"testing"
)
// a web module with one routed endpoint, the shape almost every routed module in the catalogue has.
func aRoutedWeb() Manifest {
return Manifest{
Module: "web",
Listens: []Listening{{Port: 3000, From: FromMesh}},
Contributes: map[string]map[string]any{
"route": {"label": "app", "port": 3000},
},
}
}
func reachSet(reach string) SettingsBy {
return SettingsBy{"web": {{From: "node anchor",
Values: map[string]any{ReachSetting: map[string]any{"3000": reach}}}}}
}
// namesFor renders the contribution a routed module makes and returns the two names it carries.
func namesFor(t *testing.T, m Manifest, settings SettingsBy) (public, internal string) {
t.Helper()
r := Resolution{Node: "anchor", Modules: []Manifest{m},
PublicDomain: "example.test", At: "anchor.internal"}
given, err := r.contributions(settings, nil, nil)
if err != nil {
t.Fatalf("contributions: %v", err)
}
for _, c := range given["route"] {
p, _ := c.Values["name"].(string)
i, _ := c.Values["internal-name"].(string)
return p, i
}
t.Fatal("the module contributed no route")
return "", ""
}
// **Nothing said composes both names, exactly as before.** This is the assertion that keeps every
// mesh already running identical until an assignment speaks, and it is the one that would break first
// if reach were read where it should not be.
func TestAnEndpointWithNoReachKeepsBothNames(t *testing.T) {
public, internal := namesFor(t, aRoutedWeb(), nil)
if public != "app.example.test" || internal != "app.anchor.internal" {
t.Fatalf("names are %q and %q, want both composed as before", public, internal)
}
}
// An internal endpoint has an internal name and no public one — so the proxy serves it inside, and
// the public authority is never asked for a name nobody wanted. This is what "must not be public"
// could not say before.
func TestAnInternalEndpointHasNoPublicName(t *testing.T) {
public, internal := namesFor(t, aRoutedWeb(), reachSet(ReachInternal))
if public != "" {
t.Fatalf("an internal endpoint composed the public name %q", public)
}
if internal != "app.anchor.internal" {
t.Fatalf("internal name is %q, want app.anchor.internal", internal)
}
}
// And the mirror: a public endpoint gets the public name and not the internal one, so the mesh's own
// authority is not asked to certify a name the service is not reached by.
func TestAPublicEndpointHasNoInternalName(t *testing.T) {
public, internal := namesFor(t, aRoutedWeb(), reachSet(ReachPublic))
if internal != "" {
t.Fatalf("a public endpoint composed the internal name %q", internal)
}
if public != "app.example.test" {
t.Fatalf("public name is %q, want app.example.test", public)
}
}
func TestBothComposesBothNames(t *testing.T) {
public, internal := namesFor(t, aRoutedWeb(), reachSet(ReachBoth))
if public == "" || internal == "" {
t.Fatalf("both should compose both names, got %q and %q", public, internal)
}
}
// **The filter reads the same value — for an endpoint the proxy does not serve.**
//
// A routed endpoint's port is how the proxy reaches it and nothing else (ADR 0045): a public service
// listens from the mesh, only the proxy reaches it, and it is exposed by name. So on a routed
// endpoint the reach asks for a name and the port keeps what the manifest said.
func TestAnUnroutedEndpointsPortFollowsItsReach(t *testing.T) {
// The same module with its route taken away: now the port is the only way in, so reach governs it.
bare := aRoutedWeb()
bare.Contributes = nil
for _, c := range []struct{ reach, want string }{
{ReachInternal, FromMesh},
{ReachPublic, FromEverywhere},
{ReachBoth, FromEverywhere},
{ReachMachine, FromMachine},
} {
r := Resolution{Node: "anchor", Modules: []Manifest{bare}}
rules, err := r.Rules(Rendering{Settings: reachSet(c.reach)})
if err != nil {
t.Fatalf("%s: rules: %v", c.reach, err)
}
found := false
for _, rule := range rules {
if rule.Port == 3000 {
found = true
if rule.From != c.want {
t.Fatalf("reach %q made the filter say %q, want %q", c.reach, rule.From, c.want)
}
}
}
if !found {
t.Fatalf("reach %q produced no rule for the port", c.reach)
}
}
}
// **A public name does not open the machine's port**, which is the case that found this.
//
// A module whose routed name must be public and whose machine-side port must not be had no way to say
// so while one value drove both. Under one value it could not be expressed; the port would reopen.
func TestAPublicNameLeavesARoutedPortAsTheManifestSaid(t *testing.T) {
r := Resolution{Node: "anchor", Modules: []Manifest{aRoutedWeb()},
PublicDomain: "example.test", At: "anchor.internal"}
rules, err := r.Rules(Rendering{Settings: reachSet(ReachPublic)})
if err != nil {
t.Fatal(err)
}
for _, rule := range rules {
if rule.Port == 3000 && rule.From != FromMesh {
t.Fatalf("a public reach opened a routed port to %q; the proxy is how it is reached",
rule.From)
}
}
// And the name it asked for is there, so the reach was not simply ignored.
public, internal := namesFor(t, aRoutedWeb(), reachSet(ReachPublic))
if public != "app.example.test" || internal != "" {
t.Fatalf("names are %q and %q, want the public one only", public, internal)
}
}
// A reach for a port the module does not listen on reaches nothing, and is refused where it is
// written rather than accepted and ignored.
func TestAReachForAPortTheModuleDoesNotListenOnIsRefused(t *testing.T) {
_, err := Reaches(aRoutedWeb(), []Layer{{From: "node anchor",
Values: map[string]any{ReachSetting: map[string]any{"9999": ReachInternal}}}})
if err == nil || !strings.Contains(err.Error(), "reaches nothing") {
t.Fatalf("a reach naming an undeclared port was accepted: %v", err)
}
}
// A value that is not a reach is refused, and the refusal names the four so a reader is one edit from
// right. "mesh" is the tempting wrong answer, because that is the filter's word for nearly the same
// thing.
func TestAValueThatIsNotAReachIsRefused(t *testing.T) {
for _, wrong := range []string{"mesh", "anywhere", "private", "true"} {
_, err := Reaches(aRoutedWeb(), []Layer{{From: "node anchor",
Values: map[string]any{ReachSetting: map[string]any{"3000": wrong}}}})
if err == nil || !strings.Contains(err.Error(), "a reach is") {
t.Fatalf("%q was accepted as a reach: %v", wrong, err)
}
}
}
// **A port that says both reach and expose is refused.** They say the same thing in different words,
// and accepting both would have the filter follow one while the names followed the other — the
// disagreement ADR 0138 exists to remove, reintroduced by the migration away from the older word.
func TestReachAndExposeForOnePortAreRefused(t *testing.T) {
_, err := Reaches(aRoutedWeb(), []Layer{{From: "node anchor", Values: map[string]any{
ReachSetting: map[string]any{"3000": ReachInternal},
ExposeSetting: map[string]any{"3000": FromEverywhere},
}}})
if err == nil || !strings.Contains(err.Error(), "same thing in different") {
t.Fatalf("a port set both ways was accepted: %v", err)
}
}
// A path-level refusal carries no port: it is a rule about a name, not an endpoint, and it inherits
// whatever that name turned out to be. Narrowing the endpoint must not silently drop it.
func TestARuleWithNoPortIsLeftAlone(t *testing.T) {
m := aRoutedWeb()
m.ContributesMany = map[string]map[string]map[string]any{
"route": {"refused": {"label": "app", "path": "/internal", "deny": true}},
}
r := Resolution{Node: "anchor", Modules: []Manifest{m},
PublicDomain: "example.test", At: "anchor.internal"}
given, err := r.contributions(reachSet(ReachInternal), nil, nil)
if err != nil {
t.Fatal(err)
}
var sawDeny bool
for _, c := range given["route"] {
if deny, _ := c.Values["deny"].(bool); deny {
sawDeny = true
// It keeps both, because it named no endpoint to be narrowed by.
if c.Values["name"] == nil || c.Values["internal-name"] == nil {
t.Fatalf("the path rule lost a name it shadows: %v", c.Values)
}
}
}
if !sawDeny {
t.Fatal("the path rule was dropped")
}
}
+11
View File
@@ -186,6 +186,17 @@ func UnusedSettings(m Manifest, layers []Layer) []string {
if key == PortsSetting {
continue
}
// `reach` says how far one of this module's endpoints reaches (novox/hq ADR 0138) — the
// filter's source, which names are composed, and therefore which authority certifies
// them. Validated in Reaches, so not stray.
if key == ReachSetting && len(m.Listens) > 0 {
continue
}
// `endpoints` configures a module's endpoints by name — the machine port, the subdomain and
// the reach as one block each (novox/hq ADR 0138). Validated in Endpoints, so not stray.
if key == EndpointsSetting && len(m.Listens) > 0 {
continue
}
unused = append(unused, fmt.Sprintf(
"%s sets %q, and %s has no file or contribution to merge it into",
layer.From, key, m.Module))
@@ -0,0 +1,19 @@
-- The networks a machine routes for what it hosts, beyond the container runtime's own defaults.
--
-- novox/hq ADR 0137. The derived packet filter denies forwarding by default and then allows the
-- container runtime's two default pools, named in the controller's code with a comment saying that
-- a machine configured otherwise "needs this to say so" — and no way to say it. So the filter was
-- correct only on a machine whose runtime used the defaults, and silently wrong on any other.
--
-- Measured on 2026-09-28: flipping a workstation to the derived filter cut egress for five of its
-- container networks and for every network its test beds create, because those are allocated from
-- ranges the two defaults do not cover. Nothing reported a fault; the containers simply could not
-- reach anything.
--
-- A node-level fact, beside the node's public domain and for the same reason: it is a property of
-- the machine, not of whichever module happens to load the filter today. Swapping that module must
-- not lose it.
--
-- Null for a machine that routes nothing but the runtime's defaults, which is the ordinary case and
-- what every machine held before this column existed.
alter table node add column routed_networks jsonb;
@@ -0,0 +1,26 @@
-- Which of a machine's links face outside it, replacing the networks it was told to say it routes.
--
-- novox/hq ADR 0140, superseding 0137 and 0139. The derived filter blocked everything passing
-- through a machine and then allowed the machine's own containers back by naming the address ranges
-- they sit on: two ranges fixed in the controller's source, the rest recorded by 0043's column.
--
-- Every route to a correct list fails. A constant describes one machine. A recorded range goes stale
-- in silence, and cannot tell a network the mesh made from one a predecessor left behind — measured
-- on the control-node, where six ranges fall outside the constants and two of the six belong to
-- services the mesh does not run. Generating the list from the modules put half the rule set on the
-- machine.
--
-- The list should not exist, because the mesh has no position on a container reaching outward: that
-- is not a port opened to anybody. The filter constrains what arrives from OUTSIDE the machine and
-- says nothing about what did not, which needs one fact instead of a list — which links "outside"
-- arrives on.
--
-- Reported by the machine on every apply, never recorded by hand, so it cannot go stale. Null for a
-- machine that has not reported yet; the mesh composes no filter for such a machine and leaves the
-- one it has, because a rule written around a link with no name is a rule set that does not load.
alter table node add column outward_links jsonb;
-- What 0043 recorded is not migrated into it. The ranges answered a question that no longer exists,
-- and every machine that named one keeps working without it: the traffic those ranges allowed is now
-- allowed by not having arrived from outside.
alter table node drop column routed_networks;
+55
View File
@@ -518,6 +518,61 @@ func (i *Inventory) PublicDomainOf(ctx context.Context, name string) (string, er
return *domain, nil
}
// RecordOutwardLinks keeps the links a machine reported as facing outside it.
//
// A reported fact, not a setting (novox/hq ADR 0140). It replaces the networks a machine used to be
// told to say it routes: the filter blocked everything passing through and then allowed the machine's
// own containers back by naming their address ranges, and every way of keeping that list correct
// failed — a constant describes one machine, and a recorded range goes stale in silence. The filter
// now constrains what arrives from outside and says nothing about what did not, and the one thing it
// needs is which links "outside" arrives on. The machine reads that from its own routing table on
// every apply, so it cannot go stale and nobody types it.
//
// An empty list clears it, which is what a machine with no route off itself reports. The mesh then
// composes no filter for that machine at all.
func (i *Inventory) RecordOutwardLinks(ctx context.Context, id string, links []string) error {
var kept []string
for _, name := range links {
if name = strings.TrimSpace(name); name != "" {
kept = append(kept, name)
}
}
if len(kept) == 0 {
_, err := i.store.Pool().Exec(ctx,
`update node set outward_links = null where id = $1`, id)
return err
}
body, err := json.Marshal(kept)
if err != nil {
return err
}
_, err = i.store.Pool().Exec(ctx,
`update node set outward_links = $2 where id = $1`, id, string(body))
return err
}
// OutwardLinksOf is the links a machine reported as facing outside it, empty when it has reported
// none — which is a machine the mesh composes no filter for.
func (i *Inventory) OutwardLinksOf(ctx context.Context, name string) ([]string, error) {
var body []byte
err := i.store.Pool().QueryRow(ctx,
`select outward_links from node where name = $1`, name).Scan(&body)
if errors.Is(err, pgx.ErrNoRows) {
return nil, fmt.Errorf("%w: %s", ErrNoSuchNode, name)
}
if err != nil {
return nil, err
}
if len(body) == 0 {
return nil, nil
}
var links []string
if err := json.Unmarshal(body, &links); err != nil {
return nil, fmt.Errorf("the outward links recorded for %s are not a list: %w", name, err)
}
return links, nil
}
// RecordOverlayKey keeps the public half a node generated.
func (i *Inventory) RecordOverlayKey(ctx context.Context, node, key string) error {
if strings.TrimSpace(key) == "" {
+11
View File
@@ -301,6 +301,17 @@ func (e Enrolment) Heard(ctx context.Context, report Report) (news bool, err err
return false, err
}
}
// Which of its links face outside (novox/hq ADR 0140), whenever it says so. Recorded on every
// report that carries it, adopted or converged, because the filter the mesh composes is written
// around it — and never cleared by a report that carries none, which is every bare word that the
// node is there. A machine whose routing table it could not read reports nothing rather than
// guessing, and keeps whatever it last said; a machine with genuinely no route off itself is one
// the mesh composes no filter for at all.
if len(report.Outward) > 0 {
if err := e.Inventory.RecordOutwardLinks(ctx, node.ID, report.Outward); err != nil {
return false, err
}
}
// What it says about the tunnel it carried (novox/hq ADR 0105), whenever it says it.
if report.Tunnel != nil {
if err := e.Inventory.RecordCarriedTunnel(ctx, node.ID, inventory.Carried{
+14
View File
@@ -170,6 +170,20 @@ type Report struct {
// Firewall is the firewall found on the machine — "ufw" or "none" — and empty on a node that
// was never asked, which is every converged one.
Firewall string `json:"firewall,omitempty"`
// Outward is the links on this machine that face outside it — the ones carrying a default route
// (novox/hq ADR 0140). Every node reports it, adopted or converged, because the filter the mesh
// composes for it is written around these and nothing else.
//
// **It replaces a list of addresses.** The filter used to block everything passing through the
// machine and then allow the machine's own containers back by naming the ranges they sit on. A
// range describes one machine and goes stale in silence; the link carrying the default route is
// read afresh on every report and does not change when a module is added or removed.
//
// Empty means the machine has not said. The mesh composes no filter for such a machine and
// leaves the one it has: a rule written around a link with no name is a rule set that does not
// load, and that is a machine filtering nothing while its unit reports success.
Outward []string `json:"outward,omitempty"`
// Reachable is what can be reached on the machine now: every listening socket and every
// published container port. Only an adopted node reports it; it is what converging previews.
Reachable []Reach `json:"reachable,omitempty"`
+2 -1
View File
@@ -346,7 +346,8 @@ func (s *Server) reported(ctx context.Context, m Control) {
// whenever it arrives, which is the behaviour the mesh has had all along.
func staleAgainst(report Report) string {
if report.Rekey != nil || report.Tunnel != nil || len(report.Held) > 0 ||
report.Firewall != "" || len(report.Reachable) > 0 || len(report.Carried) > 0 {
report.Firewall != "" || len(report.Reachable) > 0 || len(report.Carried) > 0 ||
len(report.Outward) > 0 {
return ""
}
return report.Declared