Correct the isLoopback javadoc: after #305 narrowing this range refuses a caller, it does not promote one
CI / contract (push) Successful in 1m16s
CI / build (push) Successful in 1m47s

This commit is contained in:
Dai Ha
2026-09-04 14:10:10 +07:00
parent de70aa38f1
commit c801851c66
@@ -90,13 +90,27 @@ public final class ConnectionIdentity {
* {@code 127.0.0.1:8765} with a source address of {@code 127.0.0.2} — measured on the Linux
* fleet host, where binding that source succeeds.
*
* <p><strong>Being strict here does not make the daemon safer; it makes it unsafe.</strong>
* That reads backwards, so it is worth stating plainly. This predicate does not decide whether
* a caller is trusted — it decides whether the caller's identity is <em>resolved at all</em>.
* Returning false means {@link #resolve} answers "no terminal", and downstream a caller with no
* terminal is treated as the primary under loopback-trust. So every address excluded here is an
* address on which a worker silently becomes the lead. Widening a check normally weakens it;
* widening this one is what closes the hole.
* <p><strong>What excluding an address costs, stated as it is today.</strong> This paragraph
* used to say that narrowing this range turned a worker into the lead, and that widening the
* check was what closed the hole. That was true only while there were <em>two</em> definitions
* that disagreed: {@code ConnectionIdentity} skipped the identity lookup for {@code 127.0.0.2}
* while {@code CallerResolver} read the same address as loopback and granted the primary role.
* #305 removed the second copy, and with one shared definition the old sentence no longer holds.
*
* <p>Measured on 2026-09-04 by narrowing this method back to exactly {@code 127.0.0.1} and
* running {@code CallerResolverTest} and {@code ConnectionIdentityTest}: a caller from
* {@code 127.0.0.2} then resolves to {@code ANONYMOUS}, not {@code PRIMARY} — for a worker
* ({@code aWorkerOnAnyLoopbackSourceAddressIsStillAWorkerNotThePrimary}) and for a non-worker
* ({@code aNonWorkerOnAnyLoopbackSourceAddressIsStillThePrimary}) alike. Excluding an address
* now <em>refuses</em> its caller; it does not promote one.
*
* <p>So keep the whole range, but for the plain reason: a genuine worker or primary that
* connects from {@code 127.0.0.2} must be identifiable at all, and narrowing this predicate
* locks it out. That is an outage, and an outage is the direction to fail in — which is exactly
* why the range must not be narrowed casually and also why doing so is no longer a security
* hole. This predicate still does not decide whether a caller is trusted; it decides whether the
* caller's identity is <em>resolved at all</em>. What makes an unresolved caller safe is
* {@link Caller#resolved()} (#317), not this method.
*/
public static boolean isLoopback(String addr) {
if (addr == null) {