Read a refused line's request before answering it, so a busy line is read as busy
mesh/merge-gate pass: builds mesh-host → ace, g14, novox, shanks; no bus step; every machine composes with the change as it did without (4 of 4 compose)
mesh/repo-check pass: its merge-check.sh passed
mesh/delivery superseded: a newer head of the same pull request

The engine answered busy and closed before the client's request had landed;
the client's write then broke and it never read the refusal waiting for it.
Under the whole suite's load this failed the busy test (novox/hq issue 434).
This commit is contained in:
2026-10-11 02:35:20 +02:00
parent bd5cc69804
commit 70cb613614
2 changed files with 76 additions and 12 deletions
+21 -6
View File
@@ -55,6 +55,11 @@ type Forward func(ctx context.Context, asked Asked) cli.Answer
// holds nothing for long. A variable so a test need not wait.
var readWithin = 30 * time.Second
// refusedWithin bounds how long the engine reads the request of a line it refuses before it is asked: long enough
// for a client that has connected to finish writing, short enough that a refused connection holds nothing for long.
// A variable so a test can say.
var refusedWithin = 5 * time.Second
// atOnce bounds how many lines are asked at the same time, and perAccount how many of them one account's: a person
// types one at a time, so more is a program on the machine. One over either bound is answered busy at once — the
// engine never stops accepting for it, so one account cannot hold the socket from every other (review of ADR 0272).
@@ -247,31 +252,41 @@ func answer(ctx context.Context, conn *net.UnixConn, held *slots, forward Forwar
say(fmt.Sprintf("mesh-cli: the answer could not be written back: %v", err))
}
}
// A refusal is written only after the request it answers is read (novox/hq issue 434). The client writes its
// request first and reads the answer after; an engine that answers and closes before that write has landed
// makes the write fail with a broken pipe, and the client never reads the refusal waiting for it — under load a
// line beyond its account's bound read as an error instead of busy. The read is bounded by refusedWithin, and
// what it says is never asked; nor does it outlast readWithin, which bounds every wait before the line is read.
refuse := func(why string) {
_ = conn.SetReadDeadline(time.Now().Add(min(refusedWithin, readWithin)))
_, _ = cli.ReadRequest(conn)
reply(cli.Refusal(why))
}
uid, pid, err := peerOf(conn)
if err != nil {
// Never sent as anyone: an account the kernel did not give is no account at all.
reply(cli.Refusal("the node-engine could not read which account is asking, so nothing was asked: " + err.Error()))
refuse("the node-engine could not read which account is asking, so nothing was asked: " + err.Error())
return
}
// The user id is this machine's account only in this machine's namespaces: a container running with the
// operator's id shares the number and nothing else (review of ADR 0272).
if err := sameNamespaces(pid); err != nil {
reply(cli.Refusal("the asking process does not run in this machine's own namespaces, so it is not this " +
"machine's account and nothing was asked: " + err.Error()))
refuse("the asking process does not run in this machine's own namespaces, so it is not this " +
"machine's account and nothing was asked: " + err.Error())
return
}
// And the login session, now: judged as the peer was when it connected, before it can exec or exit while its
// line is read (review of ADR 0272).
session := sessionOf(pid, uid)
if !held.take(uid) {
reply(cli.Refusal(fmt.Sprintf("busy: user id %d has %d lines being asked already, or the machine %d; ask "+
"again when one has ended. Nothing was asked", uid, perAccount, atOnce)))
refuse(fmt.Sprintf("busy: user id %d has %d lines being asked already, or the machine %d; ask "+
"again when one has ended. Nothing was asked", uid, perAccount, atOnce))
return
}
defer held.give(uid)
account, err := lookupAccount(uid)
if err != nil || account == "" {
reply(cli.Refusal(fmt.Sprintf("user id %d is not in this machine's user database, so nothing was asked", uid)))
refuse(fmt.Sprintf("user id %d is not in this machine's user database, so nothing was asked", uid))
return
}
req, err := cli.ReadRequest(conn)