D348 - klogsrv prints its own banner

decided · 2026-08-27

D343 got two payloads as far as calling getopt. Six functions later one of them runs:

.----------------------------------------------------------------------.
|  _      _                                                   _    __  |
| | | __ | |   ___     __ _   ___   _ __  __   __       ___  | |  / _| |
...
| v0.9                                Copyright (C) 2025 John Tornblom |
'----------------------------------------------------------------------'
main.c:278:sysctl: error 0 (orbistoun has no message table)

That is the guest's own output, and the last line is the guest naming its own next blocker - by file and line, through this project's own strerror text. CLAUDE.md calls printf "the cheapest oracle in the project" for exactly this reason; here the whole diagnostic path had to exist before the program could say anything at all.

klogsrv went from 3 imports and 3 calls to 9 and 18; ftpsrv from 1 to 8.

What was built, and the one shape they share

function	why it was next
signal	first instruction after main - SIGPIPE, so a write to a closed socket does not kill the server
getopt	second - a server that cannot parse its arguments never opens a socket
__error	errno is *__error(), so a placeholder becomes a wild pointer the guest reads
strerror	fed straight into printf unchecked - a placeholder faults inside the print
fprintf	how the error path speaks
puts	eight calls of banner, ignored

Three of the six - getopt, __error, strerror - cannot be implemented as a return value. They report through storage: the guest's own optind, a per-thread errno, a message buffer. D307 predicted this was coming ("the first time the HLE layer would own state rather than functions") and D323 built half of it; this is the other half.

orbistoun-thunk::data_symbol closes it. DataBlocks already reserved a page per data import; it now keeps the names too, and publishes them before entry. An implementation that must leave something where the guest will read it can find the guest's own slot. A guest that never imported optarg gets no write rather than an invented one.

Entering at main needed arguments to match

main is main(int argc, char **argv), and EntryArgument's options all answer a different question - what a process entry point finds in its first register. Handing that to main gives it a wild count to iterate over wild pointers, which getopt then has to refuse.

EntryArgument::MainArguments reads argc and argv out of the process image already written to the guest stack. Read rather than constructed: a second copy could disagree with the one everything else in the guest will find.

enter_guest_with_argument grew a sibling taking two, rather than a second copy of forty lines of inline assembly - rsi was always being set, to nothing in particular.

Six knowledge records, because the tree refuses without them

every_implemented_function_is_written_down failed the moment signal was bound: "signal is implemented but nothing is recorded about it". That is principle 1's accounting working exactly as designed, and it caught all six in one go.

Five are published - POSIX and FreeBSD are the citable reference. strerror is assumed, and deliberately: its message is ours, says so, and does not imitate a table nobody here has seen. A guest comparing it against a known string would fail; no caller measured does, they print it.

What is next, and the guest said it

sysctl, at main.c:278. Then getpid. Then the sockets that are the actual point.