mirror of https://github.com/jmcorgan/fips.git synced 2026-10-05 19:18:25 +00:00

Files

T

Johnathan Corgan 3a789370b9 Add an experimental native datagram API addressed by public key

A client process opens a flow to a peer's public key on a chosen port and
sends and receives datagrams on a file descriptor the daemon hands it. No
IPv6 emulation, no TUN device, no DNS: a datagram travels from key to key.
The feature is off by default and is not a stable interface.

The wire needs no change and gets none. Every FSP data packet has carried a
port pair inside its AEAD envelope since v0.2.0, and port 256 is simply the
IPv6 shim. What was missing was a way for a program to ask for a port of its
own and be handed the traffic.

Addressing is the part worth reading twice, because the obvious design is
wrong. The x-only public key is the address. An npub is that key written in
bech32, so converting between them is a local encoding rather than a lookup
or a name service. The 16-byte node address that travels on the wire is the
first half of a SHA-256 of the key: it is a truncated hash, it does not
invert, and it appears nowhere a client can see. An earlier iteration of this
work reported a peer by that hash and could supply a key only sometimes,
which is what treating a wire identifier as an identity produces.

An accepted flow therefore always knows its peer. The key is captured where
the peer is authenticated rather than looked up when a report is rendered:
every inbound datagram passes one call site inside a handler that refuses
anything whose session is not established, and the responder has already
rejected the session unless the claimed address derives from the key it
proved. Reaching for the identity cache instead gives a best-effort answer
from a structure that evicts.

A listener is a descriptor. The daemon writes one message per arrival to it,
carrying the new flow's descriptor and the peer's address, so poll, select
and epoll work on a listener and accepting is a recvmsg. That is what lets
the API be used from a program that already has an event loop, which a
command-and-reply listener could not support: an arrival could not be waited
on beside anything else. There is no accept command and no reject command.
Refusing a flow is closing the descriptor you were handed.

The Rust surface mirrors std::net. FipsStream::connect, FipsListener::bind,
incoming, accept, io::Result and an errno mapping rather than a bespoke
error type. An address is given as an npub, as a key, or as a pair, through
one parameter, the way ToSocketAddrs takes several spellings of one thing.
Each type holds its descriptor and copies of what setup told it and nothing
else, so a stream that outlives its setup connection is not representable.

set_nonblocking, AsFd and the four deadline methods carry the names and
signatures std::net uses for the same jobs. They were asked for by a user
integrating the API with tokio: AsyncFd requires a non-blocking descriptor,
and anything receiving from a peer needs a bounded wait. AsFd is the better
of the two descriptor accessors, because the borrow cannot outlive the value
that owns the descriptor, so a reactor cannot hold a registration for a
descriptor that has since been closed and its number reused by the next
open. The non-blocking flag is read, modified and written back rather than
assigned, since the flag word carries more than that one bit and a caller may
have set O_ASYNC. A zero timeout is refused with EINVAL, because the kernel
reads a zero timeval as "wait for ever", which inverts what a caller passing
zero means; std::net refuses it for the same reason. The two directions are
separate options and stay that way. FipsListener gets no timeout methods,
matching TcpListener: bounding an accept is set_nonblocking plus the caller's
own poll, which the reactor how-to builds. A flow taken from accept is
blocking whatever the listener was set to, because the two are separate
sockets and the daemon hands over a fresh one.

One rule has no counterpart in Berkeley sockets and a client author must know
it: the v1 wire carries no half-close, so nothing peer-driven ever closes a
flow. A server written to read until the flow ends waits for a signal that
cannot arrive, holding a thread and a flow per peer until its process exits.
A program decides its own termination, and the example serves one datagram
per flow.

The tests reach a live daemon rather than a stand-in. Every public item had a
unit test against a hand-written stand-in with canned replies, and the five
entry points a program actually calls first, connect, connect_from,
connect_at, bind and the SOCKET constant, had no coverage of any kind,
because the tests that appear to cover them build a Wire over a socket pair
and hand it to the private open and hold, so nothing ever resolved a socket
path or mapped its errors. examples/native-surface.rs walks all thirty-eight
items against a running daemon and reports the number of assertions it made.
The count is read from the recorder rather than written as a literal, and the
harness asserts the exit status, the completion marker and the count
together, so deleting an assertion fails the check rather than quietly
shrinking it. Watchdogs turn a hang into a named failure, which several of
the walked behaviours would otherwise produce. The shared Docker image is
built once for every integration leg, so the new binary is staged at all ten
places the existing one is, the interop builder included, which gets a stub
because those images exercise the wire between daemon versions and older refs
do not carry the example. The platform gating was tested rather than reasoned
about: flipping all eleven gates so the native API is excluded leaves the
crate compiling clean across the workspace, every target and the profiling
feature.

The shipped docs tree gains what only the LaTeX manual under design/ had,
which is not published with the daemon. A reference entry covers the whole
surface: addressing and the port tiers, the Berkeley mapping, every method on
FipsAddr, FipsStream, FipsListener and Incoming, the errno table, the
ceilings, the four places data disappears with nothing reported, the line
protocol and the command reference. The errno table gives names rather than
numbers, since the client maps each name onto the libc constant for the
platform it was built for and the supported platforms disagree on the
numbers. A tutorial side trip stands up two throwaway nodes on one machine,
peered over loopback UDP with no TUN and no DNS, then writes a listening
program and a connecting program against them; it needs neither the public
mesh nor root, because the native path is the one that does not go through
the IPv6 adapter. The obligations a client in another language carries are a
how-to of their own, since they are a task rather than a description:
reading the setup connection with recvmsg, associating a descriptor with the
last complete line, telling an empty datagram from a close, and six others.
Serving many peers from one poll loop is another, with the whole program,
because the straightforward listener spawns a thread per flow and that is
wrong at the node's ceiling of 256. The drop causes are a table mapping each
of the seven texts DropReason::as_str produces to the counter it increments,
with drop_oversize called out as the ninth counter that is not in the table.
What a daemon restart costs is a section of its own: every flow and listener
ends, descriptors do not survive, there is no resumption, and datagrams sent
but not yet forwarded are lost through a window nothing bounds.

A stack comparison diagram places the interface against the stack a reader
already knows: the same application over HTTP, TLS, TCP, IP and Ethernet on
one side, and over its own format, FSP, FMP and a FIPS transport on the
other, aligned so each row is one concern. The two columns are not
alternatives and are not drawn as such. An unmodified IPv6 program's packets
reach fips0, and the adapter hands each one to FSP as a payload, so the left
stack runs inside the right one; the left column ends at a fork, eth0 for the
ordinary internet and fips0 for the mesh, and an arrow leaves fips0 and runs
back up into FSP's input. The row where TCP would be is empty on purpose and
names Reliable Object Delivery, which is where that capability is expected to
land. ROD is a v2 capability, the box is dashed because none of it exists
yet, and the design entry says the part a reader needs most: nothing on the
surface anticipates it, so a program written today should assume it does not
exist. Both endpoints carry a scheme and a worked port,
https://<npub>.fips:443 and fips://<npub>:443, with a footnote saying the two
ports are not the same kind of thing, a TCP port inside the tunnel on the
left and an FSP port on the right. The fips:// form is a coinage: nothing in
the tree parses it, nothing registers the scheme, and the API takes a key and
a port as separate arguments rather than a URL. The diagram also says where
the right column stops, since FIPS over UDP still rides IP and Ethernet
beneath. It appears in fips-concepts.md and fips-ipv6-adapter.md, which were
making its argument in prose without a picture, and deliberately not in
fips-architecture.md, which already carries the OSI mapping and makes the
same point about the transport row.

The gateway's control socket moves onto the same bind policy this API uses,
which is the one change here that touches deployed behaviour: fips-gateway
now tightens /run/fips to 0750. That is unreachable under the packaged
deployment, where fips.service has already created the directory at that
mode, and reachable for a source build or a container that starts the gateway
alone.

One changelog entry under Added, describing the released state: what a
client opens and reads, the addressing and why the node address is not it,
the listener being a descriptor, the std::net shape of the Rust surface,
and the one rule Berkeley sockets have no counterpart for. It says in as
many words that the wire is unchanged.

2026-08-21 05:48:23 +00:00

8.7 KiB

Raw Blame History

Write a Native API Client in Another Language

Goal: speak the native datagram API's line protocol directly, from C, Python, Go or anything else, without the Rust client module.

Everything here is something the shipped Rust library already does. It is written out so an author working where there is no such library knows what they are reproducing. Each of these was a real defect before it was a rule, and each fails intermittently rather than outright.

If you are writing Rust, you do not need this guide. Use fips::native::client and see use-the-native-datagram-api.md.

For the protocol itself — the framing, the reply shapes, the commands and their refusals — see ../reference/native-api.md.

Step 1: Read the setup connection with recvmsg, never with a buffered reader

Every read on the setup connection is a recvmsg with an ancillary buffer. A plain read consumes a descriptor-bearing message's bytes with no control buffer, and the kernel then closes the descriptor rather than queueing it. The reply looks perfectly correct and the flow is silently gone.

This applies to both setup commands: a listen reply carries a descriptor as much as a connect reply does. In any language it means no buffered reader, no BufReader, no readline, and no library that wraps the socket in a stream abstraction. Keep the line buffering in your own code, over recvmsg.

Read a listener's own descriptor the same way, for the ancillary data and the close-on-exec flag. One message there is one arrival carrying exactly its own descriptor, so there is nothing to associate.

Step 2: Attach a descriptor to the last complete line of its read

A descriptor belongs to the last complete line of the read that carried it, never to the next line the reader assembles. A recvmsg returning ancillary data ends exactly at the end of the sendmsg that carried it, but it may begin with any amount of data written before it.

A client that sends one command per connection reads one line and cannot hit this. A client that pipelines two setup commands on one connection can: the first reply and the second, descriptor-bearing reply arrive as one read, and a reader that attached the descriptor to the first would hand the flow to the wrong caller.

Two corollaries:

A read that carries a descriptor and completes no line must be reported rather than held. Holding it means guessing which later line it belongs to.
A descriptor that arrives with a line you are going to discard must still be closed, or the flow leaks.

Neither can happen while the daemon writes exactly one whole line per sendmsg and treats a short write as an error, which it does. That is an invariant of two programs, though, not of the socket type.

Step 3: Treat a zero-byte read as end of file only when POLLHUP is set

An empty datagram and a closed peer both produce a zero-byte read, and MSG_EOR does not tell them apart. On Linux 6.8, recvmsg on an AF_UNIX SOCK_SEQPACKET socket returns msg_flags == 0 for a normal message, an empty message and end of file alike, so the flag carries no information.

POLLHUP does discriminate. After a zero-byte read, a queued empty datagram leaves the socket with no events pending, while a closed peer leaves POLLHUP set and latched. Poll with an events mask of zero, because POLLHUP is reported in revents whether or not it was requested. The poll costs nothing: it runs only on the zero-byte path and does not block.

Both directions of the mistake are real. Reading an empty datagram as a close lets a peer tear down a live flow by sending nothing, and presents as a spurious disconnect. Reading a close as an empty datagram leaves the caller spinning on a dead flow.

Note what a close here means: the daemon went away, never a peer finishing.

Step 4: Send with MSG_NOSIGNAL

A datagram written to a flow whose daemon half has gone, or a command written to a daemon that has exited, raises SIGPIPE, whose default disposition kills the process.

Rust ignores the signal at startup, and CPython sets it to SIG_IGN, so a program in either language sees EPIPE. A C or C++ client that has not changed the disposition simply dies. The daemon and the shipped client pass the flag on every send.

Step 5: Set a deadline on the setup socket

Set SO_RCVTIMEO on the setup connection and rewrite the resulting would-block into ETIMEDOUT. The shipped client uses five seconds.

Without it, a daemon that accepted your connection and then stopped answering blocks the setup call forever. This is also what keeps ETIMEDOUT to exactly one producer on the surface, which is what lets a caller read it.

Step 6: Keep descriptor hygiene

Five rules. Each one leaks a flow or loses one when broken.

Request close-on-exec with MSG_CMSG_CLOEXEC on the recvmsg, rather than setting it afterwards. Without it the descriptor survives an exec into a child, the child's reference holds the flow open after this process closes its own, and the flow keeps its slot against the node's ceiling until the child exits.

Walk the whole control buffer, not only the first header. Close extra descriptors rather than dropping them on the floor.

Check for truncation after taking the descriptors, not before. A MSG_CTRUNC test that returns early leaks whatever did arrive.

Lift the descriptor out of an arrival you cannot parse before discarding the message. Refusing a flow is closing its descriptor; discarding the message without taking it leaks the flow instead.

Bound the partial line. A daemon that stopped sending newlines would otherwise grow your buffer without end. The shipped client caps it at 64 KiB, well above any reply.

Step 7: Read the errno name, never the message

The refusal's data.errno is the contract. The message is for an operator reading a log.

A client that matched on English would break on a wording change. The shipped client discards the message entirely so that no caller can come to depend on it, and the daemon's own match over its error types is exhaustive precisely so a new refusal cannot reach a client without a code.

A reply carrying no errno at all should be read as ECONNREFUSED, which covers a daemon older than the field. The errno table is in ../reference/native-api.md.

Step 8: Size the receive buffer, and add no framing

Size every receive buffer at the flow's max_payload, read from the reply and never computed.

SOCK_SEQPACKET truncates a longer datagram, discards the remainder and reports success. It is detectable: recvmsg sets MSG_TRUNC in msg_flags when it dropped part of a message. A plain recv discards msg_flags and so sees none of it, which is where the belief that truncation is silent comes from. Test MSG_TRUNC as well, and a stale or misread max_payload is caught rather than quietly corrupting a payload.

Do not add framing. There is no header and no length prefix in either direction. One send is one datagram.

Step 9: Decide when a flow is over, because nothing else will

Everything above is the library's job. The termination condition is not, in any language.

The v1 wire carries no half-close. Nothing peer-driven closes the daemon's half of a live flow, so a loop that reads until the flow ends does not terminate. Your program decides when a flow is over, or nothing does.

The two shapes that work are a bounded exchange, where the program serves a known number of datagrams per flow and then drops it, and an idle deadline, where the program sets a read timeout and treats its expiry as the end. A server that reads "until the flow closes" holds a thread per peer forever and holds every flow against the node's max_flows ceiling.

Verify it

The repository's test harness drives the line protocol from Python and is the closest thing to a second implementation:

testing/native-api/client.py — a thin RPC client that runs a script of steps over one connection and checks the replies. It implements every rule above.
testing/native-api/control.py — reads show_native_flows back over the control socket while a flow is open.

Check what the node actually holds with fipsctl show native-flows, and read the per-cause drop counters with fipsctl stats metrics under native.

8.7 KiB Raw Blame History