The stress harness found this: 38 completed uploads logged as errors, and a
run that took 30s to move what should take 6.
After the last byte we waited up to 30 seconds for the downloader to close
the file connection. Soulseek does leave that close to the downloader, but a
peer that simply leaves the socket open — or whose FIN a NAT swallows — is
not a failure: it has every byte. We were marking those uploads Failed,
skipping the SendUploadSpeed report that sets our advertised speed, and
holding the upload slot the next peer was queued for.
Cutting the wait to five seconds was not enough, and stress said so: with
two slots and sixty-four queued leechers, five seconds apiece is a hundred
and sixty seconds of waiting, and ten uploads never got served inside the
run. The slot has to be free the moment the file is delivered, so the wait
is gone — Nicotine+ finishes the transfer there too and closes from its own
side. That profile went from 146s with ten uploads lost to 9.7s with none;
the default profile from 30.7s and 25 MiB/s to 5.8s and 133 MiB/s.
Also: a decompression ceiling on peer payloads, the numbers Nicotine+ uses
(128 MiB for a search or folder reply, 2 GiB for a browse listing). Deflate
reaches roughly 1000:1, so a few kilobytes from an unknown peer could
otherwise expand until the process died.
And the stress harness counted threads with `ps -M`, which is macOS
spelling: on Linux it returns one unrelated line, so the guard against
thread explosions has been reporting a single thread on CI all along.
/proc/self/status has the real number.
The client read the distributed network but never carried it: a peer
dialling with a D connection was logged and dropped, and AcceptChildren
was hardcoded false, so this client took from the tree without ever
relaying a search for anyone else.
With accept_children set, peers hang from us: each is told where our
branch sits when it connects and again when the branch moves, every
search that reaches us is passed down, and a child that goes away loses
its slot. Ten at a time, matching Nicotine+, since each costs a socket
and a copy of the network's whole search stream — which is also why it is
opt-in, through the new accept_children config key (ignored without a
listener, since a child would have nothing to dial).
A browse listed each file as a name and a size, so a searcher who saw a
bitrate in a search reply lost it the moment they opened the sharer's
files. Listings (codes 5 and 37) now carry each file's attributes the
way Nicotine+ sends them, as a SharedFileEntry with name, size and
attributes in place of the (name, size) pair; the CLI's browse record
gains bitrate and duration, and the daemon protocol moves to version 2
for the new listing shape.
The mock seeder drained every message but the TransferResponse while it
waited for the client to allow an upload, so a browse arriving in that
window was silently dropped and counted against the client.
Every ceiling this project has hit only appeared under a mixed load, and
three of the four lost work silently while single-feature e2e stayed
green, yet the benchmark only ever ran by hand on one laptop. With
STRESS_REQUIRE_FUNCTIONAL=1 the stress example exits 1 when a download,
upload, search or browse was lost; throughput never fails it because it
measures the machine, not the client. CI now runs a quarter-size profile
against soulfind on every pull request.
soulfind has reported the host's LAN address for a loopback login since
1eab2e8 (November 2025), so the client dials the mocks there and a
listener bound to 127.0.0.1 refuses it: every upload and most downloads
failed against a current soulfind while the e2e suite, whose mocks
already bind all interfaces, stayed green.
Every reply to a peer's search claimed a free upload slot, no speed and
an empty queue, whatever the sharer was doing, and searchers rank
sources on exactly those fields. A reply now says whether a slot is
free, the bytes per second of the last completed upload, and how many
uploads are waiting.
Dispatch read one byte of the code, so a code with a non-zero high byte
was handled as its low byte. Message::get_message_code now reads the
whole little-endian uint32 and handlers register by u32. PeerInit and
PierceFirewall, the only frames with a one-byte code, get their own
accessor for the listener's first frame.
Closes#54.
A download can now be cancelled whether it is queued, in progress, or
paused. Cancelled is a terminal status: the store refuses to overwrite
it, the status channel relays nothing after it, the partial file is
deleted, and the transfer thread releases its socket at its next
one-second poll instead of the 30s read deadline. A queued download that
is cancelled declines the peer's later offer with TransferResponse
"Cancelled", and an inbound file connection for it is refused. A fresh
download of the same file evicts the dead row so it is not shadowed,
and a cancelled transfer is not re-queued after a restart.
Surfaces: Client::cancel_download, download.cancel over RPC, `transfer
cancel <user> <path>` on the CLI (which also stops an upload by that
name), and x in the transfers pane, where d dismisses the row.
The slot cap merged from master defaults to 2 uploads, and the CLI capped
downloads at 5. Those are the numbers you pick when your own bandwidth is the
scarce resource, but a Soulseek transfer is paced by the peer on the other end
— usually a few hundred KiB/s — so what fills a modern link is many transfers
at once, not faster ones. Upload slots go to 10, downloads to 20. The queue
itself stays: privileged users being able to jump it is a condition of being a
tolerated third-party client.
serve_file also slept half a second after the last byte so the downloader could
drain before the socket closed. With a cap that is time no other peer can use,
and 64 uploads over 10 slots paid it six times over. A half-close says "that is
the whole file" and TCP still delivers everything written before the FIN.
Measured with 64 waiting peers: 2 slots 24.2s, 8 slots 6.2s, 32 slots 3.1s; at
10 slots the linger alone was 5.4s to 3.05s.
The client had never been driven at scale. A stress harness that mixes
searches, downloads, uploads and browses against a local soulfind found four
ceilings, three of which lose work silently rather than merely being slow.
Peers were capped at one per core. ActorSystem ran each actor's whole lifetime
as one job on a fixed-size ThreadPool, and actor loops block until stopped, so
the pool acted as a semaphore: past the cap an actor never started, never
connected and never errored, and uploads stopped working entirely. The pool had
exactly one consumer and every job it received was non-terminating, so it is
gone; actors get their own thread.
TransferResponse raced the bookkeeping it depends on. It was sent straight back
to the peer while UpdateDownloadTokens was still queued for the client loop —
but the response is what invites the peer's file connection, and that connection
is matched by token. Under load the client loop lost, the file connection
matched nothing, and the download sat on Queued forever. The response now goes
out after the token is recorded.
Download tokens were md5(filename) truncated to 20 bits, and the store deletes
every entry matching a token, so two downloads sharing one destroyed each other
— which fetching the same filename from two peers does every time. They now
come from a counter.
Port 0 from GetPeerAddress means the server does not know where a user listens.
It was cached and dialled, and the upload waiting on it was dropped. It is no
longer cached or served against; the connect attempt stays, because its failure
is what makes an unreachable peer fall back to server brokering.
Mixed profile at 1024 peers: 512/512 downloads, 256/256 uploads, 400-850 MiB/s
in ~7s, against 71/96 downloads at 7.9 MiB/s and a 60s timeout before.
Michel de Graaf