strfry 1.1.3 backs out an epoll teardown guard

What the pointer move contains

strfry does not vendor its WebSocket stack directly. It carries golpe, and golpe carries external/uWebSockets. The tag moves golpe from 72477956 to 73ba63b5, and that commit moves two submodules of its own: external/uWebSockets from 1ef91686 to 45816de8, and external/hoytech-cpp from 722e3dc9 to 1d33d359. The hoytech-cpp bump arrives from an earlier golpe commit dated 28 August and is not part of the revert.

The uWebSockets move is what the changelog line refers to. 45816de8 is a child of 1ef91686, and its message is Revert "Prevent duplicate teardown from stale epoll events".

The guard that came out

The reverted work reached uWebSockets through pull request #6, merged on 20 July 2026. It added two hunks, both short, both carrying their own explanation in a comment.

The first sits in the dispatch loop in src/Epoll.cpp:

if (poll->isClosed()) {
    continue;
}

Its comment states the reason: a poll closed by an earlier callback in the same batch is still present in readyEvents, because its deletion is deferred to loop->closing. Dispatching to it would operate on a closed fd of -1 and run the teardown path a second time.

The second is an early return at the top of WebSocket<isServer>::onEnd in src/WebSocket.cpp:

if (webSocket->isClosed()) {
    return;
}

That comment sets out the sequence in more detail. terminate() can be called from a different poll's callback, and the comment names the slow-client back-pressure cap as a caller that does exactly that. closeSocket() sets the fd to -1 but defers deleting the object, so the stale event still sitting in the same epoll batch re-enters onEnd. Without the guard, disconnectionHandler fires twice and the socket is closed twice, which the comment describes as double-freeing both the user data and the WebSocket itself.

Both hunks are absent from 45816de8, and the two files return to the form they had before the merge.

Which releases carried it

golpe picked the merge up in 3cbace13 on 20 July, and strfry 1.1.1, tagged the following day, points at that golpe commit. 1.1.2 kept it: its own golpe bump moved uWebSockets from 85d00e24, the merge commit, to 1ef91686, a descendant of it. The guard therefore shipped in 1.1.1 and 1.1.2, and 1.1.3 is the first tag in that line without it.

What did not come out

The fix that led 1.1.2 is untouched. That release's uWebSockets bump carried 1ef91686, which added a check of the accumulated fragment buffer against maxPayload and closed a path where a client could keep appending continuation frames, each one under the per-frame limit, and grow relay memory without a ceiling. 45816de8 is built on top of that commit rather than reverting it, so 1.1.3 still has the fragment check. The epoll teardown guard is the only thing removed.

Where it stands

45816de8 is the current head of hoytech/uWebSockets and 73ba63b5 the current head of golpe. Both strfry's master and its release branch point at that golpe commit, and the 1.1.3 tag is the tip of release. The guard is not present anywhere in the published chain, and the changelog line above is the only statement about it in the artifacts.

For an operator, the direction matters: moving from 1.1.1 or 1.1.2 to 1.1.3 removes the guard rather than adding it. The golpe pointer changed, so a build has to update the submodule instead of reusing a cached checkout, the same condition that applied to 1.1.2.