[MDA2AV]

All eleven published packages share one version, as always. The bump is only their Version elements; IoxideRuntime.Version is generated from ioxide.csproj. ioxide.redis.stackexchange is bumped too, so a manual pack matches, but CI does not pack it.

The minor goes to 17 for the breaks listed below: ioxide.redis is gone, and one core signature changed.

Carries everything since 0.16.281:

• Reactor:
  • reactor: a timer queued before the op table grows keeps its deadline, and the table starts at a configurable 4096 #297: a timer queued before the op table grows keeps its deadline, and the table starts at a configurable 4096.
  • reactor: a post from another thread is never left queued behind a wake that was already spent #321: a post from another thread is never left queued behind a wake that was already spent.
  • reactor: OnFault's doc says it runs before the teardown, which is when it does #323: OnFault's doc says it runs before the teardown, which is when it does.
  • reactor, clients: the wait asks only for what is left until the next deadline, and the clock is read through the vDSO #329: the wait asks only for what is left until the next deadline, and the clock is read through the vDSO.
• TCP:
  • tcp: an accept that runs out of file descriptors waits for the next tick instead of spinning #287: an accept that runs out of file descriptors waits for the next tick instead of spinning (tcp: accept re-arms immediately on EMFILE/ENFILE, so fd exhaustion pins every reactor at 100% CPU and logs millions of lines per second #222).
  • tcp: accepted sockets take TCP_NODELAY from the listener instead of a setsockopt per accept #288: accepted sockets take TCP_NODELAY from the listener instead of a setsockopt per accept.
  • tcp: an incremental recv ended by a full completion queue keeps its part-filled buffer instead of recycling it #298: an incremental recv ended by a full completion queue keeps its part-filled buffer instead of recycling it.
  • tcp: an incremental accept short of locked memory for its buffer ring sheds the connection instead of ending the reactor #299: an incremental accept short of locked memory for its buffer ring sheds the connection instead of ending the reactor.
  • tcp: a recv-queue overflow reaches a Stream reader as an error, not a clean end #304: a recv-queue overflow reaches a Stream reader as an error, not a clean end.
  • tcp: a flush from another thread that races the connection's close completes #305: a flush from another thread that races the connection's close completes.
  • tcp: a write the slab cannot grow to fails instead of spinning the reactor #306: a write the slab cannot grow to fails instead of spinning the reactor.
  • tcp: a handler that lets go and then throws is released once, so teardown never closes a descriptor it gave back (#242) #334: a handler that lets go and then throws is released once, so teardown never closes a descriptor it gave back (Releasing an fd number mid-run kills an unrelated live connection #242).
• Core and files: core, file: every descriptor ioxide opens is close-on-exec, so a child process inherits none #290: every descriptor ioxide opens is close-on-exec, so a child process inherits none.
• TLS:
  • tls: a flush or Complete from another thread encrypts on the reactor, and a pipe flush allocates nothing #285: a flush or Complete from another thread encrypts on the reactor, and a pipe flush allocates nothing (tls: a flush from a thread other than the reactor runs SSL_write concurrently with the decrypt pump's SSL_read and breaks the session #249).
  • tls: a refused handshake reports its own reason, not another connection's or none #319: a refused handshake reports its own reason, not another connection's or none.
  • tls: a finished handshake does not keep its connection reachable behind a stalled one #320: a finished handshake does not keep its connection reachable behind a stalled one.
• HTTP/2 (pure C#):
  • http2: an HPACK table size update past the 4096 the server allows is a COMPRESSION_ERROR #308: an HPACK table size update past the 4096 the server allows is a COMPRESSION_ERROR.
  • http2: a frame inside an open header block, or a CONTINUATION outside one, is a PROTOCOL_ERROR #309: a frame inside an open header block, or a CONTINUATION outside one, is a PROTOCOL_ERROR.
  • http2: a WINDOW_UPDATE past 2^31-1 is a FLOW_CONTROL_ERROR, not a window that wraps negative #310: a WINDOW_UPDATE past 2^31-1 is a FLOW_CONTROL_ERROR, not a window that wraps negative.
  • http2: DATA on stream 0 is a PROTOCOL_ERROR #311: DATA on stream 0 is a PROTOCOL_ERROR.
  • http2: a reset stream or a dropped connection no longer hands a running streamed handler's request and chunk back to the pool #331: a reset stream or a dropped connection no longer hands a running streamed handler's request and chunk back to the pool.
  • http2: a streamed request's trailers are decoded and dropped, never served as a second request nor written into the running handler's request #332: a streamed request's trailers are decoded and dropped, never served as a second request nor written into the running handler's request.
  • http2: a streamed response waiting on credit no longer hangs its handler when the client closes the connection cleanly #333: a streamed response waiting on credit no longer hangs its handler when the client closes the connection cleanly.
• nghttp2:
  • nghttp2: a request body past MaxRequestBytes is reset with ENHANCE_YOUR_CALM instead of reaching the handler #295: a request body past MaxRequestBytes is reset with ENHANCE_YOUR_CALM instead of reaching the handler.
  • nghttp2: request bodies can stream through Nghttp2Request.BodyReader, credited as the handler reads #330: request bodies can stream through Nghttp2Request.BodyReader, credited as the handler reads.
• HTTP/3 and QUIC:
  • quic: a Handshake packet that lands on another reactor is forwarded to the one that owns it #282: a Handshake packet that lands on another reactor is forwarded to the one that owns it.
  • nghttp3: a late answer to a cancelled request is dropped instead of closing the connection #291: nghttp3 drops a late answer to a cancelled request instead of closing the connection.
  • quic: a released claim cancels its receive, so its socket closes and the address can connect again #293: a released claim cancels its receive, so its socket closes and the address can connect again.
  • quic: a pipe flush waits at the send-retention high-water, and reports IsCompleted once the connection is gone #294: a pipe flush waits at the send-retention high-water, and reports IsCompleted once the connection is gone.
  • http3, nghttp3: an upload a streaming handler never reads gives its connection credit back #296: an upload a streaming handler never reads gives its connection credit back, on ioxide.http3 and nghttp3.
  • quic: credit opened outside an engine cycle is flushed, so a paced upload does not stall #324: credit opened outside an engine cycle is flushed, so a paced upload does not stall.
  • quic: a connection accepted on a migrated peer's pin socket sends from the serving socket #325: a connection accepted on a migrated peer's pin socket sends from the serving socket.
  • quic: a junk long-header datagram is refused before ~75 KB of connection state is built #326: a junk long-header datagram is refused before ~75 KB of connection state is built.
  • quic: a send on a stream the peer already finished no longer leaks send retention #327: a send on a stream the peer already finished no longer leaks send retention.
  • quic: a client socket's ephemeral port is never handed to another socket #328: a client socket's ephemeral port is never handed to another socket.
• HTTP client:
  • httpclient: requests by URL, names and IPv6, TLS by host, connect, idle and lifetime limits, and a pool that waits on the reactor #300: requests by URL, names and IPv6, TLS by host, connect, idle and lifetime limits, and a pool that waits on the reactor.
  • httpclient: a malformed h1 chunk size fails its response instead of being read as another size #301: a malformed h1 chunk size fails its response instead of being read as another size.
  • httpclient: bytes past an h1 response's end retire its connection instead of being read as the next response #302: bytes past an h1 response's end retire its connection instead of being read as the next response.
  • httpclient: a CR, LF or NUL in an h1 request's head is refused instead of written onto the connection #303: a CR, LF or NUL in an h1 request's head is refused instead of written onto the connection.
• Postgres:
  • pg: a timed-out connection is shut down, and freed only once its armed recv has returned #292: a timed-out connection is shut down, and freed only once its armed recv has returned.
  • pg: a statement whose Bind or Execute fails stays cached, because its Parse succeeded #312: a statement whose Bind or Execute fails stays cached, because its Parse succeeded.
  • pg: the timeout sweep removes the connection that timed out, not whatever its index holds afterwards #313: the timeout sweep removes the connection that timed out, not whatever its index holds afterwards.
  • pg: a backend message longer than MaxReceiveBytes fails the connection, and its length can no longer wrap #315: a backend message longer than MaxReceiveBytes fails the connection, and its length can no longer wrap.
  • pg: a SCRAM iteration count above 1,000,000 is refused instead of stalling the reactor #316: a SCRAM iteration count above 1,000,000 is refused instead of stalling the reactor.
  • pg: a connect and its startup end at CommandTimeoutMs instead of waiting forever #318: a connect and its startup end at CommandTimeoutMs instead of waiting forever.
• Redis:
  • redis.stackexchange: StackExchange.Redis with its connections on the reactor's ring (experimental) #335: ioxide.redis.stackexchange, StackExchange.Redis with its connections on the reactor's ring. It is experimental and not published.
  • redis: drop ioxide.redis #337: ioxide.redis is dropped.
  • redis.stackexchange: the transport in one file per concern, receives into native memory, and a bench against StackExchange.Redis's own sockets #338: the transport is split into one file per concern, receives land in native memory, and bench/redis-matrix.sh runs it against StackExchange.Redis's own sockets.
• Bench and samples:
  • bench: wait for every reactor's socket before loading, and run the sample's own binary #283: the bench waits for every reactor's socket before loading, and runs the sample's own binary.
  • samples: the nghttp3 streamed-response sample reuses its two responses #284: the nghttp3 streamed-response sample reuses its two responses.
  • bench: a cell measured while the conntrack table was full is discarded #289: a cell measured while the conntrack table was full is discarded.

Breaking

• Removed package: ioxide.redis has no 0.17 version; its last is 0.16.281 (redis: drop ioxide.redis #337). StackExchange.Redis over ioxide (ioxide.redis.stackexchange, redis.stackexchange: StackExchange.Redis with its connections on the reactor's ring (experimental) #335) takes its place, unpublished for now.
• Signature: RingSocket.CreateTcp(IRingHost) is now CreateTcp(IRingHost, AddressFamily family = AddressFamily.InterNetwork) (httpclient: requests by URL, names and IPv6, TLS by host, connect, idle and lifetime limits, and a pool that waits on the reactor #300). Calls still compile; code built against 0.16.281 has to be recompiled.
• Behaviour:
  • httpclient: a CR, LF or NUL in an h1 request's head is refused instead of written onto the connection #303: an h1 request whose head carries a CR, LF or NUL now throws instead of being sent.
  • core, file: every descriptor ioxide opens is close-on-exec, so a child process inherits none #290: a child process no longer inherits ioxide's descriptors.

Package validation against the 0.16.281 packages on NuGet (EnablePackageValidation, baseline 0.16.281) reports the CreateTcp change as the only API break in the eleven packages.

Native

• libioxide_nghttp2: rebuilt for nghttp2: request bodies can stream through Nghttp2Request.BodyReader, credited as the handler reads #330 with scripts/build-nghttp2-native.sh: ih2_server_new takes a manual_flow flag, and ih2_consume is new. The package ships the library; consumers rebuild nothing.

On this branch's tree (main at b647653, plus the bump), all eight suites pass with none skipped: E2E 285, Unit 78, Http 83, Tls 174, Chaos 79, File 5, Pg 18 and Redis 6. The tls module was loaded, so the kTLS tests ran.

Some tests stay pending. Each one pins a known gap and fails the run once the gap is fixed: E2E 6, Unit 1, Http 4, Tls 8, Chaos 1, and Redis 1 (#336).