What I found
MsQuic walks decoded ACK ranges from low packet number to high. A syntactically valid multi-range ACK can acknowledge a real outstanding packet in a lower range, then include a deliberately skipped packet number in a later range.
The later range correctly raises PROTOCOL_VIOLATION. The early return skips completion and pool-return of metadata that the lower range already spliced off the connection-owned lists. Shutdown cannot find those nodes. Repeating the sequence across connections retains process memory.
This is a medium-severity authenticated remote DoS. Closing the attacker’s connection does not reclaim the leak.
flowchart TD
ACK["Authenticated multi-range ACK"] --> Low["Low range: real outstanding PN"]
ACK --> High["High range: skipped PN"]
Low --> Detach["Splice metadata into local AckedPackets"]
High --> Err["PROTOCOL_VIOLATION and return"]
Detach --> Leak["AckedPackets never returned to the pool"]
Err --> Leak
Leak --> Shut["Uninitialize walks SentPackets / LostPackets only"]
The ownership bug
QuicConnRecvFrames accepts the ACK. Ranges are stored low-to-high in DecodedAckRanges. QuicLossDetectionProcessAckBlocks iterates that order.
The skipped-packet check runs inside the same loop that mutates ownership. For the lower real range, matching nodes move from SentPackets / LostPackets onto a stack-local AckedPackets list. When a later range covers Connection->Send.SkippedPacketNumber, the function calls QuicConnTransportError(..., QUIC_ERROR_PROTOCOL_VIOLATION) and returns.
The return sits before the loop that completes acknowledgements and returns metadata to the pool. QuicLossDetectionUninitialize only walks the member lists. Frame-owned stream or datagram references on those nodes stay held too.
The failed invariant: a rejecting validation must dominate ownership changes, or the error path must release every node already detached.
sequenceDiagram
participant Client as 1-RTT peer
participant Loss as Loss detection
participant Pool as Sent-packet pool
Client->>Loss: ACK range real PN
Loss->>Loss: Detach metadata from SentPackets
Client->>Loss: ACK range skipped PN
Loss->>Client: PROTOCOL_VIOLATION, close
Note over Loss,Pool: Detached nodes never returned
Attack shape
- Complete a normal QUIC/TLS handshake.
- Pace responses so at least one lower sent or lost packet stays outstanding while the server sends across a scheduled packet-number skip.
- Send a valid encrypted ACK whose low range covers the real packet and whose later range covers the skipped number.
- The connection closes. The detached allocation remains in the shared process.
- Open a fresh connection and repeat.
QUIC packet numbers and keys are per-connection, so a second independent client cannot ACK the first connection’s packets. One authenticated peer is the whole attacker.
A positive E2E run against an unmodified server retained one 120-byte metadata allocation: gets 81 / returns 80 after shutdown, versus a balanced no-attack control.
Impact
Retained memory after N triggers is the sum of each detached metadata allocation plus up to 12 frame references per packet. Per connection the yield is bounded by outstanding sent/lost state. Across attacker-created connections it accumulates until process restart.
No reviewed path gives an OOB, write primitive, control-flow influence, or privilege-boundary cross. RCE and confidentiality/integrity impact are not supported. The cost of a completed handshake plus enough server output to cross a skip is why this stays Medium.
Fix
Validate every ACK range, including skipped-PN rejection, before splicing metadata — or on the error path walk AckedPackets and return every node. Treating an ACK of an unsent packet as a protocol violation is correct (RFC 9000). The defect is the non-transactional placement of that check.
What this is not
Not “rejecting a skipped PN is the bug.” That rejection is intentional. The bug is leaking the earlier range’s detached objects when the later range fails. Connection close is not cleanup for process-lifetime pool nodes.