Files
moloapp/audit/MOLO-H2-004/README.md
T

39 lines
1.9 KiB
Markdown

# MOLO-H2-004: HTTP/2 control-frame flood, unbounded ack queue
**Scope:** [audit](../README.md) · [README](../../README.md) · [AGENTS](../../AGENTS.md)
CVE-2019-9511/9512/9515/9518, the 2019 Netflix/Google "HTTP/2 DoS" disclosure batch
("Ping Flood" is CVE-2019-9512 specifically); the class, not a MoloApp-specific CVE.
## Source
- [CISA advisory on the 2019 HTTP/2 implementation flaws](https://www.cisa.gov/news-events/alerts/2019/08/13/http2-implementations-vulnerable-denial-service)
## Relevance to MoloApp
Every PING frame a client sends must be answered with a PING ack; a client that keeps
sending PINGs but never reads its socket can make the server queue an unbounded
amount of outbound ack data, if nothing caps either the rate of control frames
processed or the size of the output queue built up for a non-reading peer. This is
the "Ping Flood" shape from the 2019 disclosure batch.
`src/runtime/http/v2/connection.cpp`'s `EnqueueControl` (around line 87) queues an ack
for every PING with no per-connection rate cap on control frames and no ceiling on
`out_queued_bytes_` tied to a slow/non-reading peer (confirmed in the audit); the only
backpressure path found applies to file-backed response bodies, not control frames.
## What the check does
Opens one h2c connection to `build/bin/runtime_interop_server`, sends up to 300,000
PING frames without ever reading the socket, and checks whether the server cuts the
connection (GOAWAY/close) before accepting all of them. PASS means it did; FAIL means
every PING was accepted with no limit on the resulting queue of unread ack bytes.
## Caveats
This measures whether a floor exists on a single connection's queued-but-unread
output, not the server's behavior across many connections at once. Capped at 300,000
frames (about 5 MB of acks) so the check cannot itself exhaust the test host's memory.
Run: `make test` in this directory, or `make -C audit test` from the repo root.