Reproducible Apache mod_http2 CVE-2026-119618 source, HTTP/2, and debugger A/B lab
The first-stage watchpoint result localized the later nulling store. The separate Phase A record covers the bounded free-head/EOS follow-up and its inconclusive schedules.
This private diagnostic seeks the first watched write after the observed
eight-byte C0 copy into B=A+32. The normal chosen-pointer runner is
unchanged unless LAB_WATCHPOINT=1 is set. No address is resolved over
HTTP. The native-literal PID/pointer oracle and all debugger output stay
under ignored output/.
The observer calls chosen_watchpoint_ready() only for an exact pair,
after the C2 table copy and event 63, while the C0 buffer is still held
before stock APR-util's memcpy. GDB launches httpd -X as its own
inferior and stops at this hook. It requests a hardware eight-byte
write watchpoint on B, records the initial write to the native literal
pointer, records the first later observed change to that slot, then
continues with the watchpoint active to an instrumented pre-callback
scan of the same table object and array address, or until the
client timeout. The event
file records classifications and stack symbol names, not addresses.
The full GDB transcript, including the selected instruction, symbol,
and bt 12, remains private. A debugger can alter timing:
only a run whose APR trace confirms exact placement and copy order may
be used to interpret a watchpoint hit.
The debugger stops after a watched store; its reported PC may be the
next instruction. The pre-callback scan sample does not by itself
prove that the push callback consumed the value or returned it.
The calibrated image must provide GDB with Python scripting support. The runner checks this before any traffic and fails closed otherwise. Preflight verifies the debugger script's ready event and the inferior PID, but it cannot prove that the host will grant a hardware watchpoint until an exact pair reaches the hook. The runner therefore also uses a tiny local GDB smoke program and fails before HTTP traffic unless a hardware watchpoint stops on a known eight-byte write in the same container profile. The pinned Apache image does not include GDB. Build a disposable child image from the included Containerfile, recording its base image ID and installed GDB package version. The Apache module and APR-util hashes must still match the pinned profile when the runner starts:
podman build --pull=never \
-f experiments/chosen_pointer/watchpoint/Containerfile \
-t localhost/apache-h2-affected-watch:local .
The disposable container also needs permission for GDB to trace its
child process; run it with --cap-add=SYS_PTRACE and, if the container
seccomp profile blocks ptrace, --security-opt=seccomp=unconfined.
These options apply only to the isolated lab container. No host port is
published and --network none leaves only container loopback.
Build the repository's pinned Apache images and fetch APR as described in the parent experiment. Then, from the repository root, run a no-traffic preflight before any trial:
podman run --rm --pull=never --network none --memory 2g --pids-limit 512 \
--cap-add=SYS_PTRACE --security-opt=seccomp=unconfined \
-e LAB_WATCHPOINT=1 -e LAB_PREFLIGHT_ONLY=1 \
-v "$PWD:/lab:Z" --entrypoint /bin/bash \
localhost/apache-h2-affected-watch:local \
/lab/experiments/chosen_pointer/run_in_container.sh
If preflight succeeds, one bounded affected trial uses the same command
without LAB_PREFLIGHT_ONLY=1; LAB_CONNECTIONS and LAB_PAIRS are
each capped at 8. The runner keeps the client at 40 seconds and the
startup/oracle check at 10 seconds. Put an outer host timeout around the
container command (for example, timeout 180s podman run ...), because
APR compilation precedes client startup.
The private result directory contains:
apr.trace, chosen.ptr, raw/: process and HTTP/2 evidence.gdb.log: raw debugger output, which may contain addresses.watch-smoke.log: raw debugger output from the no-traffic hardware check.watch-events.jsonl: pointer-free hit classifications and short
stack symbol lists.watchpoint-summary.json: pointer-free booleans and counts.The summary is generated by
watchpoint/summarize_watch.py output/RESULT/watch-events.jsonl.
Do not publish the raw debugger log, oracle, or APR trace. Before using
stack symbols in a public note, verify the paired trace, exact write
order, and whether the mutation reflects the same table lifetime. If
GDB reports no hardware watchpoint, a ptrace error, no exact pair, or
a timeout, record the run as an instrumentation or scheduling limit.
The runner returns code 6 for an inconclusive watchpoint run. It
requires a normal debugger exit, a distinct pre-copy value, the
first watched write of the target from the APR-util copy stack, a
later mutation, and a subsequent matching table-and-array
pre-callback sample in the same trace before reporting the internal
writer-before-scan order. client_exit is reported separately; GDB
may intentionally stop the inferior before an HTTP response completes.
A writer-only record is narrower and remains inconclusive for a
selected read.
LAB_WATCH_PHASE_A=1 is an additional opt-in diagnostic; it requires
LAB_WATCHPOINT=1 and leaves the preceding watch mode unchanged. After
the initial C0 target-pointer write, the hardware watchpoint stays armed.
At the subsequent nulling write-watch stop, GDB checks, as booleans
only, that the cleanup-record register equals B and the same pool's
free-cleanup head equals Q. The pinned APR disassembly maps this stop to
the instruction immediately after the free-head update and nulling store.
The pool identity is retained only in debugger memory and private output.
Before resuming, the driver also checks all other stopped thread stacks and
refuses to continue if another thread is already inside either cleanup
registration function beyond the new entry guards.
Before resuming, the driver arms conditional entry breakpoints on both
apr_pool_cleanup_register() and apr_pool_pre_cleanup_register().
GDB automatically continues unrelated calls; the condition checks for a
free-cleanup head equal to Q. The breakpoints remain armed through the
matching pre-callback scan and later stream activity. If a registration
would reuse Q, GDB stops at function entry,
before its free-list load or any possible write to that read-only literal.
Phase A then intentionally kills the still-paused inferior before batch
GDB exits. A positive Phase A result requires a debugger event confirming
the inferior PID was cleared. This intentional termination is not a natural
Apache crash. The no-traffic GDB smoke first stops at a guarded function
entry, confirms its body has not written a marker, kills the paused child,
and checks that the inferior is gone.
On bounded-run teardown, the shell also kills a still-live, verified GDB
child before stopping GDB; this covers a debugger blocked in continue.
The report distinguishes a same-pool EOS h2_bucket_eos_create() stack
from other registrations and distinguishes entries before versus after the
matching scan. It does not execute the candidate dereference or write.
Use the same no-traffic preflight command above with both environment flags, then a bounded affected run with an outer timeout: