
Reproducible Apache mod_http2 CVE-2026-119618 source, HTTP/2, and debugger A/B lab
This is a local A/B lab for the assign_headers() fix in Apache HTTP Server
2.4.71. It compares the released affected 2.4.69 source with 2.4.71,
sends three single-request HTTP/2 cases to a header-echo endpoint, records
a paired GDB trace of the changed table argument, and includes a bounded
Helgrind comparison of the APR allocator race. A separate stress comparison
records an affected-build allocator fault with header bytes in its invalid
pointer argument and a healthy fixed-build control. A subsequent
uninstrumented core identifies the faulting
bucket-free pointer inside the header table's recorded backing array. A
private allocation-history trace records a
tight cross-thread return of that same address geometry. The later
receive trace and core show 1,130 client
HTTP/2 bytes entering an address still recorded as the original table's
backing array. A separate GET/push route reads a damaged original table.
Neither run demonstrates an arbitrary-address primitive or code execution.
A later instrumented inbound-beam run
returned five sampled process-pointer words in HTTP/2 DATA through a lab-only
echo handler. That disclosure depends on a diagnostic APR timing pause; it
also includes a pointer to a static mod_http2 literal in the archived
affected build. It does not establish natural exploit reliability or RCE.
A separate gated reverse-order run
changed one table value-pointer byte and returned the exact predicted suffix
of an attacker-supplied User-Agent in a PUSH_PROMISE. It demonstrates a
constrained relative read, with a matched fixed control and bounded misses.
An eight-byte local-oracle follow-up
sampled a full pointer substitution immediately after the original copy;
the slot was null at a later scan, and no chosen-string response returned.
A hardware-watchpoint follow-up
localized the first later nulling store to APR cleanup registration. The
watched instruction is shared by fresh-record allocation and free-list reuse;
the free-cleanup-head consequence is a conditional hypothesis, not a measured
write through the selected pointer.
A source-gated APR follow-up
resolved that branch in one later exact-overlap trial: a fresh 32-byte
cleanup allocation returned the selected slot in S1's table pool, then filled
it. A repeat placed the C0 copy, two same-pool 32-byte allocations at A and
B, cleanup fill, and null pre-callback scan in one trace. No free-list pop or
chosen-pointer read was observed.
An opt-in 56-byte later-slot trial
placed two selected read-only pointers beyond that cleanup record in an exact
overlap. Both matched immediately after the copy, but a subsequent
h2_headers_create() allocation occupied those slots before the original
table scan. No selected value was returned; the matched fixed control had no
overlap and returned its normal push headers.
A one-shot APR scan observer then
recorded set_push_header returning for original-table entries 0–5 and
starting entry 6 before the affected worker faulted. Entry 6's key slot
overlaps two response-table hash indices. The exact faulting instruction was
not captured in that run, and no selected value was returned. A later
exact-pair GDB stop captured the
libc string-read fault: RDI and the fault address both matched entry 6's
key pointer. The fixed 4×4 control had no exact pair or scan arm. Neither
trial returned a selected value.
A later allocator-order diagnostic
measured why the full-width pointer was lost: when C1 reset the shared APR
node after C2 advanced it, two later allocations occupied the selected
table slots. In one bounded affected run with C1's reset ordered first,
the copied pointer was the actual entry-1 callback argument and a complete
PUSH_PROMISE returned the known mod_http2 literal. Apache aborted after
the response. The pointer came from a private same-process oracle; this is
not an HTTP-only address chain or RCE.
A bounded Phase A follow-up left the
free-head and natural EOS-consumer hypothesis unmeasured under its GDB
schedules; exact-pair misses and an entry-breakpoint storm are recorded.
A same-connection survival study validated
the health-check client on the fixed build, but seven affected runs had no
qualifying disclosure, so post-leak survival remains unmeasured.
Two further trials with the unchanged archived disclosure client also
returned no changed echo; they are separate schedule controls.