On Tue, Nov 17, 2026 at 10:08:31PM +0100, Willy Tarreau wrote:
> Hi Daniel,
>
> On Tue, Nov 17, 2026 at 12:30:01PM -0500, Daniel Corbett wrote:
> > Hi,
> >
> > Since 2aed0c9be ("BUG/MINOR: h3: fix missing EOM after trailers without
> > FIN"), an HTTP/3 client can crash haproxy on master by sending a
> > request with a body and trailers and never closing the stream. It isn't
> > in any stable branch yet, but it is tagged for backport up to 2.8, so
> > it would be good to have this fixed before it goes further. The patch
> > is attached, and its commit message has the full analysis.
> >
> > On a stock build of master (68a8abcbe):
> >
> > FATAL: bug condition "blk == ((void *)0)" matched at src/htx.c:937
> > call trace(15):
> > | 0x706884 <42 90 00 e8 ac 76 f4 ff]: htx_xfer+0x634/0x668 >
> > complain_with_dbg
> > | 0x52dd31 <4c 89 ef e8 1f 85 1d 00]:
> > qcs_http_rcv_buf+0x221/0x484 > htx_xfer
> > | 0x4c6d41 <48 89 df e8 cf 6d 06 00]: qcm_strm_rcv_buf+0x51/0x6b3
> > > qcs_http_rcv_buf
> > | 0x63ca3a <d9 0f 29 45 a0 ff 50 10]: sc_conn_recv+0x47a/0xae1
> > | 0x640503 <89 75 e8 e8 bd c0 ff ff]: sc_conn_io_cb+0x93/0xf2 >
> > sc_conn_recv
> >
> > The patch keeps the unterminated trailers in the QCS Rx buffer until the
> > FIN adds the EOT. I picked that over the flag approach mentioned in the
> > commit message because the HTX tail already holds that state, but I'm
> > fine either way if you'd rather have it explicit.
> >
> > Bisect and versions, stock builds (make TARGET=linux-glibc
> > USE_OPENSSL=1 USE_QUIC=1), with the attached reproducer:
> >
> > 68a8abcbe (master) died 10 of 10
> > 2aed0c9be died 3 of 3
> > 14fa727d4 (its parent) died 0 of 5
> > 3.4.7 died 0 of 5
> > 68a8abcbe, no buffer-request died 2 of 10
> > 68a8abcbe + patch died 0 of 10, also 0 of 10 without
> > buffer-request
> >
> > With the patch, trailers without FIN followed by a late FIN, a
> > RESET_STREAM or a CONNECTION_CLOSE, with and without "option
> > http-buffer-request", gave no crash in 36 runs (18 stock, 18 ASAN with
> > no ASAN report). With the late FIN, the backend got the body first and
> > the trailers once the FIN arrived. reg-tests on the ASAN build give the
> > same result with and without the patch (357 passed, 4 failed in
> > ssl_client_auth / ssl_client_samples on this host's crypto policy).
> >
> > Reproducer, attached as plain text (repro.cfg, repro.py, run.sh). It
> > needs openssl and python3 with aioquic. run.sh makes a throwaway cert
> > and starts haproxy, then repro.py does a GET (baseline), the POST
> > without FIN, and another GET:
> >
> > HAPROXY=/path/to/haproxy ./run.sh
> >
> > baseline: GET answered
> > sent: POST with HEADERS + DATA + trailers HEADERS, no FIN
> > after: GET not answered, haproxy is gone
> > haproxy: died, wait status 132, FATAL: bug condition "blk == ((void *)0)"
> > matched at src/htx.c:937
> >
> > I found this with AI-assisted tooling. The reproducer and the patch
> > were tested on real builds as described.
>
> Thanks for the report, we're about to issue stable releases hopefully
> tomorrow if the storm of AI reports stops for a while, so we'll try to
> address that one before.
So for now we'll revert the minor fix. It probably uncovers some other
unexpected invariants that need to be revalidated across other muxes,
such as the relevance today of this BUG_ON() in htx.c (maybe the null
block can be ignored or skipped, I don't know). The initial issue is
not a big deal and there's no reason the reasonably simple fix for it
makes everything else more brittle.
Thanks again for the report!
Willy