Hi,
While testing QMux, I found that a client which does not read the
response until the socket buffer is full crashes haproxy, as
qcc_qmux_send_frames() does not support partial writes.
I gave a try at a fix: the first patch keeps the unsent part of the
record in the Tx buffer and retries later, the second one keeps the
connection open until this data is flushed so that the end of the
response is not lost. I am well aware this may not be the right
approach, so any feedback is welcome and I will adapt the patches as
needed.
Reproducer, with any HTTP server on 127.0.0.1:8000 serving a 4MB file
as /big.
haproxy.cfg:
global
expose-experimental-directives
defaults
mode http
timeout client 30s
timeout server 30s
timeout connect 5s
frontend fe
bind :8443 proto qmux
default_backend be
backend be
server s1 127.0.0.1:8000
repro.py:
import socket, sys, time
def vi(v): # QUIC varint
for n, p in ((1, 0), (2, 0x4000), (4, 0x80000000), (8,
0xC000000000000000)):
if v < 1 << (8 * n - 2):
return (v | p).to_bytes(n, "big")
def rec(b): # QMux record
return vi(len(b)) + b
tp = b"".join(vi(i) + vi(len(vi(v))) + vi(v)
for i, v in ((4, 1 << 40), (5, 1 << 40), (6, 1 << 40),
(7, 1 << 40), (8, 100), (9, 100)))
qpack = b"\x00\x00\xd1\xd6\x50\x01x\x51\x04/big" # GET http://x/big
hdrs = vi(1) + vi(len(qpack)) + qpack
s = socket.create_connection((sys.argv[1], int(sys.argv[2])))
s.sendall(rec(vi(0x3F5153300D0A0D0A) + vi(len(tp)) + tp) +
rec(vi(0x0b) + vi(0) + vi(len(hdrs)) + hdrs))
time.sleep(5) # never read the response
$ python3 repro.py 127.0.0.1 8443
FATAL: bug condition "sent != b_data(buf)" matched at src/qcm_qmux.c:385
Both patches should be backported to 3.4.
William
William Dauchy (2):
BUG/MEDIUM: qmux: retry on partial write instead of aborting
BUG/MINOR: qmux: do not release connection with pending Tx record data
include/haproxy/qcm_qmux.h | 9 ++++++++
src/mux_quic.c | 40 +++++++++++++++++++++++-------------
src/qcm_qmux.c | 42 +++++++++++++++++++-------------------
3 files changed, 56 insertions(+), 35 deletions(-)
--
2.56.1