From: Yeonggi Kim <[email protected]>
Hi,
Since 5da704b55 ("MEDIUM: log: Update the log applet to use its own
buffers"), merged in 3.3-dev4, syslog_io_handler() returns early on an
empty input buffer and no longer reaches applet_getchar(), which is
where it noticed the client's shutdown. A TCP syslog session of a
log-forward section whose client closes after its last message was
consumed therefore stays open. A log-forward section does not inherit
"timeout client" from the defaults section, so without one set in the
section, the session never ends.
Impact: each such client leaves a session behind, and a soft-stop then
waits for "hard-stop-after" (forever without it). We hit it on 3.4 with
a "ring" forwarding to a log-forward section of the same process: on
reload, the ring's forwarder closes its connection, the log-forward
session stays, and the old process is only stopped by
"hard-stop-after".
A reproducer is in the commit message of 1/2.
The series:
- 1/2 lets an empty buffer reach applet_getchar() again, and only
leaves early when no input buffer is available, like the dns and
peers applets do.
- 2/2 adds reg-tests/log/log_forward_tcp_close.vtc. It fails (timeout)
without 1/2 on master bd884a2, 3.4.6 and 3.3.16, and passes with
it, as well as on 3.2.25, which is not affected. The other
reg-tests/log tests and reg-tests/ssl/log_forward_ssl.vtc pass with
and without the series.
Both should be backported as far as 3.3.
Thanks,
Yeonggi Kim
Yeonggi Kim (2):
BUG/MEDIUM: log: do not ignore the peer's shutdown in the syslog
applet
REGTESTS: log: check that a log-forward TCP session ends on client
close
reg-tests/log/log_forward_tcp_close.vtc | 60 +++++++++++++++++++++++++
src/log.c | 7 ++-
2 files changed, 65 insertions(+), 2 deletions(-)
create mode 100644 reg-tests/log/log_forward_tcp_close.vtc
--
2.52.0
