On 9/21/26 17:25, Denis V. Lunev via Devel wrote:
> daemonStreamEvent() holds priv->lock across its whole body and closes
> the client from every error path. The close hook runs in the same
> thread and takes priv->lock again, so the daemon main loop deadlocks
> against itself and the host stops answering RPC entirely while its
> guests keep running. Only a restart of the daemon recovers.
>
> We hit this in the field on 9.5.0, and current master is affected the
> same way. Getting there needs a client already marked wantClose while
> one of its streams still delivers events, which is presumably why it
> has gone unnoticed since 9.2.0.
>
> Patch 1 closes the client after the lock guard has gone out of scope.
> Patch 2 adds a test for it, which needs remote_daemon_stream.c compiled
> into a test binary; that is the only build change and it pulls in the
> generated protocol headers via remote_daemon.h. The test deadlocks
> without patch 1 and passes with it.
>
> Signed-off-by: Denis V. Lunev <den@openvz.org>
>
> Denis V. Lunev (2):
> remote: fix daemon deadlock when a stream event closes its client
> tests: add daemonstreamtest covering the stream close deadlock
>
> src/remote/meson.build | 7 +-
> src/remote/remote_daemon_stream.c | 58 ++++----
> tests/daemonstreamtest.c | 212 ++++++++++++++++++++++++++++++
> tests/meson.build | 9 ++
> 4 files changed, 262 insertions(+), 24 deletions(-)
> create mode 100644 tests/daemonstreamtest.c
>
>
> base-commit: 91d4618734bd6d02174b909aadb7d84a99560ebe
Reviewed-by: Michal Privoznik <mprivozn@redhat.com>
and merged.
Michal