On Tue, Jul 24, 2018 at 03:35:51PM +0100, Daniel P. Berrangé wrote: > On Tue, Jul 24, 2018 at 03:25:04PM +0100, Stefan Hajnoczi wrote: > > The simple trace backend spawns a write-out thread which is used to > > asynchronously flush the in-memory ring buffer to disk. > > > > fork(2) does not clone all threads, only the thread that invoked > > fork(2). As a result there is no write-out thread in the child process! > > > > This causes a hang during shutdown when atexit(3) handler installed by > > the simple trace backend waits for the non-existent write-out thread. > > > > This patch uses pthread_atfork(3) to terminate the write-out thread > > before fork and restart it in both the parent and child after fork. > > This solves a hang in qemu-iotests 147 due to qemu-nbd --fork usage. > > I'm not convinced this is safe, as it looks like it has a window in > which both the parent and child processes will be doing write-out to > the same file. > > In particular in the main QEMU system emulators it means that any > time we fork() in QEMU, eg for spawning commands with migration > exec: URI, or TAP devuce ifup scripts, etc, we'll be starting a > write-out thread in the child.
I'd be more inclined to have the pthread_atfork() handle simply terminate the tracing process, reversing all effects of trace_init_backends(). Then after qemu-nbd has called fork(), it can simply call trace_init_backends() explicitly to start it running again. This avoids unecessarily starting tracing in child processes that are not requiring/expecting it. Regards, Daniel -- |: https://berrange.com -o- https://www.flickr.com/photos/dberrange :| |: https://libvirt.org -o- https://fstop138.berrange.com :| |: https://entangle-photo.org -o- https://www.instagram.com/dberrange :|
