[libvirt PATCH v2 0/7] qemu: introduce a per-VM event loop thread

[Daniel P. Berrangé]

On Fri, Mar 06, 2020 at 02:42:45PM +0100, Michal Prívozník wrote:
> On 5. 3. 2020 13:51, Daniel P. Berrangé wrote:
> > This series changes the way we manage the QEMU monitor and
> > QEMU agent, such that all I/O is processed by a dedicated
> > event loop thread.
> > 
> > Many times in the past years people are reported issues
> > where long running monitor event callbacks block the main
> > libvirtd event loop for an unacceptably long period of
> > time. In the best case, this delays other work being
> > completed, but in bad cases it leads to mgmt app failures
> > when keepalive times trigger a client disconnect.
> > 
> > With this series, when we spawn QEMU, we also spawn a
> > dedicated thread running a GMainLoop instance. Then QEMU
> > monitor and QEMU agent UNIX sockets are switched to use
> > GMainContext for events instead of the traditional libvirt
> > event loop APIs. We kill off the event thread when we see
> > EOF on the QEMU monitor during shutdown.
> > 
> > The cost of this approach is one extra thread per VM,
> > which incurs a new OS process and a new stack allocation.
> > 
> > The QEMU driver already delegates some QMP event handling
> > to a thread pool for certain types of event. This was a
> > previous hack to mitigate the impact on the main event
> > loop. It is likely that we can remove this thread pool
> > from the QEMU driver & rely on the per-VM event threads
> > to do all the work. This will, however, require careful
> > analysis of each handler we pushed into the thread pool
> > to make sure its work doesn't have a dependency on the
> > event loop running in parallel.
> > 
> > This is one step towards eliminating the need to have the
> > libvirt event loop registered when using the embedded QEMU
> > driver. A further step is using a thread to dispatch the
> > lifecycle events, since that currently relies on a zero
> > second timer being registered with the event loop.
> > 
> > Changed in v2:
> > 
> >  - Fixed race accessing free'd memory causing crash
> >  - Fixed unused variables
> >  - Merged first acked patches
> > 
> > Daniel P. Berrangé (7):
> >   src: introduce an abstraction for running event loops
> >   qemu: start/stop an event loop thread for domains
> >   qemu: start/stop an event thread for QMP probing
> >   tests: start/stop an event thread for QEMU monitor/agent tests
> >   qemu: convert monitor to use the per-VM event loop
> >   qemu: fix variable naming in agent code
> >   qemu: convert agent to use the per-VM event loop
> > 
> >  po/POTFILES.in               |   1 +
> >  src/libvirt_private.syms     |   5 +
> >  src/qemu/qemu_agent.c        | 600 ++++++++++++++++++-----------------
> >  src/qemu/qemu_agent.h        |   1 +
> >  src/qemu/qemu_domain.c       |  33 ++
> >  src/qemu/qemu_domain.h       |   6 +
> >  src/qemu/qemu_monitor.c      | 145 ++++-----
> >  src/qemu/qemu_monitor.h      |   3 +-
> >  src/qemu/qemu_process.c      |  43 ++-
> >  src/qemu/qemu_process.h      |   2 +
> >  src/util/Makefile.inc.am     |   2 +
> >  src/util/vireventthread.c    | 190 +++++++++++
> >  src/util/vireventthread.h    |  31 ++
> >  tests/qemumonitortestutils.c |  14 +
> >  14 files changed, 700 insertions(+), 376 deletions(-)
> >  create mode 100644 src/util/vireventthread.c
> >  create mode 100644 src/util/vireventthread.h
> > 
> 
> Reviewed-by: Michal Privoznik <mprivozn at redhat.com>
> 
> Should we also delete the event loop from the qemu_shim.c then?

I've got patches for that, but its not possible until we can make the
virObjectEventState  struct use a thread for dispatch, instead of
using a event loop zero-second timer.

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 :|