Re: [Qemu-devel] [RFC PATCH 00/10] monitor: Split monitor.c in core/HMP/QMP/misc

[Kevin Wolf <kwolf@redhat.com>]

Am 07.06.2019 um 16:03 hat Daniel P. Berrangé geschrieben:
> On Fri, Jun 07, 2019 at 03:54:20PM +0200, Kevin Wolf wrote:
> > monitor.c mixes a lot of different things in a single file: The core
> > monitor infrastructure, HMP infrastrcture, QMP infrastructure, and the
> > implementation of several HMP and QMP commands. Almost worse, struct
> > Monitor mixes state for HMP, for QMP, and state actually shared between
> > all monitors. monitor.c must be linked with a system emulator and even
> > requires per-target compilation because some of the commands it
> > implements access system emulator state.
> > 
> > The reason why I care about this is that I'm working on a protoype for a
> > storage daemon, which wants to use QMP (but probably not HMP) and
> > obviously doesn't have any system emulator state. So I'm interested in
> > some core monitor parts that can be linked to non-system-emulator tools.
> > 
> > This series first creates separate structs MonitorQMP and MonitorHMP
> > which inherit from Monitor, and then moves the associated infrastructure
> > code into separate source files.
> > 
> > While the split is probably not perfect, I think it's an improvement of
> > the current state even for QEMU proper, and it's good enough so I can
> > link my storage daemon against just monitor/core.o and monitor/qmp.o and
> > get a useless QMP monitor that parses the JSON input and rejects
> > everything as an unknown command.
> > 
> > Next I'll try to teach it a subset of QMP commands that can actually be
> > supported in a tool, but while there will be a few follow-up patches to
> > achieve this, I don't expect that this work will bring up much that
> > needs to be changed in the splitting process done in this series.
> > 
> > Kevin Wolf (10):
> >   monitor: Remove unused password prompting fields
> >   monitor: Split monitor_init in HMP and QMP function
> >   monitor: Make MonitorQMP a child class of Monitor
> >   monitor: Create MonitorHMP with readline state
> >   monitor: Move cmd_table to MonitorHMP
> >   Move monitor.c to monitor/misc.c
> >   monitor: Create monitor_int.h with common definitions
> >   monitor: Split out monitor/qmp.c
> >   monitor: Split out monitor/hmp.c
> >   monitor: Split out monitor/core.c
> > 
> >  include/monitor/monitor.h |    8 +-
> >  monitor/monitor_int.h     |  207 ++
> >  hmp.c                     |    4 +-
> >  monitor.c                 | 4727 -------------------------------------
> >  monitor/core.c            |  604 +++++
> >  monitor/hmp.c             | 1351 +++++++++++
> >  monitor/misc.c            | 2406 +++++++++++++++++++
> >  monitor/qmp.c             |  404 ++++
> >  Makefile.objs             |    1 +
> >  Makefile.target           |    3 +-
> >  monitor/Makefile.objs     |    2 +
> 
> It will be nice to have the monitor code split up a bit more.
> 
> I'm not a fan, however, of having both $ROOT/qmp.c and $ROOT/monitor/qmp.c
> Likwise  $ROOT/hmp.c and $ROOT/monitor/hmp.c.  Can we move those other
> existing files out of the root dir, into monitor/, so we don't have two
> files with the same name in different dirs.

$ROOT/hmp.c and $ROOT/qmp.c contain various command implementations,
just as $ROOT/monitor/misc.c. This is still a bit of a mess. I'll have
to address this at least partially in the next step because I need to
separate commands that can be linked with tools from those that require
a system emulator.

My plan involves at least creating some monitor/qmp-cmds-*.c, which
might already make $ROOT/qmp.c empty. Even though I don't strictly need
it, there's no reason not to do the same for HMP, too. In any case, I'd
rather address this in a separate follow-up series.

But if people prefer, I can move the existing files in the root
directory to monitor/{qmp,hmp}-cmds.c temporarily in this series and
then work from there with follow-ups until they are empty (or maybe I
don't even have to make them completely empty then).

Kevin