Re: [PATCH 2/2] Support monitor chardev hotswap with QMP

[Markus Armbruster <armbru@redhat.com>]

Li Zhang <zhlcindy@gmail.com> writes:

> From: Li Zhang <li.zhang@cloud.ionos.com>
>
> For some scenarios, it needs to hot-add a monitor device.
> But QEMU doesn't support hotplug yet. It also works by adding
> a monitor with null backend by default and then change its
> backend to socket by QMP command "chardev-change".
>
> So this patch is to support monitor chardev hotswap with QMP.
>
> Signed-off-by: Li Zhang <li.zhang@cloud.ionos.com>

I think what what you're trying to say is that chardev-change does not
work when the character device changes is used by a QMP monitor.
Correct?

If yes, how exactly does it misbehave?

Does it work with an HMP monitor?

> ---
>  monitor/monitor-internal.h |  3 +++
>  monitor/monitor.c          |  2 +-
>  monitor/qmp.c              | 42 +++++++++++++++++++++++++++++++++++---
>  3 files changed, 43 insertions(+), 4 deletions(-)
>
> diff --git a/monitor/monitor-internal.h b/monitor/monitor-internal.h
> index 40903d6386..2df6dd21de 100644
> --- a/monitor/monitor-internal.h
> +++ b/monitor/monitor-internal.h
> @@ -186,4 +186,7 @@ int hmp_compare_cmd(const char *name, const char *list);
>  void qmp_query_qmp_schema(QDict *qdict, QObject **ret_data,
>                                   Error **errp);
>  
> +gboolean monitor_unblocked(GIOChannel *chan, GIOCondition cond,
> +                               void *opaque);
> +
>  #endif
> diff --git a/monitor/monitor.c b/monitor/monitor.c
> index e94f532cf5..2d255bab18 100644
> --- a/monitor/monitor.c
> +++ b/monitor/monitor.c
> @@ -157,7 +157,7 @@ static inline bool monitor_is_hmp_non_interactive(const Monitor *mon)
>  
>  static void monitor_flush_locked(Monitor *mon);
>  
> -static gboolean monitor_unblocked(GIOChannel *chan, GIOCondition cond,
> +gboolean monitor_unblocked(GIOChannel *chan, GIOCondition cond,
>                                    void *opaque)
>  {
>      Monitor *mon = opaque;
> diff --git a/monitor/qmp.c b/monitor/qmp.c
> index 2326bd7f9b..55cfb230d9 100644
> --- a/monitor/qmp.c
> +++ b/monitor/qmp.c
> @@ -44,6 +44,7 @@ struct QMPRequest {
>      Error *err;
>  };
>  typedef struct QMPRequest QMPRequest;
> +static void monitor_qmp_set_handlers_bh(void *opaque);
>  
>  QmpCommandList qmp_commands, qmp_cap_negotiation_commands;
>  
> @@ -480,7 +481,35 @@ void monitor_data_destroy_qmp(MonitorQMP *mon)
>      g_queue_free(mon->qmp_requests);
>  }
>  
> -static void monitor_qmp_setup_handlers_bh(void *opaque)
> +static int monitor_qmp_change(void *opaque)
> +{
> +    MonitorQMP *mon = opaque;
> +
> +    mon->common.use_io_thread =
> +        qemu_chr_has_feature(mon->common.chr.chr, QEMU_CHAR_FEATURE_GCONTEXT);
> +
> +    if (mon->common.use_io_thread) {
> +        aio_bh_schedule_oneshot(iothread_get_aio_context(mon_iothread),
> +                                monitor_qmp_set_handlers_bh, mon);
> +    } else {
> +        qemu_chr_fe_set_handlers(&mon->common.chr, monitor_can_read,
> +                                 monitor_qmp_read, monitor_qmp_event,
> +                                 monitor_qmp_change, &mon->common, NULL, true);
> +    }
> +
> +    if (mon->common.out_watch) {
> +        g_source_remove(mon->common.out_watch);

All other updates of @out_watch are under @mon_lock.  Why not this one?

I have no idea whether g_source_remove() is the right function to call.
Its documentation says "You must use g_source_destroy() for sources
added to a non-default main context."  The qemu_chr_fe_set_handlers()
contract is of no help.

Documentation of g_source_destroy() confuses some more: "This does not
unref the GSource: if you still hold a reference, use g_source_unref()
to drop it.

Marc-André, can you help?

> +        qemu_mutex_lock(&mon->common.mon_lock);
> +        mon->common.out_watch =
> +        qemu_chr_fe_add_watch(&mon->common.chr, G_IO_OUT | G_IO_HUP,
> +                               monitor_unblocked, &mon->common);

Bad indentation.  Better:

        mon->common.out_watch =
            qemu_chr_fe_add_watch(&mon->common.chr, G_IO_OUT | G_IO_HUP,
                                   monitor_unblocked, &mon->common);

or

        mon->common.out_watch = qemu_chr_fe_add_watch(&mon->common.chr,
                                                      G_IO_OUT | G_IO_HUP,
                                                      monitor_unblocked,
                                                      &mon->common);

or

        mon->common.out_watch = qemu_chr_fe_add_watch(&mon->common.chr,
                                        G_IO_OUT | G_IO_HUP,
                                        monitor_unblocked, &mon->common);

> +        qemu_mutex_unlock(&mon->common.mon_lock);
> +    }
> +
> +    return 0;
> +}

This function copies from monitor_data_init(), monitor_init_qmp(), and
monitor_flush_locked().  Feels like a refactoring would be in order.
Possibly on top.

> +
> +static void monitor_qmp_set_handlers_bh(void *opaque)
>  {
>      MonitorQMP *mon = opaque;
>      GMainContext *context;
> @@ -490,7 +519,14 @@ static void monitor_qmp_setup_handlers_bh(void *opaque)
>      assert(context);
>      qemu_chr_fe_set_handlers(&mon->common.chr, monitor_can_read,
>                               monitor_qmp_read, monitor_qmp_event,
> -                             NULL, &mon->common, context, true);
> +                             monitor_qmp_change, &mon->common, context, true);
> +
> +}
> +
> +static void monitor_qmp_setup_handlers_bh(void *opaque)
> +{
> +    MonitorQMP *mon = opaque;
> +    monitor_qmp_set_handlers_bh(mon);
>      monitor_list_append(&mon->common);
>  }
>  
> @@ -531,7 +567,7 @@ void monitor_init_qmp(Chardev *chr, bool pretty, Error **errp)
>      } else {
>          qemu_chr_fe_set_handlers(&mon->common.chr, monitor_can_read,
>                                   monitor_qmp_read, monitor_qmp_event,
> -                                 NULL, &mon->common, NULL, true);
> +                                 monitor_qmp_change, &mon->common, NULL, true);
>          monitor_list_append(&mon->common);
>      }
>  }