Re: [PATCH v2 1/2] qdev-monitor: avoid QemuOpts in QMP device_add

[Markus Armbruster <armbru@redhat.com>]

Stefan Hajnoczi <stefanha@redhat.com> writes:

> The QMP device_add monitor command converts the QDict arguments to
> QemuOpts and then back again to QDict. This process only supports scalar
> types. Device properties like virtio-blk-pci's iothread-vq-mapping (an
> array of objects) are silently dropped by qemu_opts_from_qdict() during
> the QemuOpts conversion even though QAPI is capable of validating them.
> As a result, hotplugging virtio-blk-pci devices with the
> iothread-vq-mapping property does not work as expected (the property is
> ignored).
>
> Get rid of the QemuOpts conversion in qmp_device_add() and call
> qdev_device_add_from_qdict() with from_json=true. Using the QMP
> command's QDict arguments directly allows non-scalar properties.
>
> The HMP is also adjusted since qmp_device_add()'s now expects properly
> typed JSON arguments and cannot be used from HMP anymore. Move the code
> that was previously in qmp_device_add() (with QemuOpts conversion and
> from_json=false) into hmp_device_add() so that its behavior is
> unchanged.
>
> This patch changes the behavior of QMP device_add but not HMP
> device_add. QMP clients that sent incorrectly typed device_add QMP
> commands no longer work. This is a breaking change but clients should be
> using the correct types already. See the netdev_add QAPIfication in
> commit db2a380c8457 for similar reasoning and object-add in commit
> 9151e59a8b6e. Unlike those commits, we continue to rely on 'gen': false
> for the time being.
>
> Move the drain_call_rcu() invocation into qdev_device_add_from_qdict()
> so all callers benefit from it automatically. This avoids code
> duplication.
>
> Markus helped me figure this out and even provided a draft patch. The
> code ended up very close to what he suggested.
>
> Suggested-by: Markus Armbruster <armbru@redhat.com>
> Cc: Daniel P. Berrangé <berrange@redhat.com>
> Signed-off-by: Stefan Hajnoczi <stefanha@redhat.com>
> ---
>  system/qdev-monitor.c | 56 ++++++++++++++++++++++---------------------
>  1 file changed, 29 insertions(+), 27 deletions(-)
>
> diff --git a/system/qdev-monitor.c b/system/qdev-monitor.c
> index 6af6ef7d66..8a756b1a91 100644
> --- a/system/qdev-monitor.c
> +++ b/system/qdev-monitor.c
> @@ -725,6 +725,17 @@ err_del_dev:
>      if (dev) {
>          object_unparent(OBJECT(dev));
>          object_unref(OBJECT(dev));
> +
> +        /*
> +         * Drain all pending RCU callbacks. This is done because
> +         * some bus related operations can delay a device removal
> +         * (in this case this can happen if device is added and then
> +         * removed due to a configuration error)
> +         * to a RCU callback, but user might expect that this interface
> +         * will finish its job completely once qmp command returns result
> +         * to the user
> +         */
> +        drain_call_rcu();
>      }
>      return NULL;
>  }

Moving this from qmp_device_add() adds RCU draining to call chains not
going through qmp_device_add().

Can adding it hurt?  I guess it can't.

Can it fix bugs?  I don't know.

Let's review the callers of qdev_device_add_from_qdict():

* qdev_device_add()

  - called from qmp_device_add()

    No change.

  - called from device_init_func() called from qemu_create_cli_devices()

    See below.

  - called from usbback_portid_add() called from usbback_process_port()
    called from usbback_backend_changed()

    · called from usbback_init()

    · called as XenDevOps method backend_changed()

    This is Xen.  We now drain pending RCU callbacks.  Impact?  Beats
    me.

* qemu_create_cli_devices() called from qmp_x_exit_preconfig()

  - as QMP command with -preconfig, phase must be
    PHASE_MACHINE_INITIALIZED

  - called from qemu_init() without -preconfig

  We now drain pending RCU callbacks.  Can any be pending at this
  early point?  If not, the change is a no-op.

* failover_add_primary() called from virtio_net_set_features() called as
  VirtioDeviceClass method set_features()

  This is virtio-net failover.  We now drain pending RCU callbacks.
  Impact?  Beats me.

My gut feeling is "improvement, possibly even a bug fix".  It deserves
its own commit, doesn't it?

> @@ -849,34 +860,10 @@ void hmp_info_qdm(Monitor *mon, const QDict *qdict)
>  
>  void qmp_device_add(QDict *qdict, QObject **ret_data, Error **errp)
>  {
> -    QemuOpts *opts;
>      DeviceState *dev;
>  
> -    opts = qemu_opts_from_qdict(qemu_find_opts("device"), qdict, errp);
> -    if (!opts) {
> -        return;
> -    }
> -    if (!monitor_cur_is_qmp() && qdev_device_help(opts)) {
> -        qemu_opts_del(opts);
> -        return;
> -    }
> -    dev = qdev_device_add(opts, errp);
> -    if (!dev) {
> -        /*
> -         * Drain all pending RCU callbacks. This is done because
> -         * some bus related operations can delay a device removal
> -         * (in this case this can happen if device is added and then
> -         * removed due to a configuration error)
> -         * to a RCU callback, but user might expect that this interface
> -         * will finish its job completely once qmp command returns result
> -         * to the user
> -         */
> -        drain_call_rcu();
> -
> -        qemu_opts_del(opts);
> -        return;
> -    }
> -    object_unref(OBJECT(dev));
> +    dev = qdev_device_add_from_qdict(qdict, true, errp);
> +    object_unref(dev);
>  }
>  
>  static DeviceState *find_device_state(const char *id, Error **errp)
> @@ -967,8 +954,23 @@ void qmp_device_del(const char *id, Error **errp)
>  void hmp_device_add(Monitor *mon, const QDict *qdict)
>  {
>      Error *err = NULL;
> +    QemuOpts *opts;
> +    DeviceState *dev;
>  
> -    qmp_device_add((QDict *)qdict, NULL, &err);
> +    opts = qemu_opts_from_qdict(qemu_find_opts("device"), qdict, &err);
> +    if (!opts) {
> +        goto out;
> +    }
> +    if (qdev_device_help(opts)) {
> +        qemu_opts_del(opts);
> +        return;
> +    }
> +    dev = qdev_device_add(opts, &err);
> +    if (!dev) {
> +        qemu_opts_del(opts);
> +    }
> +    object_unref(dev);
> +out:
>      hmp_handle_error(mon, err);
>  }

Remainder looks good to me.