Re: [PATCH 09/11] qdev: Avoid QemuOpts in QMP device_add

[Damien Hedde <damien.hedde@greensocs.com>]

On 10/5/21 16:37, Kevin Wolf wrote:
> Am 27.09.2021 um 13:39 hat Kevin Wolf geschrieben:
>> Am 27.09.2021 um 13:06 hat Damien Hedde geschrieben:
>>> On 9/24/21 11:04, Kevin Wolf wrote:
>>>> Directly call qdev_device_add_from_qdict() for QMP device_add instead of
>>>> first going through QemuOpts and converting back to QDict.
>>>>
>>>> Note that this changes the behaviour of device_add, though in ways that
>>>> should be considered bug fixes:
>>>>
>>>> QemuOpts ignores differences between data types, so you could
>>>> successfully pass a string "123" for an integer property, or a string
>>>> "on" for a boolean property (and vice versa).  After this change, the
>>>> correct data type for the property must be used in the JSON input.
>>>>
>>>> qemu_opts_from_qdict() also silently ignores any options whose value is
>>>> a QDict, QList or QNull.
>>>>
>>>> To illustrate, the following QMP command was accepted before and is now
>>>> rejected for both reasons:
>>>>
>>>> { "execute": "device_add",
>>>>     "arguments": { "driver": "scsi-cd",
>>>>                    "drive": { "completely": "invalid" },
>>>>                    "physical_block_size": "4096" } }
>>>>
>>>> Signed-off-by: Kevin Wolf <kwolf@redhat.com>
>>>> ---
>>>>    softmmu/qdev-monitor.c | 18 +++++++++++-------
>>>>    1 file changed, 11 insertions(+), 7 deletions(-)
>>>>
>>>> diff --git a/softmmu/qdev-monitor.c b/softmmu/qdev-monitor.c
>>>> index c09b7430eb..8622ccade6 100644
>>>> --- a/softmmu/qdev-monitor.c
>>>> +++ b/softmmu/qdev-monitor.c
>>>> @@ -812,7 +812,8 @@ void hmp_info_qdm(Monitor *mon, const QDict *qdict)
>>>>        qdev_print_devinfos(true);
>>>>    }
>>>> -void qmp_device_add(QDict *qdict, QObject **ret_data, Error **errp)
>>>> +static void monitor_device_add(QDict *qdict, QObject **ret_data,
>>>> +                               bool from_json, Error **errp)
>>>>    {
>>>>        QemuOpts *opts;
>>>>        DeviceState *dev;
>>>> @@ -825,7 +826,9 @@ void qmp_device_add(QDict *qdict, QObject **ret_data, Error **errp)
>>>>            qemu_opts_del(opts);
>>>>            return;
>>>>        }
>>>> -    dev = qdev_device_add(opts, errp);
>>>> +    qemu_opts_del(opts);
>>>> +
>>>> +    dev = qdev_device_add_from_qdict(qdict, from_json, errp);
>>>
>>> Hi Kevin,
>>>
>>> I'm wandering if deleting the opts (which remove it from the "device" opts
>>> list) is really a no-op ?
>>
>> It's not exactly a no-op. Previously, the QemuOpts would only be freed
>> when the device is destroying, now we delete it immediately after
>> creating the device. This could matter in some cases.
>>
>> The one case I was aware of is that QemuOpts used to be responsible for
>> checking for duplicate IDs. Obviously, it can't do this job any more
>> when we call qemu_opts_del() right after creating the device. This is
>> the reason for patch 6.
>>
>>> The opts list is, eg, traversed in hw/net/virtio-net.c in the function
>>> failover_find_primary_device_id() which may be called during the
>>> virtio_net_set_features() (a TYPE_VIRTIO_NET method).
>>> I do not have the knowledge to tell when this method is called. But If this
>>> is after we create the devices. Then the list will be empty at this point
>>> now.
>>>
>>> It seems, there are 2 other calling sites of
>>> "qemu_opts_foreach(qemu_find_opts("device"), [...]" in net/vhost-user.c and
>>> net/vhost-vdpa.c
>>
>> Yes, you are right. These callers probably need to be changed. Going
>> through the command line options rather than looking at the actual
>> device objects that exist doesn't feel entirely clean anyway.
> 
> So I tried to have a look at the virtio-net case, and ended up very
> confused.
> 
> Obviously looking at command line options (even of a differrent device)
> from within a device is very unclean. With a non-broken, i.e. type safe,
> device-add (as well as with the JSON CLI option introduced by this
> series), we can't have a QemuOpts any more that is by definition unsafe.
> So this code needs a replacement.
> 
> My naive idea was that we just need to look at runtime state instead.
> Don't search the options for a device with a matching 'failover_pair_id'
> (which, by the way, would fail as soon as any other device introduces a
> property with the same name), but search for actual PCIDevices in qdev
> that have pci_dev->failover_pair_id set accordingly.
> 
> However, the logic in failover_add_primary() suggests that we can have a
> state where QemuOpts for a device exist, but the device doesn't, and
> then it hotplugs the device from the command line options. How would we
> ever get into such an inconsistent state where QemuOpts contains a
> device that doesn't exist? Normally devices get their QemuOpts when they
> are created and device_finalize() deletes the QemuOpts again. >

Just read the following from docs/system/virtio-net-failover.rst

 > Usage
 > -----
 >
 > The primary device can be hotplugged or be part of the startup
 > configuration
 >
 >   -device virtio-net-pci,netdev=hostnet1,id=net1,
 >           mac=52:54:00:6f:55:cc,bus=root2,failover=on
 >
 > With the parameter failover=on the VIRTIO_NET_F_STANDBY feature
 > will be enabled.
 >
 > -device vfio-pci,host=5e:00.2,id=hostdev0,bus=root1,
 >         failover_pair_id=net1
 >
 > failover_pair_id references the id of the virtio-net standby device.
 > This is only for pairing the devices within QEMU. The guest kernel
 > module net_failover will match devices with identical MAC addresses.
 >
 > Hotplug
 > -------
 >
 > Both primary and standby device can be hotplugged via the QEMU
 > monitor.  Note that if the virtio-net device is plugged first a
 > warning will be issued that it couldn't find the primary device.

So maybe this whole primary device lookup can happen during the -device 
CLI option creation loop. And we can indeed have un-created devices 
still in the list ?

Damien

> Any suggestions how to get rid of the QemuOpts abuse in the failover
> code?
> 
> If this is a device that we previously managed to rip out without
> deleting its QemuOpts, can we store its dev->opts (which is a type safe
> QDict after this series) somewhere locally instead of looking at global
> state? Preferably I would even like to get rid of dev->opts because we
> really should look at live state rather than command line options after
> device creation, but I guess one step at a time.
> 
> (Actually, I'm half tempted to just break it because no test cases seem
> to exist, so apparently nobody is really interested in it.)
> 
> Kevin
>