Re: [Qemu-devel] [PATCH v11 01/12] qmp: delete qemu opts when delete an object

[Markus Armbruster]

Yang Hongyang <address@hidden> writes:

> On 09/24/2015 05:42 PM, Markus Armbruster wrote:
>> Yang Hongyang <address@hidden> writes:
>>
>>> On 09/24/2015 03:43 PM, Markus Armbruster wrote:
>>>> This has finally reached the front of my review queue.  I apologize for
>>>> the loooong delay.
>>>>
>>>> Copying Paolo for another pair of eyeballs (he wrote this code).
>>>>
>>> [...]
>>>>> +
>>>>> +    opts = qemu_opts_find(qemu_find_opts_err("object", NULL), id);
>>>>> +    qemu_opts_del(opts);
>>>>
>>>> qemu_find_opts_err("object", &error_abort) please, because when it
>>>> fails, we want to die right away, not when the null pointer it returns
>>>> gets dereferenced.
>>>
>>> Thanks for the review.
>>> Jason, do you want me to propose a fix on top of this series or simply drop
>>> this for now because this patch is an independent bug fix and won't
>>> affect the
>>> other filter patch series.
>>>
>>>>
>>>> Same sloppiness in netdev_del_completion() and qmp_netdev_del(), not
>>>> your patch's fault.
>>>>
>>>> Elsewhere, we store the QemuOpts in the object just so we can delete it:
>>>> DeviceState, DriveInfo.  Paolo, what do you think?
>>>
>>> I don't get it. Currently, only objects created at the beginning through
>>> QEMU command line will be stored in the QemuOpts, objects that created
>>> with object_add won't stored in QemuOpts. Do you mean for DeviceState,
>>> DriveInfo they store there QemuOpts explicity so that they can delete it?
>>> Why don't we just delete it from objects directly instead?
>>
>> Let me elaborate.
>
> Thanks very much for the elaboration.
>
>>
>> We have the same pattern in multiple places: some kind of object gets
>> configured via QemuOpts, and an object's QemuOpts need to stay around
>> until the object dies.
>>
>> Example 1: Block device backends
>>
>>      DriveInfo has a member opts.
>>
>>      drive_new() stores the QemuOpts in dinfo->opts.
>>
>>      drive_info_del() destroys dinfo->opts.
>>
>>      Note: DriveInfo member opts is always non-null.  But not every
>>      BlockBackend has a DriveInfo.
>>
>> Example 2: Device frontends
>>
>>      DeviceState has a member opts.
>>
>>      qdev_device_add() stores the QemuOpts in dev->opts.
>>
>>      device_finalize() destroys dev->opts.
>>
>>      Note: DeviceState member opts may be null (not every device is
>>      created by qdev_device_add()).  Fine, because qemu_opts_del(NULL) is
>>      a no-op.
>>
>> Example 3: Character device backends
>>
>>      CharDriverState has a member opts.
>>
>>      qemu_chr_new_from_opts() stores the QemuOpts in chr->opts.
>>
>>      qemu_chr_delete() destroys chr->opts.
>
> 1-3 store there ops in there own state, not in global ops group right?

Both!  But keeping a pointer in their own state simplifies calling
qemu_opts_del() on destruction, and also makes it more obvious what is
keeping the QemuOpts alive.

>> Example 4: Network device backends
>>
>>      Two cases
>>
>>      A. netdev
>>
>>         qmp_netdev_add() does not store the QemuOpts.
>>
>>         qmp_netdev_del() still needs to destroy it.  It has to find it
>>         somehow.  Here's how it does it:
>>
>>             opts = qemu_opts_find(qemu_find_opts_err("netdev", NULL), id);
>>             if (!opts) {
>>                 error_setg(errp, "Device '%s' is not a netdev", id);
>>                 return;
>>             }
>>
>>         The !opts condition is a non-obvious way to test "not created
>>         with -netdev", see commit 645c949.  Note that the commit's claim
>>         that qemu_opts_del(NULL) crashes is no longer true since commit
>>         4782183.
>>
>>      B. Legacy net
>>
>>         hmp_host_net_add() does not store the QemuOpts.
>
> I'm afraid it does store the QemuOpts, but not in it's own state.
> net/net.c:
> 1088     qemu_opt_set(opts, "type", device, &error_abort);
> This will store the QemuOpts, or am I misunderstood it?

Doesn't store opts anywhere, actually.  It merely modifies it (adds a
parameter "type")

>>
>>         hmp_host_net_remove() still needs to destroy it.  I can't see
>>         where that happens, and I'm not sure it does.
>>
>> Example 5: Generic object
>>
>>      object_create() does not store the QemuOpts.
>>
>>      It still needs to be destroyed along with the object.  It isn't, and
>>      your patch fixes it.
>>
>> Personally, I find the technique in example 1-3 easier to understand
>> than the one in example 4-5.
>
> I agree that opts should not be used to determine not created something
> while there's case when something created but Opts not stored.