On 06/15/20 17:02, Philippe Mathieu-Daudé wrote:
> On 6/15/20 4:45 PM, Gerd Hoffmann wrote:
>> Hi,
>>
>>>> I can explain the rationale for that change, but I'm not sure of the
>>>> answer to your question. That changes makes sure that the fw_cfg data
>>>> remains exactly the same even on newer versions of qemu if the machine
>>>> is set the same. This way you can do migrations to newer qemu versions
>>>> and anything using fw_cfg won't get confused because the data changes.
>>>>
>>>> The reason that change was so complex was preserving the order for
>>>> migrating from older versions.
>>>>
>>>> This is only about migration. I'm not sure what gen_id is, but if it's
>>>> migrated, it better be future proof.
>>>
>>> Whenever introducing a new fw_cfg file (*any* new named file), how do we
>>> decide whether we need fw_cfg_set_order_override()?
>>
>> The whole point of the sorting is to make sure the fw_cfg directory
>> listing entry (FW_CFG_FILE_DIR) is stable and doesn't change underneath
>> the guest when it gets live-migrated.
>>
>> That sorting was added in qemu 2.6, to make sure things don't chance by
>> accident in case the initialization order changes. Now you've got a
>> problem when you migrate from qemu 2.5 (+older) to qemu 2.6 (+newer),
>> because 2.5 has the entries in initialization order and 2.6 has the
>> entries in alphabetical order. To deal with that machine types for 2.5
>> & older keep the old sort order. This is the reason why
>> legacy_fw_cfg_order exists.
>>
>> For new features and files you can completely ignore the whole legacy
>> sorting mess. cross-version live migration works only for features
>> supported by both qemu versions, therefore the legacy sorting is only
>> relevant for features & files already supported by qemu 2.5.
>
> Thanks you Gerd for the whole explanation. I added an entry to
> my TODO list to document this, based on your comment (and Corey's).
Yes, please!
Apparently, I've been confused by commit bab47d9a75a3 ("Sort the fw_cfg
file list", 2016-04-07) before (in January 2018):
http://mid.mail-archive.com/5367c8a4-91bd-7712-525d-0a1ed6e6acab@redhat.com
(See in particular my question which I believe remains relevant:
"is the idea that the same machine type on a new QEMU release may only
reorder the additions of the preexistent fw_cfg files across the source
code, but must not expose *new* fw_cfg files?"
And I think Gerd just answered that above (in the positive), namely,
"cross-version live migration works only for features supported by both
qemu versions". So indeed we must not have a new fw_cfg file appear in
an old machine type on a new QEMU release, without the user explicitly
asking for it on the command line.)
> I'll address it later, as you confirmed it doesn't impact this
> series.
That's my understanding too. Thanks for explaining, Gerd!
Laszlo
