Re: [PATCH v6 1/3] multifd: Create property multifd-flush-after-each-section

[Markus Armbruster <armbru@redhat.com>]

Juan Quintela <quintela@redhat.com> writes:

> Markus Armbruster <armbru@redhat.com> wrote:
>> Juan Quintela <quintela@redhat.com> writes:
>>
>>>>> @@ -478,6 +478,24 @@
>>>>>  #                    should not affect the correctness of postcopy migration.
>>>>>  #                    (since 7.1)
>>>>>  #
>>>>> +# @multifd-flush-after-each-section: flush every channel after each
>>>>> +#                                    section sent.  This assures that
>>>>> +#                                    we can't mix pages from one
>>>>> +#                                    iteration through ram pages with
>>
>> RAM
>
> OK.
>
>>>>> +#                                    pages for the following
>>>>> +#                                    iteration.  We really only need
>>>>> +#                                    to do this flush after we have go
>>
>> to flush after we have gone
>
> OK
>
>>>>> +#                                    through all the dirty pages.
>>>>> +#                                    For historical reasons, we do
>>>>> +#                                    that after each section.  This is
>
>> we flush after each section
>
> OK
>
>>>>> +#                                    suboptimal (we flush too many
>>>>> +#                                    times).
>
>> inefficient: we flush too often.
>
> OK
>
>>>>> +#                                    Default value is false.
>>>>> +#                                    Setting this capability has no
>>>>> +#                                    effect until the patch that
>>>>> +#                                    removes this comment.
>>>>> +#                                    (since 8.0)
>>>>
>>>> IMHO the core of this new "cap" is the new RAM_SAVE_FLAG_MULTIFD_FLUSH bit
>>>> in the stream protocol, but it's not referenced here.  I would suggest
>>>> simplify the content but highlight the core change:
>>>
>>> Actually it is the other way around.  What this capability will do is
>>> _NOT_ use RAM_SAVE_FLAG_MULTIFD_FLUSH protocol.
>>>
>>>>  @multifd-lazy-flush:  When enabled, multifd will only do sync flush after
>>
>> Spell out "synchronrous".
>
> ok.
>
>>>>                        each whole round of bitmap scan.  Otherwise it'll be
>>
>> Suggest to scratch "whole".
>
> ok.
>
>>>>                        done per RAM save iteration (which happens with a much
>>>>                        higher frequency).
>>
>> Less detail than Juan's version.  I'm not sure how much detail is
>> appropriate for QMP reference documentation.
>>
>>>>                        Please consider enable this as long as possible, and
>>>>                        keep this off only if any of the src/dst QEMU binary
>>>>                        doesn't support it.
>>
>> Clear guidance on how to use it, good!
>>
>> Perhaps state it more forcefully: "Enable this when both source and
>> destination support it."
>>
>>>>
>>>>                        This capability is bound to the new RAM save flag
>>>>                        RAM_SAVE_FLAG_MULTIFD_FLUSH, the new flag will only
>>>>                        be used and recognized when this feature bit set.
>>
>> Is RAM_SAVE_FLAG_MULTIFD_FLUSH visible in the QMP interface?  Or in the
>> migration stream?
>
> No.  Only migration stream.

Doc comments should be written for readers of the QEMU QMP Reference
Manual.  Is RAM_SAVE_FLAG_MULTIFD_FLUSH relevant for them?

Perhaps the relevant part is "the peer needs to enable this capability
too, or else", for a value of "or else".

What happens when the source enables, and the destination doesn't?

What happens when the destination enables, and the source doesn't?

Any particular reason for having the destination recognize the flag only
when the capability is enabled?

>> I'm asking because doc comments are QMP reference documentation, but
>> when writing them, it's easy to mistake them for internal documentation,
>> because, well, they're comments.
>
>>> Name is wrong.  It would be multifd-non-lazy-flush.  And I don't like
>>> negatives.  Real name is:
>>>
>>> multifd-I-messed-and-flush-too-many-times.
>>
>> If you don't like "non-lazy", say "eager".
>
> more than eager it is unnecesary.

"overeager"?

>>>> I know you dislike multifd-lazy-flush, but that's still the best I can come
>>>> up with when writting this (yeah I still like it :-p), please bare with me
>>>> and take whatever you think the best.
>>>
>>> Libvirt assumes that all capabilities are false except if enabled.
>>> We want RAM_SAVE_FLAG_MULTFD_FLUSH by default (in new machine types).
>>>
>>> So, if we can do
>>>
>>> capability_use_new_way = true
>>>
>>> We change that to
>>>
>>> capability_use_old_way = true
>>>
>>> And then by default with false value is what we want.
>>
>> Eventually, all supported migration peers will support lazy flush.  What
>> then?  Will we flip the default?  Or will we ignore the capability and
>> always flush lazily?
>
> I have to take a step back.  Cope with me.
>
> How we fix problems in migration that make the stream incompatible.
> We create a property.
>
> static Property migration_properties[] = {
>     ...
>     DEFINE_PROP_BOOL("decompress-error-check", MigrationState,
>                       decompress_error_check, true),
>     ....
> }
>
> In this case it is true by default.
>
> GlobalProperty hw_compat_2_12[] = {
>     { "migration", "decompress-error-check", "off" },
>     ...
> };
>
> We introduced it on whatever machine that is newer than 2_12.
> Then we make it "off" for older machine types, that way we make sure
> that migration from old qemu to new qemu works.
>
> And we can even left libvirt, if they know that both qemus are new, they
> can setup the property to true even for old machine types.
>
> So, what we have:
>
> Machine 2_13 and newer use the new code.
> Machine 2_12 and older use the old code (by default).
> We _can_ migrate machine 2_12 with new code, but we need to setup it
> correctly on both sides.
> We can run the old code with machine type 2_13.  But I admit than that
> is only useful for testing, debugging, meassuring performance, etc.
>
> So, the idea here is that we flush a lot of times for old machine types,
> and we only flush when needed for new machine types.  Libvirt (or
> whoever) can use the new method if it sees fit just using the
> capability.

Got it.

What should be done, if anything, when all machines defaulting to the
old code are gone?  Getting rid of the old code along with the
capability is desirable, isn't it?  But if the capability is a stable
interface, the deprecation process applies.  Should it be stable?  Or
should it be just something we use to do the right thing depending on
the machine types (primary purpose), and also enable experimentation
(secondary purpose)?

> Now that I am telling this, I can switch back to a property instead of a
> capability:
> - I can have the any default value that I want
> - So I can name it multifd_lazy_flush or whatever.
>
> Later, Juan.