On Thu, Jun 07, 2018 at 01:24:29PM +0800, Wei Wang wrote:
> On 06/06/2018 07:02 PM, Peter Xu wrote:
> > On Wed, Jun 06, 2018 at 06:04:23PM +0800, Wei Wang wrote:
> > > On 06/06/2018 01:42 PM, Peter Xu wrote:
> > > > IMHO migration states do not suite here. IMHO bitmap syncing is too
> > > > frequently an operation, especially at the end of a precopy migration.
> > > > If you really want to introduce some notifiers, I would prefer
> > > > something new rather than fiddling around with migration state. E.g.,
> > > > maybe a new migration event notifiers, then introduce two new events
> > > > for both start/end of bitmap syncing.
> > > Please see if below aligns to what you meant:
> > >
> > > MigrationState {
> > > ...
> > > + int ram_save_state;
> > >
> > > }
> > >
> > > typedef enum RamSaveState {
> > > RAM_SAVE_BEGIN = 0,
> > > RAM_SAVE_END = 1,
> > > RAM_SAVE_MAX = 2
> > > }
> > >
> > > then at the step 1) and 3) you concluded somewhere below, we change the
> > > state and invoke the callback.
> > I mean something like this:
> >
> > 1693c64c27 ("postcopy: Add notifier chain", 2018-03-20)
> >
> > That was a postcopy-only notifier. Maybe we can generalize it into a
> > more common notifier for the migration framework so that we can even
> > register with non-postcopy events like bitmap syncing?
>
> Precopy already has its own notifiers: git 99a0db9b
> If we want to reuse, that one would be more suitable. I think mixing
> non-related events into one notifier list isn't nice.
I think that's only for migration state changes?
>
> > >
> > > Btw, the migration_state_notifiers is already there, but seems not really
> > > used (I only tracked spice-core.c called
> > > add_migration_state_change_notifier). I thought adding new migration
> > > states
> > > can reuse all that we have.
> > > What's your real concern about that? (not sure how defining new events
> > > would
> > > make a difference)
> > Migration state is exposed via control path (QMP). Adding new states
> > mean that the QMP clients will see more. IMO that's not really
> > anything that a QMP client will need to know, instead we can keep it
> > internally. That's a reason from compatibility pov.
> >
> > Meanwhile, it's not really a state-thing at all for me. It looks
> > really more like hook or event (start/stop of sync).
>
> Thanks for sharing your concerns in detail, which are quite helpful for the
> discussion. To reuse 99a0db9b, we can also add sub-states (or say events),
> instead of new migration states.
> For example, we can still define "enum RamSaveState" as above, which can be
> an indication for the notifier queued on the 99a0db9b notider_list to decide
> whether to call start or stop.
> Does this solve your concern?
Frankly speaking I don't fully understand how you would add that
sub-state. If you are confident with the idea, maybe you can post
your new version with the change, then I can read the code.
>
>
> >
> > > > > I would suggest to focus on the supplied interface and its usage in
> > > > > live
> > > > > migration. That is, now we have two APIs, start() and stop(), to
> > > > > start and
> > > > > stop the optimization.
> > > > >
> > > > > 1) where in the migration code should we use them (do you agree with
> > > > > the
> > > > > step (1), (2), (3) you concluded below?)
> > > > > 2) how should we use them, directly do global call or via notifiers?
> > > > I don't know how Dave and Juan might think; here I tend to agree with
> > > > Michael that some notifier framework should be nicer.
> > > >
> > > What would be the advantages of using notifiers here?
> > Isolation of modules? Then migration/ram.c at least won't need to
> > include something like "balloon.h".
> >
> > And I think it's also possible too if some other modules would like to
> > hook at these places someday.
>
> OK, I can implement it with notifiers, but this (framework) is usually added
> when someday there is a second user who needs a callback at the same place.
>
>
>
> > >
> > >
> > > > This is not that obvious to me. For now I think it's true, since when
> > > > we call stop() we'll take the mutex, meanwhile the mutex is actually
> > > > always held by the iothread (in the big loop in
> > > > virtio_balloon_poll_free_page_hints) until either:
> > > >
> > > > - it sleeps in qemu_cond_wait() [1], or
> > > > - it leaves the big loop [2]
> > > >
> > > > Since I don't see anyone who will set dev->block_iothread to true for
> > > > the balloon device, then [1] cannot happen;
> > > there is a case in virtio_balloon_set_status which sets
> > > dev->block_iothread
> > > to true.
> > >
> > > Did you mean the free_page_lock mutex? it is released at the bottom of the
> > > while() loop in virtio_balloon_poll_free_page_hint. It's actually released
> > > for every hint. That is,
> > >
> > > while(1){
> > > take the lock;
> > > process 1 hint from the vq;
> > > release the lock;
> > > }
> > Ah, so now I understand why you need the lock to be inside the loop,
> > since the loop is busy polling actually. Is it possible to do this in
> > an async way?
>
> We need to use polling here because of some back story in the guest side
> (due to some locks being held) that makes it a barrier to sending
> notifications for each hints.
Any link to the "back story" that I can learn about? :) If it's too
complicated a problem and you think I don't need to understand at all,
please feel free to do so. Then I would assume at least Michael has
fully acknowledged that idea, and I can just stop putting more time on
this part.
Besides, if you are going to use a busy loop, then I would be not
quite sure about whether you really want to share that iothread with
others, since AFAIU that's not how iothread is designed (which is
mostly event-based and should not welcome out-of-control blocking in
the handler of events). Though I am not 100% confident about my
understaning on that, I only raise this question up. Anyway you'll
just take over the thread for a while without sharing, and after the
burst IOs it's mostly never used (until the next bitmap sync). Then
it seems a bit confusing to me on why you need to share that after
all.
>
> > I'm a bit curious on how much time will it use to do
> > one round of the free page hints (e.g., an idle guest with 8G mem, or
> > any configuration you tested)? I suppose during that time the
> > iothread will be held steady with 100% cpu usage, am I right?
>
> Compared to the time spent by the legacy migration to send free pages, that
> small amount of CPU usage spent on filtering free pages could be neglected.
> Grinding a chopper will not hold up the work of cutting firewood :)
Sorry I didn't express myself clearly.
My question was that, have you measured how long time it will take
from starting of the free page hints (when balloon state is set to
FREE_PAGE_REPORT_S_REQUESTED), until it completes (when QEMU receives
the VIRTIO_BALLOON_FREE_PAGE_REPORT_STOP_ID, then set the status to
FREE_PAGE_REPORT_S_STOP)?
--
Peter Xu
