Re: [Qemu-devel] [PATCH] migration: optimize the downtime

[Jay Zhou]

On 2017/7/24 23:35, Dr. David Alan Gilbert wrote:
> * Jay Zhou (address@hidden) wrote:
> > Hi Dave,
> >
> > On 2017/7/21 17:49, Dr. David Alan Gilbert wrote:
> > > * Jay Zhou (address@hidden) wrote:
> > > > Qemu_savevm_state_cleanup() takes about 300ms in my ram migration tests
> > > > with a 8U24G vm(20G is really occupied), the main cost comes from
> > > > KVM_SET_USER_MEMORY_REGION ioctl when mem.memory_size = 0 in
> > > > kvm_set_user_memory_region(). In kmod, the main cost is
> > > > kvm_zap_obsolete_pages(), which traverses the active_mmu_pages list to
> > > > zap the unsync sptes.
> > >
> > > Hi Jay,
> > >     Is this actually increasing the real downtime when the guest isn't
> > > running, or is it just the reported time? I see that the s->downtime
> > > value is calculated right after where we currently call
> > > qemu_savevm_state_cleanup.
> >
> > It actually increased the real downtime, I used the "ping" command to
> > test. Reason is that the source side libvirt sends qmp to qemu to query
> > the status of migration, which needs the BQL. qemu_savevm_state_cleanup
> > is done with BQL, qemu can not handle the qmp if qemu_savevm_state_cleanup
> > has not finished. And the source side libvirt delays about 300ms to notify
> > the destination side libvirt to send the "cont" command to start the vm.
> >
> > I think the value of s->downtime is not accurate enough, maybe we could
> > move the calculation of end_time after qemu_savevm_state_cleanup has done.
>
> I'm copying in Paolo, Radim and Andrea- is there anyway we can make the
> teardown of KVMs dirty tracking not take so long? 300ms is a silly long time
> on only a small VM.
>
> > >     I guess the biggest problem is that 300ms happens before we restart
> > > the guest on the source if a migration fails.
> >
> > 300ms happens even if a migration succeeds.
>
> Hmm, OK, this needs fixing then - it does explain a result I saw a while
> ago where the downtime was much bigger with libvirt than it was with
> qemu on it's own.
>
> > > > I think it can be optimized:
> > > > (1) source vm will be destroyed if the migration is successfully done,
> > > >       so the resources will be cleanuped automatically by the system
> > > > (2) delay the cleanup if the migration failed
> > >
> > > I don't like putting it in qmp_cont; that shouldn't have migration magic
> > > in it.
> >
> > Yes, it is not a ideal place. :(
> >
> > > I guess we could put it in migrate_fd_cleanup perhaps? It gets called on
> > > a bh near the end -  or could we just move it closer to the end of
> > > migration_thread?
> >
> > I have tested putting it in migrate_fd_cleanup, but the downtime is not
> > optimized. So I think it is the same to move it closer to the end of
> > migration_thread if it holds the BQL.
> > Could we put it in migrate_init?
>
> Your explanation above hints as to why migrate_fd_cleanup doesn't help;
> it's because we're still going to be doing it with the BQL taken.

Yes, it is.

> Can you tell me which version of libvirt you're using?

I'm using 1.3.4

> I thought the newer ones were supposed to use events so they did't
> have to poll qemu.

After checking the codes of the newest libvirt, I think it is the same
in the qemuMigrationWaitForCompletion function, which is used to poll
qemu every 50ms.

Thanks,
Jay

>   If we move qemu_savevm_state_cleanup is it still safe? Are there
> some things we're supposed to do at that point which are wrong if
> we don't.
>
> I wonder about something like;  take a mutex in
> memory_global_dirty_log_start, release it in
> memory_global_dirty_log_stop.  Then make ram_save_cleanup start
> a new thread that does the call to memory_global_dirty_log_stop.
>
> Dave