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

[Dr. David Alan Gilbert]

* Paolo Bonzini (address@hidden) wrote:
> On 24/07/2017 17:35, 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.
> 
> Xiao Guangrong is working on something vaguely related (but different
> and simpler because it's entirely contained within KVM), which is to
> make log_sync faster.
> 
> The Intel folks working on clear containers also would like
> MemoryListeners to have a better complexity, but that's again separate
> from the "zapping" of SPTEs.

They do keep popping up; I remember they're a pain in COLO.

> > Can you tell me which version of libvirt you're using?
> > I thought the newer ones were supposed to use events so they did't
> > have to poll qemu.
> > 
> >  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.
> 
> I don't like having such a long-lived mutex (it seems like a recipe for
> deadlocks with the BQL), plus memory_region_transaction_commit (the
> expensive part of memory_global_dirty_log_stop) needs to be under the
> BQL itself because it calls MemoryListeners.
> 
> Maybe memory_global_dirty_log_stop can delay itself to the next vm_start
> if it's called while runstate_running() returns false (which should be
> always the case)?
> 
> It could even be entirely enclosed within memory.c if you do it with a
> VMChangeStateHandler.

This still causes the BQL to be held for quite a while; albeit at a less
critical point.

In this and the existing case we don't actually need efficiency - what we need 
is just
not to be holding onto the BQL for so long;  could we do a less
efficient commit here, removing one region at a time, yielding the lock
and retaking it?


> Thanks,
> 
> Paolo
> 
> > Dave
> > 
> >> Thanks,
> >> Jay
> >>
> >>> However, we would need to be a bit careful of anything that needs
> >>> cleaning up before the source restarts on failure; I'm not sure of
> >>> the semantics of all the current things wired into save_cleanup.
> >>>
> >>> Dave
> >>>
> >>>
> >>>> Signed-off-by: Jay Zhou <address@hidden>
> >>>> ---
> >>>>   migration/migration.c | 16 +++++++++-------
> >>>>   qmp.c                 | 10 ++++++++++
> >>>>   2 files changed, 19 insertions(+), 7 deletions(-)
> >>>>
> >>>> diff --git a/migration/migration.c b/migration/migration.c
> >>>> index a0db40d..72832be 100644
> >>>> --- a/migration/migration.c
> >>>> +++ b/migration/migration.c
> >>>> @@ -1877,6 +1877,15 @@ static void *migration_thread(void *opaque)
> >>>>           if (qemu_file_get_error(s->to_dst_file)) {
> >>>>               migrate_set_state(&s->state, current_active_state,
> >>>>                                 MIGRATION_STATUS_FAILED);
> >>>> +            /*
> >>>> +             * The resource has been allocated by migration will be 
> >>>> reused in
> >>>> +             * COLO process, so don't release them.
> >>>> +             */
> >>>> +            if (!enable_colo) {
> >>>> +                qemu_mutex_lock_iothread();
> >>>> +                qemu_savevm_state_cleanup();
> >>>> +                qemu_mutex_unlock_iothread();
> >>>> +            }
> >>>>               trace_migration_thread_file_err();
> >>>>               break;
> >>>>           }
> >>>> @@ -1916,13 +1925,6 @@ static void *migration_thread(void *opaque)
> >>>>       end_time = qemu_clock_get_ms(QEMU_CLOCK_REALTIME);
> >>>>
> >>>>       qemu_mutex_lock_iothread();
> >>>> -    /*
> >>>> -     * The resource has been allocated by migration will be reused in 
> >>>> COLO
> >>>> -     * process, so don't release them.
> >>>> -     */
> >>>> -    if (!enable_colo) {
> >>>> -        qemu_savevm_state_cleanup();
> >>>> -    }
> >>>>       if (s->state == MIGRATION_STATUS_COMPLETED) {
> >>>>           uint64_t transferred_bytes = qemu_ftell(s->to_dst_file);
> >>>>           s->total_time = end_time - s->total_time;
> >>>> diff --git a/qmp.c b/qmp.c
> >>>> index b86201e..0e68eaa 100644
> >>>> --- a/qmp.c
> >>>> +++ b/qmp.c
> >>>> @@ -37,6 +37,8 @@
> >>>>   #include "qom/object_interfaces.h"
> >>>>   #include "hw/mem/pc-dimm.h"
> >>>>   #include "hw/acpi/acpi_dev_interface.h"
> >>>> +#include "migration/migration.h"
> >>>> +#include "migration/savevm.h"
> >>>>
> >>>>   NameInfo *qmp_query_name(Error **errp)
> >>>>   {
> >>>> @@ -200,6 +202,14 @@ void qmp_cont(Error **errp)
> >>>>       if (runstate_check(RUN_STATE_INMIGRATE)) {
> >>>>           autostart = 1;
> >>>>       } else {
> >>>> +        /*
> >>>> +         * Delay the cleanup to reduce the downtime of migration.
> >>>> +         * The resource has been allocated by migration will be reused
> >>>> +         * in COLO process, so don't release them.
> >>>> +         */
> >>>> +        if (runstate_check(RUN_STATE_POSTMIGRATE) && 
> >>>> !migrate_colo_enabled()) {
> >>>> +            qemu_savevm_state_cleanup();
> >>>> +        }
> >>>>           vm_start();
> >>>>       }
> >>>>   }
> >>>> --
> >>>> 1.8.3.1
> >>>>
> >>>>
> >>> --
> >>> Dr. David Alan Gilbert / address@hidden / Manchester, UK
> >>>
> >>> .
> >>>
> >>
> > --
> > Dr. David Alan Gilbert / address@hidden / Manchester, UK
> > 
> 
--
Dr. David Alan Gilbert / address@hidden / Manchester, UK