Re: [PATCH 08/13] migration: Thread-ify precopy vmstate load process

[Fabiano Rosas <farosas@suse.de>]

Peter Xu <peterx@redhat.com> writes:

> On Mon, Jan 12, 2026 at 04:04:12PM -0300, Fabiano Rosas wrote:
>> Peter Xu <peterx@redhat.com> writes:
>> 
>> > On Thu, Jan 08, 2026 at 05:27:37PM -0300, Fabiano Rosas wrote:
>> >> Peter Xu <peterx@redhat.com> writes:
>> >> 
>> >> > Migration module was there for 10+ years.  Initially, it was in most cases
>> >> > based on coroutines.  As more features were added into the framework, like
>> >> > postcopy, multifd, etc.. it became a mixture of threads and coroutines.
>> >> >
>> >> > I'm guessing coroutines just can't fix all issues that migration want to
>> >> > resolve.
>> >> >
>> >> > After all these years, migration is now heavily based on a threaded model.
>> >> >
>> >> > Now there's still a major part of migration framework that is still not
>> >> > thread-based, which is precopy load.  We do load in a separate thread in
>> >> > postcopy since the 1st day postcopy was introduced, however that requires a
>> >> > separate state transition from precopy loading all devices first, which
>> >> > still happens in the main thread of a coroutine.
>> >> >
>> >> > This patch tries to move the migration incoming side to be run inside a
>> >> > separate thread (mig/dst/main) just like the src (mig/src/main).  The
>> >> > entrance to be migration_incoming_thread().
>> >> >
>> >> > Quite a few things are needed to make it fly..  One note here is we need to
>> >> > change all these things in one patch to not break anything.  The other way
>> >> > to do this is add code to make all paths (that this patch touched) be ready
>> >> > for either coroutine or thread.  That may cause confusions in another way.
>> >> > So reviewers, please take my sincere apology on the hardness of reviewing
>> >> > this patch: it covers a few modules at the same time, and with some risky
>> >> > changes.
>> >> >
>> >> > BQL Analysis
>> >> > ============
>> >> >
>> >> > Firstly, when moving it over to the thread, it means the thread cannot take
>> >> > BQL during the whole process of loading anymore, because otherwise it can
>> >> > block main thread from using the BQL for all kinds of other concurrent
>> >> > tasks (for example, processing QMP / HMP commands).
>> >> >
>> >> > Here the first question to ask is: what needs BQL during precopy load, and
>> >> > what doesn't?
>> >> >
>> >> 
>> >> I just noticed that the BQL held at process_incoming_migration_co is
>> >> also responsible for stopping qmp_migrate_set_capabilities from being
>> >> dispatched.
>> >
>> > I don't know if it is by design, or even if it will be guaranteed to work..
>> >
>> 
>> Regardless, we shouldn't rely on the BQL for this. The BQL should be
>> left as last resort for things that interact across subsystems. If
>> someone is issuing a migration command during a migration, the migration
>> code is exquisitely positioned to handle that itself.
>
> Yes I agree.
>
>> 
>> > Consider the migration incoming rocoutine runs into qemu_get_byte(), and
>> > then proactively yield the migration coroutine (qemu_coroutine_yield())
>> > when the incoming port is blocked on read..
>> >
>> > AFAIU, a proper fix for that (note, this will currently break tests) is:
>> >
>> > bool migration_is_running(void)
>> >  {
>> > -    MigrationState *s = current_migration;
>> > +    MigrationStatus state;
>> >  
>> > -    if (!s) {
>> > -        return false;
>> > +    if (runstate_check(RUN_STATE_INMIGRATE)) {
>> > +        MigrationIncomingState *mis = migration_incoming_get_current();
>> > +
>> > +        if (!mis) {
>> > +            return false;
>> > +        }
>> > +
>> > +        state = mis->state;
>> > +    } else {
>> > +        MigrationState *s = migrate_get_current();
>> > +
>> > +        if (!s) {
>> > +            return false;
>> > +        }
>> > +
>> > +        state = s->state;
>> >      }
>> >  
>> > -    switch (s->state) {
>> > +    switch (state) {
>> >      case MIGRATION_STATUS_ACTIVE:
>> >      case MIGRATION_STATUS_POSTCOPY_DEVICE:
>> >      case MIGRATION_STATUS_POSTCOPY_ACTIVE:
>> >
>> 
>> LGTM
>> 
>> >> 
>> >> Any point during incoming migration when BQL is unlocked we have a
>> >> window where a capability could be changed. Same for parameters, for
>> >> that matter.
>> >> 
>> >> To make matters worse, the -incoming cmdline will trigger
>> >> qmp_migrate_incoming->...->migration_transport_compatible early on, but
>> >> until the channels finally connect and process_incoming_migration_co
>> >> starts it's possible to just change a capability in an incompatible way
>> >> and the transport will never be validated again.
>> >
>> > Right.  Above should fix it, but I believe it also means after "-incoming
>> > tcp:xxx" (or anything not "defer") we should forbid changing migration caps
>> > or params on destination.
>> >
>> 
>> Parameters are never forbidden, right? And we cannot forbid them with
>> is_running because some parameters are allowed to be changed while
>> running.
>
> Right, my above statement was definitely inaccurate.
>
> After merging caps and params we only have params.  We should only allow
> some params to be changed anytime.  Most of the params (including caps)
> should not allow changing during migration live on either src/dst.
>
>> 
>> I feel we should have a more fine grained way of saying "this option
>> cannot be set at this moment", instead of just using the state as a
>> proxy. States can change, while the fact that from a certain point on,
>> certain options should not be touched anymore doesn't change.
>
> IIUC that's what migration_is_running() about?
>

At a high level, yes, but I think that's actually a downside of
migration_is_running. It's not explicit.

E.g.: qmp_migrate -> migration_transport_compatible is the first time
capabilities are checked in the code. Guess what migration_is_running
returns at that point? (it's false)

If there's ever a change in BQL behavior, there'll be a bug and we'll
see another patch extending the scope of migration_is_running, either
adding existing states to it, or as Prasad proposed, adding a new state
that is set a little later/earlier.

What we technically want is to stop accepting new capabilities as soon
as we're about to use them for the first time. Not as soon as (state ==
ACTIVE || SETUP || ...) which is a more high level definition.

> It's a matter of if such finer granularity is necessary for us. Let's
> assume the simple scheme is:
>
>   (a) Some params are allowed to be modified anytime, examples,
>       downtime-limit, max-bandwidth, max-postcopy-bandwidth, etc.
>

What about during migration_cleanup() or migrate-cancel? Won't there
ever be something in the path of setting these params that would require
a "healthy" migration?

>   (b) All the rest params and all capabilities are not allowed to be modified
>       when migration_is_running() returns true (my fixed version above)
>
> The whitelisted (a) should really be the smallest set of params, and
> justified one by one or it should fall into (b).
>
> My hope is the simple scheme should work for us already.  After merging
> caps, it's only about some params that can be set anytime, rest params can
> only be set when migration is not running.
>

Sure, I'd just like to avoid adding another set of "if
(s->parameters->has_foobar)" if possible.

>> 
>> Maybe a little infra like bdrv_op_is_blocked, i.e, a list of blocked
>> operations. It could be set in qmp_migrate and checked in
>> qmp_set_parameters/caps.
>
> Any example you can think of, that above simple scheme won't work for us?
>

It works. I think it's fragile due to:

- the reliance on BQL to avoid concurrent invocations;

- the possible addition of new states for unrelated reasons
  risks regressions and it's just extra work to make sure everything is
  still blocked at the time it needs to be blocked.

- any bugs need to be fixed by moving states around, which is dubious
  since the migration flow itself (which the states should represent)
  has not changed in a while.

  We're still seeing patches such as dc487044d5 ("migration: Make
  migration_has_failed() work even for CANCELLING") and the recent patch
  from Prasad. So there is something changing elsewhere and exposing
  these bugs and it's not immediately clear what that something is (or
  else we would have avoided the bug).

To be clear, I'm not asking we solve these issues at this moment, it's
fine if we go forward with the simple scheme. I'm just trying to clarify
what I think the problems are.

>> 
>> > As discussed above, that'll at least break our qtests.  But frankly
>> > speaking I think that's the right thing to do..  I hope libvirt always
>> > works with "defer" and never update any caps/params after QMP
>> > migrate_incoming.
>> >
>> > So I wonder if I should continue with above patch, and then fix our qtests.
>> > Your work from the other "merge caps+params" might also work here,
>> > actually, if we make sure everything will be set alone with the QMP
>> > migrate_incoming single command.
>> >
>> 
>> For incoming, yes. And this is maybe a point in favor of adding the
>> 'config'.
>> 
>> For outgoing, there's still the point I mentioned above about how to
>> restrict _some_ options to be allowed at runtime and others not.
>> 
>> > Let me know your initial thoughts, then I'll see what I can do..
>> >
>> 
>> We should fix the bug, I think your patch is good for that.
>> 
>> Although this kind of overlaps with some things we've been discussing
>> with Prasad. I'd be super happy if the code magically stopped using
>> QAPI's MigrationStatus for internal tracking of migration state and
>> blocking of commands and so on.
>> 
>> Whatever comes first =)
>
> I didn't follow as closely on the discussion there.  I don't know if
> changing MigrationStatus is a good idea..  we should have some really good
> reason to make libvirt and all mgmt need a change.. but maybe I misread the
> discussion.  I can wait until I read something solid if Prasad is going to
> propose something.
>

We're moving towards decoupling MigrationStatus from these ordinary code
checks. MigrationStatus should be used for informing management about a
relevant change in the migration stage of progression.

As it stands, we're either "exposing implementation details" or
"inadvertently changing the API". Both are undesirable.

Of course, we might introduce the "internal state is off-sync with
externally visible states" issue, but hopefully we'll catch that during
review. =)

>> 
>> ---
>> Side note, did we ever discuss something like this?
>> 
>> struct MigrationState {
>>    <state>
>>    union {
>>      <outgoing>
>>      <incoming>
>>    }
>> }
>> 
>> there's so much stuff in these structs...
>
> Yeah..  When merging state, we'll also need to be careful on overlapped
> fields, e.g. expecting a COMPLETED state (from a completed incoming
> migration) when starting a new outgoing migration.
>
> We can likely leave this for later.

No worries, I'm just collecting your opinions on it.