On 28/02/2024 5:04, Peter Xu wrote:
External email: Use caution opening links or attachments
On Wed, Feb 28, 2024 at 02:00:26AM +0200, Avihai Horon wrote:
On 27/02/2024 9:41, Peter Xu wrote:
External email: Use caution opening links or attachments
On Thu, Feb 22, 2024 at 05:56:27PM +0200, Avihai Horon wrote:
Currently, migration code serializes device data sending during pre-copy
iterative phase. As noted in the code comment, this is done to prevent
faster changing device from sending its data over and over.
Frankly speaking I don't understand the rational behind 90697be889 ("live
migration: Serialize vmstate saving in stage 2"). I don't even think I
noticed this logic before even if I worked on migration for a few years...
I was thinking all devices should always get its chance to run for some
period during iterations. Do you know the reasoning behind? And I must
confess I also know little on block migration, which seems to be relevant
to this change. Anyway, I also copy Jan just in case he'll be able to chim
in.
I am not 100% sure either, but I can guess:
This commit is pretty old (dates to 2009), so maybe back then the only
iterative devices were block devices and RAM.
Block devices didn't change as fast as RAM (and were probably much bigger
than RAM), so it made sense to send them first and only then send RAM, which
changed constantly.
Makes sense. For some reason I read it the other way round previously.
If there is a fast changing device, even if we don't proceed with other
device iterators and we stick with the current one, assuming it can finally
finish dumping all data, but then we'll proceed with other devices and the
fast changing device can again accumulate dirty information?
I guess this logic only makes sense for the case where we only have block
devices and a RAM device, because the block devices wouldn't change that
much?
However, with switchover-ack capability enabled, this behavior can be
problematic and may prevent migration from converging. The problem lies
in the fact that an earlier device may never finish sending its data and
thus block other devices from sending theirs.
Yes, this is a problem.
This bug was observed in several VFIO migration scenarios where some
workload on the VM prevented RAM from ever reaching a hard zero, not
allowing VFIO initial pre-copy data to be sent, and thus destination
could not ack switchover. Note that the same scenario, but without
switchover-ack, would converge.
Fix it by not serializing device data sending during pre-copy iterative
phase if switchover was not acked yet.
I am still not fully convinced that it's even legal that one device can
consume all iterator's bandwidth, ignoring the rest.. Though again it's
not about this patch, but about commit 90697be889.
Yes, I agree. As I wrote above, maybe this was valid back then when the only
iterative devices were block and RAM.
I'm thinking whether we should allow each device to have its own portion of
chance to push data for each call to qemu_savevm_state_iterate(),
irrelevant of vfio's switchover-ack capability.
I wasn't sure for 100% why we have this logic in the first place, so I wrote
my patch as little invasive as possible, keeping migration behavior as is
(except for switchover-ack).
But I tend to agree with you for three reasons:
1. I think block migration is deprecated (see commits 66db46ca83b8,
40101f320d6b and 8846b5bfca76).
Instead, users are instructed to use blockdev-mirror and co. [1]. If I'm not
mistaken, this operates using a different infrastructure than migration.
So this logic is not relevant anymore.
2. As you pointed out earlier, the fast changing device can accumulate dirty
data over and over. VFIO devices come after RAM, so this logic doesn't
achieve its goal in this case (we may sync fast changing RAM over and over).
3. My fix in this patch won't solve a similar problem that could happen,
where a VFIO device with a lot of pre-copy data (not necessarily initial
data) may never be able to send it, thus not realizing the full potential of
pre-copy for VFIO.
(I personally have not encountered this problem yet, but maybe it can happen
with a vGPU).
Thanks for a summary.
If you agree, I can send a v2 that simply removes this logic and gives every
device an equal chance to send its data (like Joao showed) with some
explanation why we do it.
Let's see whether others have an opinion, but to me I think we can give it
a shot. In that case we can "break" in the previous "ret < 0" check
already.
Sure.
So I will wait some more and may send v2 next week, if no special
feedback is received.
One more thing to mention is then I think we need to calculate the case of
"all iterators returned 1" (aka, "all completes") scenario. With the old
check it is guaranteed if the loop iterates over all iterators then all
iterators have completed. Now we allow ret==0 to keep iterating, then the
check needs to be done explicitly.
In general, it could be something like:
all_done = 1;
loop {
...
ret = se->ops->save_live_iterate(f, se->opaque);
if (ret < 0) {
/* error handling.. */
...
break;
} else if (ret == 0) {
all_done = 0;
}
}
return all_done;
Yes, this looks good.
We could also give RAM precedence over other devices only during the first
iteration of sending RAM (i.e., only until first dirty sync), but I don't
know how much benefit this would give.
That sounds a bit tricky. We can leave that for later until justified to
be anything useful.
I agree.
Now when I checked VFIO code I found that VFIO still may have two issues:
====
1) VFIO doesn't yet respect migration_rate_exceeded()
That's the knob we use to limit send throughput when user specified the bw
to be something not zero, then throttling will apply. Currently it seems
VFIO always sends limited data per iteration - a little bit more than
1MB-ish?
Correct.
I'm mostly only looking at VFIO_MIG_DEFAULT_DATA_BUFFER_SIZE,
which seems fine, because in qemu_savevm_state_iterate() there's one more
migration_rate_exceeded() check anyway:
if (migration_rate_exceeded(f)) {
return 0;
}
However it may be good that VFIO will also use migration_rate_exceeded()
inside the iterator (before we can have some better way to do throttling..).
You mean that you want to add a migration_rate_exceeded() check inside
vfio_save_iterate() before we send any data?
E.g.:
static int vfio_save_iterate(QEMUFile *f, void *opaque)
{
VFIODevice *vbasedev = opaque;
VFIOMigration *migration = vbasedev->migration;
ssize_t data_size;
+ if (migration_rate_exceeded()) {
+ // return early
+ }
data_size = vfio_save_block(f, migration);
if (data_size < 0) {
return data_size;
}
Or check migration_rate_exceeded() not only here, but also during
sending of the 1MB buffer (inside vfio_save_block())?
Either way, I think this would only complicate things for no real
benefit: even if we do exceed BW, it would only be by 1MB max, and IMHO
this shouldn't make any real difference.
2) vfio_save_iterate() always return 1
IIUC right now "1" means "all data is sent"? But that seems not true for
VFIO if returned 1 every time.
I think it's not a major issue, e.g. the main migration routine even
ignored the retval of qemu_savevm_state_iterate(), the rest VFIO data
should rely on the upcoming loops over ->save_live_complete_precopy().
However I think it could already break bg_migration_iteration_run() as it
doesn't do that (it stores device state _before_ normal iterators). But I
agree bg-snapshot is never ready with VFIO anyway and the feature itself is
probably still flaky, so not a major concern. Just want to raise it up
when we're there.
Sure.
I think this "always return 1" in vfio_save_iterate() was a small hack
to avoid one VFIO device blocking the others due to
qemu_savevm_state_iterate() serialization behavior.
Now with the serialization removed, I can add a patch that changes the
"always return 1" to a proper return value based on remaining data left.
====
Besides all above, I think we are just still not good enough to do proper
bandwidth provisioning over the iterators if the max-bandwidth is set
pretty low: even with above change to allow iterators to proceed with
ret==0, we can also hit the case where RAM is very busy, it consumes all
bandwidth, and migration_rate_exceeded() reported true, then VFIO will be
skipped again as it's after RAM iterators. We may at least want to
remember the last time we iterated here, and then for the next iteration we
go with the handler next to it. But that should not happen with high /
unlimited bandwidth setup, and can be another story.
Indeed.
As you said, I think it's more of a problem with low BW. If it's fine by
you, this can be addressed later if we see real scenarios encountering
this situation.
Thanks.
