On 1.10.2024 23:30, Peter Xu wrote:
On Tue, Oct 01, 2024 at 10:41:14PM +0200, Maciej S. Szmigiero wrote:
On 30.09.2024 23:57, Peter Xu wrote:
On Mon, Sep 30, 2024 at 09:25:54PM +0200, Maciej S. Szmigiero wrote:
On 27.09.2024 02:53, Peter Xu wrote:
On Fri, Sep 27, 2024 at 12:34:31AM +0200, Maciej S. Szmigiero wrote:
On 20.09.2024 18:45, Peter Xu wrote:
On Fri, Sep 20, 2024 at 05:23:08PM +0200, Maciej S. Szmigiero wrote:
On 19.09.2024 23:11, Peter Xu wrote:
On Thu, Sep 19, 2024 at 09:49:10PM +0200, Maciej S. Szmigiero wrote:
On 9.09.2024 22:03, Peter Xu wrote:
On Tue, Aug 27, 2024 at 07:54:27PM +0200, Maciej S. Szmigiero wrote:
From: "Maciej S. Szmigiero" <maciej.szmigi...@oracle.com>
load_finish SaveVMHandler allows migration code to poll whether
a device-specific asynchronous device state loading operation had finished.
In order to avoid calling this handler needlessly the device is supposed
to notify the migration code of its possible readiness via a call to
qemu_loadvm_load_finish_ready_broadcast() while holding
qemu_loadvm_load_finish_ready_lock.
Signed-off-by: Maciej S. Szmigiero <maciej.szmigi...@oracle.com>
---
include/migration/register.h | 21 +++++++++++++++
migration/migration.c | 6 +++++
migration/migration.h | 3 +++
migration/savevm.c | 52 ++++++++++++++++++++++++++++++++++++
migration/savevm.h | 4 +++
5 files changed, 86 insertions(+)
diff --git a/include/migration/register.h b/include/migration/register.h
index 4a578f140713..44d8cf5192ae 100644
--- a/include/migration/register.h
+++ b/include/migration/register.h
@@ -278,6 +278,27 @@ typedef struct SaveVMHandlers {
int (*load_state_buffer)(void *opaque, char *data, size_t data_size,
Error **errp);
+ /**
+ * @load_finish
+ *
+ * Poll whether all asynchronous device state loading had finished.
+ * Not called on the load failure path.
+ *
+ * Called while holding the qemu_loadvm_load_finish_ready_lock.
+ *
+ * If this method signals "not ready" then it might not be called
+ * again until qemu_loadvm_load_finish_ready_broadcast() is invoked
+ * while holding qemu_loadvm_load_finish_ready_lock.
[1]
+ *
+ * @opaque: data pointer passed to register_savevm_live()
+ * @is_finished: whether the loading had finished (output parameter)
+ * @errp: pointer to Error*, to store an error if it happens.
+ *
+ * Returns zero to indicate success and negative for error
+ * It's not an error that the loading still hasn't finished.
+ */
+ int (*load_finish)(void *opaque, bool *is_finished, Error **errp);
The load_finish() semantics is a bit weird, especially above [1] on "only
allowed to be called once if ..." and also on the locks.
The point of this remark is that a driver needs to call
qemu_loadvm_load_finish_ready_broadcast() if it wants for the migration
core to call its load_finish handler again.
It looks to me vfio_load_finish() also does the final load of the device.
I wonder whether that final load can be done in the threads,
Here, the problem is that current VFIO VMState has to be loaded from the main
migration thread as it internally calls QEMU core address space modification
methods which explode if called from another thread(s).
Ahh, I see. I'm trying to make dest qemu loadvm in a thread too and yield
BQL if possible, when that's ready then in your case here IIUC you can
simply take BQL in whichever thread that loads it.. but yeah it's not ready
at least..
Yeah, long term we might want to work on making these QEMU core address space
modification methods somehow callable from multiple threads but that's
definitely not something for the initial patch set.
Would it be possible vfio_save_complete_precopy_async_thread_config_state()
be done in VFIO's save_live_complete_precopy() through the main channel
somehow? IOW, does it rely on iterative data to be fetched first from
kernel, or completely separate states?
The device state data needs to be fully loaded first before "activating"
the device by loading its config state.
And just curious: how large is it
normally (and I suppose this decides whether it's applicable to be sent via
the main channel at all..)?
Config data is *much* smaller than device state data - as far as I remember
it was on order of kilobytes.
then after
everything loaded the device post a semaphore telling the main thread to
continue. See e.g.:
if (migrate_switchover_ack()) {
qemu_loadvm_state_switchover_ack_needed(mis);
}
IIUC, VFIO can register load_complete_ack similarly so it only sem_post()
when all things are loaded? We can then get rid of this slightly awkward
interface. I had a feeling that things can be simplified (e.g., if the
thread will take care of loading the final vmstate then the mutex is also
not needed? etc.).
With just a single call to switchover_ack_needed per VFIO device it would
need to do a blocking wait for the device buffers and config state load
to finish, therefore blocking other VFIO devices from potentially loading
their config state if they are ready to begin this operation earlier.
I am not sure I get you here, loading VFIO device states (I mean, the
non-iterable part) will need to be done sequentially IIUC due to what you
said and should rely on BQL, so I don't know how that could happen
concurrently for now. But I think indeed BQL is a problem.
Consider that we have two VFIO devices (A and B), with the following order
of switchover_ack_needed handler calls for them: first A get this call,
once the call for A finishes then B gets this call.
Now consider what happens if B had loaded all its buffers (in the loading
thread) and it is ready for its config load before A finished loading its
buffers.
B has to wait idle in this situation (even though it could have been already
loading its config) since the switchover_ack_needed handler for A won't
return until A is fully done.
This sounds like a performance concern, and I wonder how much this impacts
the real workload (that you run a test and measure, with/without such
concurrency) when we can save two devices in parallel anyway; I would
expect the real diff is small due to the fact I mentioned that we save >1
VFIO devices concurrently via multifd.
Do you think we can start with a simpler approach?
I don't think introducing a performance/scalability issue like that is
a good thing, especially that we already have a design that avoids it.
Unfortunately, my current setup does not allow live migrating VMs with
more than 4 VFs so I can't benchmark that.
/me wonders why benchmarking it requires more than 4 VFs.
My point here was that the scalability problem will most likely get more
pronounced with more VFs.
But I almost certain that with more VFs the situation with devices being
ready out-of-order will get even more likely.
If the config space is small, why loading it in sequence would be a
problem?
Have you measured how much time it needs to load one VF's config space that
you're using? I suppose that's vfio_load_device_config_state() alone?
It's not the amount of data to load matters here but that these address
space operations are slow.
The whole config load takes ~70 ms per device - that's time equivalent
of transferring 875 MiB of device state via a 100 GBit/s link.
What's the downtime of migration with 1/2/4 VFs? I remember I saw some
data somewhere but it's not in the cover letter. It'll be good to mention
these results in the cover letter when repost.
Downtimes with the device state transfer being disabled / enabled:
4 VFs 2 VFs 1 VF
Disabled: 1783 ms 614 ms 283 ms
Enabled: 1068 ms 434 ms 274 ms
Will add these numbers to the cover letter of the next patch set version.
Thanks.
I'm guessing 70ms isn't a huge deal here, if your NIC has 128GB internal
device state to migrate.. but maybe I'm wrong.
It's ~100 MiB of device state per VF here.
Ouch..
I watched your kvm forum talk recording, I remember that's where I get that
128 number but probably get the unit wrong.. ok that makes sense.
And it's 70ms of downtime *per device*:
so with 4 VF it's ~280ms of downtime taken by the config loads.
That's a lot - with perfect parallelization this downtime should
*reduce by* 210ms.
Yes, in this case it's a lot. I wonder why it won't scale as good even
with your patchset.
Did you profile why? I highly doubt in your case network is an issue, as
there's only 100MB per-dev data, so even on 10gbps it takes 100ms only to
transfer for each, while now assuming it can run concurrently. I think you
mentioned you were using 100gbps, right?
Right, these 2 test machines are connected via a 100 GBbps network.
Logically when with multiple threads, VFIO read()s should happen at least
concurrently per-device. Have you checked that there's no kernel-side
global VFIO lock etc. that serializes portions of the threads read()s /
write()s on the VFIO fds?
For these devices the kernel side has been significantly improved a year ago:
https://lore.kernel.org/kvm/20230911093856.81910-1-yish...@nvidia.com/
In the mlx5 driver the in-kernel device reading task (work) is separated
from the userspace (QEMU) read()ing task via a double/multi buffering scheme.
If there was indeed some global lock serializing all device accesses we
wouldn't be seeing that much improvement from this patch set as we are
seeing - especially that the improvement seems to *increase* with the
increased VF count in a single PF.
It's just a pity that you went this far, added all these logics, but
without making it fully concurrent at least per device.
AFAIK NVIDIA/Mellanox are continuously working on improving the mlx5 driver,
but to benefit from the driver parallelism we need parallelism in QEMU
too so the userspace won't become the serialization point/bottleneck.
In other words, it's kind of a chicken and egg problem.
That's why I want to preserve as much parallelism in this patch set as
possible to avoid accidental serialization which (even if not a problem
right now) may become the bottleneck at some point.
I'm OK if you want this in without that figured out, but if I were you I'll
probably try to dig a bit to at least know why.
I also wonder whether you profiled a bit on how that 70ms contributes to
what is slow.
I think that's something we can do after we have parallel config loads
and it turns out their downtime for some reason still scales strongly
linearly with the number of VFIO devices (rather than taking roughly
constant time regardless of the count of these devices if running perfectly
in parallel).
Similarly, I wonder whether the config space load() can involves something
globally shared. I'd also dig a bit here, but I'll leave that to you to
decide.
Making config loads thread-safe/parallelizable is definitely on my future
TODO list.
Just wanted to keep the amount of changes in the first version of this
patch set within reasonable bounds - one has to draw a line somewhere
otherwise we'll keep working on this patch set forever, with the
QEMU code being a moving target meanwhile.
So what I'm thinking could be very clean is, we just discussed about
MIG_CMD_SWITCHOVER and looks like you also think it's an OK approach. I
wonder when with it why not we move one step further to have
MIG_CMD_SEND_NON_ITERABE just to mark that "iterable devices all done,
ready to send non-iterable". It can be controlled by the same migration
property so we only send these two flags in 9.2+ machine types.
Then IIUC VFIO can send config data through main wire (just like most of
other pci devices! which is IMHO a good fit..) and on destination VFIO
holds off loading them until passing the MIG_CMD_SEND_NON_ITERABE phase.
Starting the config load only on MIG_CMD_SEND_NON_ITERABE would (in addition
to the considerations above) also delay starting the config load until all
iterable devices were read/transferred/loaded and also would complicate
future efforts at loading that config data in parallel.
However I wonder whether we can keep it simple in that VFIO's config space
is still always saved in vfio_save_state(). I still think it's easier we
stick with the main channel whenever possible. For this specific case, if
the config space is small I think it's tricky you bypass this with:
if (migration->multifd_transfer) {
/* Emit dummy NOP data */
qemu_put_be64(f, VFIO_MIG_FLAG_END_OF_STATE);
return;
}
Then squash this as the tail of the iterable data.
On the src, I think it could use a per-device semaphore, so that iterable
save() thread will post() only if it finishes dumping all the data, then
that orders VFIO iterable data v.s. config space save().
In the future we want to not only transfer but also load the config data
in parallel.
How feasible do you think this idea is? E.g. does it involve BQL so far
(e.g. memory updates, others)? What's still missing to make it concurrent?
My gut feeling is that is feasible overall but it's too much of a rabbit
hole for the first version of this device state transfer feature.
I think it will need some deeper QEMU core address space management changes,
which need to be researched/developed/tested/reviewed/etc. on their own.
If it was an easy task I would have gladly included such support in this
patch set version already for extra downtime reduction :)
Yes I understand.
Note that it doesn't need to be implemented and resolved in one shot, but I
wonder if it'll still be good to debug the issue and know where is not
scaling.
Considering that your design is fully concurrent as of now on iterable data
from QEMU side, it's less persuasive to provide perf numbers that still
doesn't scale that much; 1.78s -> 1.06s is a good improvement, but it
doesn't seem to solve the scalability issue that this whole series wanted
to address in general.
An extreme (bad) example is if VFIO has all ioctl()/read()/write() take a
global lock, then any work in QEMU trying to run things in parallel will be
a vain. Such patchset cannot be accepted because the other issue needs to
be resolved first.
Now it's in the middle of best/worst condition, where it did improve but it
still doesn't scale that well. I think it can be accepted, but still I
feel like we're ignoring some of the real issues. We can choose to ignore
the kernel saying that "it's too much to do together", but IMHO the issues
should be tackled in the other way round.. the normal case is one should
work out the kernel scalability issues, then QEMU should be on top.. Simply
because any kernel change that might scale >1 device save()/load() can
affect future QEMU change and design, not vice versa.
Again, I know you wished we make some progress, so I don't have a strong
opinion. Just FYI.
As I wrote above, the kernel side of things are being taken care of by
the mlx5 driver maintainers.
And these performance numbers suggest that there isn't some global lock
serializing all device accesses as otherwise it would quickly become
the bottleneck and we would be seeing diminishing improvement from
increased VF count instead of increased improvement.
(..)
that I feel like perhaps can be replaced by a sem (then to drop the
condvar)?
Once we have ability to load device config state outside main migration
thread replacing "load_finish" handler with a semaphore should indeed be
possible (that's internal migration API so there should be no issue
removing it as not necessary anymore at this point).
But for now, the devices need to have ability to run their config load
code on the main migration thread, and for that they need to be called
from this handler "load_finish".
A sem seems a must here to notify the iterable data finished loading, but
that doesn't need to hook to the vmstate handler, but some post-process
tasks, like what we do around cpu_synchronize_all_post_init() time.
If per-device vmstate handler hook version of load_finish() is destined to
look as weird in this case, I'd rather consider a totally separate way to
enqueue some jobs that needs to be run after all vmstates loaded. Then
after one VFIO device fully loads its data, it enqueues the task and post()
to one migration sem saying that "there's one post-process task, please run
it in migration thread". There can be a total number of tasks registered
so that migration thread knows not to continue until these number of tasks
processed. That counter can be part of vmstate handler, maybe, reporting
that "this vmstate handler has one post-process task".
Maybe you have other ideas, but please no, let's avoid this load_finish()
thing..
I can certainly implement the task-queuing approach instead of the
load_finish() handler API if you like such approach more.
I have an even simpler solution now. I think you can reuse precopy
notifiers.
You can add one new PRECOPY_NOTIFY_INCOMING_COMPLETE event, invoke it after
vmstate load all done.
As long as VFIO devices exist, VFIO can register with that event, then it
can do whatever it wants in the main loader thread with BQL held.
You can hide that sem post() / wait() all there, then it's completely VFIO
internal. Then we leave vmstate handler alone; it just doesn't sound
suitable when the hooks need to be called out of order.
I can certainly implement this functionality via a new
precopy_notify(PRECOPY_NOTIFY_INCOMING_COMPLETE) notifier, for example
by having a single notify handler registered by the VFIO driver, which
handler will be common to all VFIO devices.
This handler on the VFIO driver side will then take care of proper operation
ordering between the existing VFIO devices.
- How qemu_loadvm_load_finish_ready_broadcast() interacts with all
above..
So if you really think it matters to load whatever VFIO device who's
iterable data is ready first, then let's try come up with some better
interface.. I can try to think about it too, but please answer me
questions above so I can understand what I am missing on why that's
important. Numbers could help, even if 4 VF and I wonder how much diff
there can be. Mostly, I don't know why it's slow right now if it is; I
thought it should be pretty fast, at least not a concern in VFIO migration
world (which can take seconds of downtime or more..).
IOW, it sounds more reasonalbe to me that no matter whether vfio will
support multifd, it'll be nice we stick with vfio_load_state() /
vfio_save_state() for config space, and hopefully it's also easier it
always go via the main channel to everyone. In these two hooks, VFIO can
do whatever it wants to sync with other things (on src, sync with
concurrent thread pool saving iterable data and dumping things to multifd
channels; on dst, sync with multifd concurrent loads). I think it can
remove the requirement on the load_finish() interface completely. Yes,
this can only load VFIO's pci config space one by one, but I think this is
much simpler, and I hope it's also not that slow, but I'm not sure.
To be clear, I made a following diagram describing how the patch set
is supposed to work right now, including changing per-device
VFIO_MIG_FLAG_DEV_DATA_STATE_COMPLETE into a common MIG_CMD_SWITCHOVER.
Time flows on it left to right (->).
----------- DIAGRAM START -----------
Source overall flow:
Main channel: live VM phase data -> MIG_CMD_SWITCHOVER -> iterable
-> non iterable
Multifd channels: \ multifd device state
read and queue (1) -> multifd config data read and queue (1) /
Target overall flow:
Main channel: live VM phase data -> MIG_CMD_SWITCHOVER -> iterable -> non iterable
-> config data load operations
Multifd channels: \ multifd device state (1)
-> multifd config data read (1)
Target config data load operations flow:
multifd config data read (1) -> config data load (2)
Notes:
(1): per device threads running in parallel
Here I raised this question before, but I'll ask again: do you think we can
avoid using a separate thread on dest qemu, but reuse multifd recv threads?
Src probably needs its own threads because multifd sender threads takes
request, so it can't block on its own.
However dest qemu isn't like that, it's packet driven so I think maybe it's
ok VFIO directly loads the data in the multifd threads. We may want to
have enough multifd threads to make sure IO still don't block much on the
NIC, but I think tuning the num of multifd threads should work in this
case.
We need to have the receiving threads decoupled from the VFIO device state
loading threads at least because otherwise:
1) You can have a deadlock if device state for multiple devices arrives
out of order, like here:
Time flows left to right (->).
Multifd channel 1: (VFIO device 1 buffer 2) (VFIO device 2 buffer 1)
Multifd channel 2: (VFIO device 2 buffer 2) (VFIO device 1 buffer 1)
Both channel receive/load threads would be stuck forever in this case,
since they can't load buffer 2 for devices 1 and 2 until they load
buffer 1 for each of these devices.
2) If devices are loading buffers at different speeds you don't want
to block the faster device from receiving new buffer just because
the slower one hasn't finished its loading yet.
I don't see why it can't be avoided. Let me draw this in columns.
How I picture this is:
multifd recv thread 1 multifd recv thread 2
--------------------- ---------------------
recv VFIO device 1 buffer 2 recv VFIO device 2 buffer 2
-> found that (dev1, buf1) missing, -> found that (dev2, buf1)
missing,
skip load skip load
recv VFIO device 2 buffer 1 recv VFIO device 1 buffer 1
-> found that (dev2, buf1+buf2) ready, -> found that (dev1, buf1+buf2)
ready,
load buf1+2 for dev2 here load buf1+2 for dev1 here
Here right after one multifd thread recvs a buffer, it needs to be injected
into the cache array (with proper locking), so that whoever receives a full
series of those buffers will do the load (again, with proper locking..).
Would this not work?
For sure but that's definitely more complicated logic than just having
a simple device loading thread that naturally loads incoming buffers
for that device in-order.
That thread isn't even in the purview of the migration code since
it's a VFIO driver internal implementation detail.
And we'd still lose parallelism if it happens that two buffers that
are to be loaded next for two devices happen to arrive in the same
multifd channel:
Multifd channel 1: (VFIO device 1 buffer 1) (VFIO device 2 buffer 1)
Multifd channel 2: (VFIO device 2 buffer 2) (VFIO device 1 buffer 2)
Now device 2 buffer 1 has to wait until loading device 1 buffer 1
finishes even thought with the decoupled loading thread implementation
from this patch set these would be loaded in parallel.
Thanks.
Thanks,
Maciej
