On Fri, Oct 06, 2023 at 09:58:30AM +0100, Joao Martins wrote:
On 03/10/2023 15:01, Michael S. Tsirkin wrote:
On Wed, Sep 27, 2023 at 12:14:28PM +0100, Joao Martins wrote:
On setups with one or more virtio-net devices with vhost on,
dirty tracking iteration increases cost the bigger the number
amount of queues are set up e.g. on idle guests migration the
following is observed with virtio-net with vhost=on:
48 queues -> 78.11% [.] vhost_dev_sync_region.isra.13
8 queues -> 40.50% [.] vhost_dev_sync_region.isra.13
1 queue -> 6.89% [.] vhost_dev_sync_region.isra.13
2 devices, 1 queue -> 18.60% [.] vhost_dev_sync_region.isra.14
With high memory rates the symptom is lack of convergence as soon
as it has a vhost device with a sufficiently high number of queues,
the sufficient number of vhost devices.
On every migration iteration (every 100msecs) it will redundantly
query the *shared log* the number of queues configured with vhost
that exist in the guest. For the virtqueue data, this is necessary,
but not for the memory sections which are the same. So
essentially we end up scanning the dirty log too often.
To fix that, select a vhost device responsible for scanning the
log with regards to memory sections dirty tracking. It is selected
when we enable the logger (during migration) and cleared when we
disable the logger.
The real problem, however, is exactly that: a device per vhost worker/qp,
when there should be a device representing a netdev (for N vhost workers).
Given this problem exists for any Qemu these days, figured a simpler
solution is better to increase stable tree's coverage; thus don't
change the device model of sw vhost to fix this "over log scan" issue.
Signed-off-by: Joao Martins <joao.m.martins@oracle.com>
---
I am not fully sure the heuristic captures the myriad of different vhost
devices -- I think so. IIUC, the log is always shared, it's just whether
it's qemu head memory or via /dev/shm when other processes want to
access it.
Thanks for working on this.
I don't think this works like this because different types of different
vhost devices have different regions - see e.g. vhost_region_add_section
I am also not sure all devices are running at the same time - e.g.
some could be disconnected, and vhost_sync_dirty_bitmap takes this
into account.
Good point. But this all means logic in selecting the 'logger' to take into
considering whether vhost_dev::log_enabled or vhost_dev::started right?
With respect to regions it seems like this can only change depending on whether
one of the vhost devices, backend_type is VHOST_BACKEND_TYPE_USER *and* whether
the backend sets vhost_backend_can_merge?
With respect to 'could be disconnected' during migration not devices can be
added or removed during migration, so might not be something that occurs during
migration.
I placed this in log_sync exactly to just cover migration, unless
there's some other way that disconnects the vhost and changes these variables
during migration.
The *frontend* can't be added or removed (ATM - this is just because we lack
good ways to describe devices that can be migrated, so all we
came up with is passing same command line on both sides,
and this breaks if you add/remove things in the process).
We really shouldn't bake this assumption into code if we can
help it though.
But I digress.
The *backend* can disconnect at any time as this is not guest visible.
But the idea is I think a good one - I just feel more refactoring is
needed.
Can you expand on what refactoring you were thinking for this fix?
Better separate the idea of logging from device. then we can
have a single logger that collects data from devices to decide
what needs to be logged.
My thinking on this bug was mostly to address the inneficiency with the smallest
intrusive fix (if at all possible!) given that virtually all multiqueue vhost
supported QEMU have this problem. And then move into a 'vhost-device for all
queues' as it feels like the problem here is the 'device per queue pair' doesn't
scale.
At the end of the day the problem on this is the vhost object model in log_sync
not scaling to amount of queues. But you could also argue that if the log is
shared that you can just log once for all, plus another one for each deviation
of normal behaviour, like the points you made in the earlier paragraph, and thus
the thinking behind this patch would still apply?
The thinking is good, but not the implementation.