---Problem Description---
A vhost performance patch was introduced in the 4.10 kernel upstream, and is currently included in the Zesty 4.10 kernel:
commit 809ecb9bca6a9424ccd392d67e368160f8b76c92
Author: Jason Wang <email address hidden>
Date: Mon Dec 12 14:46:49 2016 +0800
vhost: cache used event for better performance
--
However I recently hit a functional issue linked to this patch which would cause random guests to lose their network connection under stress. This is not architecture specific and more likely to be hit with high network stress (i.e. lots of uperf instances).
which reads:
"
Revert "vhost: cache used event for better performance"
This reverts commit 809ecb9bca6a9424ccd392d67e368160f8b76c92. Since it
was reported to break vhost_net. We want to cache used event and use
it to check for notification. The assumption was that guest won't move
the event idx back, but this could happen in fact when 16 bit index
wraps around after 64K entries.
Signed-off-by: Jason Wang <email address hidden>
Acked-by: Michael S. Tsirkin <email address hidden>
Signed-off-by: David S. Miller <email address hidden>
"
I am requesting this patch to revert the problematic one be pulled into Ubuntu Zesty (anything 4.10+).
---uname output---
Linux p82qvirt 4.10.0-32-generic #36~16.04.1-Ubuntu SMP Wed Aug 9 09:19:19 UTC 2017 ppc64le ppc64le ppc64le GNU/Linux
Machine Type = 8247-22L
---Steps to Reproduce---
I can recreate the scenario with the following setup:
- on a 20core host, start 20 1core VMs
- I have a single linux bridge assigned to all guests using virtio
- start a uperf benchmark between each guest pair (10 total) using a high number of uperf nprocs (32)
---Problem Description---
A vhost performance patch was introduced in the 4.10 kernel upstream, and is currently included in the Zesty 4.10 kernel:
commit 809ecb9bca6a942 4ccd392d67e3681 60f8b76c92
Author: Jason Wang <email address hidden>
Date: Mon Dec 12 14:46:49 2016 +0800
vhost: cache used event for better performance
--
However I recently hit a functional issue linked to this patch which would cause random guests to lose their network connection under stress. This is not architecture specific and more likely to be hit with high network stress (i.e. lots of uperf instances).
The patch author has now reverted this patch upstream: /git.kernel. org/pub/ scm/linux/ kernel/ git/torvalds/ linux.git/ commit/ drivers/ vhost?id= 8d65843c44269c2 1e95c98090d9bb4 848d473853
https:/
which reads: 4ccd392d67e3681 60f8b76c92. Since it
"
Revert "vhost: cache used event for better performance"
This reverts commit 809ecb9bca6a942
was reported to break vhost_net. We want to cache used event and use
it to check for notification. The assumption was that guest won't move
the event idx back, but this could happen in fact when 16 bit index
wraps around after 64K entries.
Signed-off-by: Jason Wang <email address hidden>
Acked-by: Michael S. Tsirkin <email address hidden>
Signed-off-by: David S. Miller <email address hidden>
"
I am requesting this patch to revert the problematic one be pulled into Ubuntu Zesty (anything 4.10+).
---uname output---
Linux p82qvirt 4.10.0-32-generic #36~16.04.1-Ubuntu SMP Wed Aug 9 09:19:19 UTC 2017 ppc64le ppc64le ppc64le GNU/Linux
Machine Type = 8247-22L
---Steps to Reproduce---
I can recreate the scenario with the following setup:
- on a 20core host, start 20 1core VMs
- I have a single linux bridge assigned to all guests using virtio
- start a uperf benchmark between each guest pair (10 total) using a high number of uperf nprocs (32)