| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2025-11-12 | |||
| 13:29:38 | gibi | `with threading.Lock()` is wrong while `with Guard(names)` is the correct way is non intuitive | |
| 13:29:49 | gibi | hence we need documentation | |
| 13:29:59 | sean-k-mooney | right | |
| 13:30:03 | sean-k-mooney | that the delta in usage | |
| 13:30:26 | sean-k-mooney | i need to push a patch to cyborg and venos to fix a requriement issue that i just fixed in watcher | |
| 13:30:38 | sean-k-mooney | but when im done ill update the patch again. thanks for reviewing | |
| 13:31:27 | gibi | cool thanks | |
| 14:21:53 | opendevreview | Vasyl Saienko proposed openstack/nova master: Set vnic model and queue for VDPA https://review.opendev.org/c/openstack/nova/+/966855 | |
| 14:39:42 | opendevreview | Vasyl Saienko proposed openstack/nova master: Set vnic model and queue for VDPA https://review.opendev.org/c/openstack/nova/+/966855 | |
| 14:40:50 | vsaienko | sean-k-mooney: I found that queues parameter is not passed to domain xml interface section in case of VDPA, this was a reason of slow iperf test. I've reported a bug https://bugs.launchpad.net/nova/+bug/2131150 and created a fix https://gerrit.mcp.mirantis.com/c/packaging/sources/nova/+/247944 please add to your review queue | |
| 14:41:06 | sean-k-mooney | vsaienko: yep i just -1 your patch | |
| 14:41:17 | sean-k-mooney | vsaienko: this si not a bug | |
| 14:41:28 | sean-k-mooney | it was not possibel to use multi queu whne we added vdpa supprot | |
| 14:41:37 | sean-k-mooney | and it was not implemented intentionally as a result | |
| 14:42:03 | vsaienko | hm... but now its possible | |
| 14:42:09 | sean-k-mooney | so 1 this is a new feature, two if we are got supprot this we will need much more testing the what is in the patch and mroe docs as well | |
| 14:42:28 | sean-k-mooney | vsaienko: rigth so as this is a feature requst it woudl normally need a bluefpirnt or a spec | |
| 14:42:53 | sean-k-mooney | we coudl do it under a bug in very limited cases but its not actully a bug from an upstream point of view | |
| 14:42:54 | vsaienko | ack | |
| 14:43:12 | sean-k-mooney | the main gaps are when did this supprot land in the kernel and qemu | |
| 14:43:37 | sean-k-mooney | we need to document that both in a release note and in the vdpa docs | |
| 14:43:48 | sean-k-mooney | depending o nthe qemu veriosn we may also need to add a version check | |
| 14:44:51 | sean-k-mooney | vsaienko: have you tested packed ring format ? | |
| 14:45:20 | sean-k-mooney | vsaienko: one of the open quetion too | |
| 14:45:37 | sean-k-mooney | is what happen if you have more cpu then the VF that backs the vdpa device supprot in hardware queue | |
| 14:46:09 | sean-k-mooney | i.e. is it ok for the number of queue we pass to qemu to exceed or even not match exactly the number of queue supproted by the hardware | |
| 14:48:15 | sean-k-mooney | vsaienko: if we can answer those quetions and ensure there is no upgrade impact when live migrating between hsot with and without this supprot we may be fine to proceed with a specless bluepritn but those quetions are why this woudl normally require a spec | |
| 14:49:38 | sean-k-mooney | for example it is poosble to set the queue for each vdpa device when they are created https://github.com/os-net-config/os-net-config/commit/384f46f8e0dfb847ab543e82f8cd08f5065edceb#diff-f43ec4c1cd48d8d241917be63e57e70516ef2bbba6e9ba786ce2ad5f1a92ca7dR628-R642 | |
| 14:50:01 | sean-k-mooney | now we dont offically suprpot that in nova but we also dont not supprot it | |
| 14:50:36 | sean-k-mooney | in the same way the number of quese for a sriov vf is determiend by how you allocate teh vf | |
| 14:51:20 | sean-k-mooney | that is howe it work today for vdpa as well so we need to make sure we dont break that flow if you dont enabel multi queu vis teh flavor. | |
| 14:52:31 | sean-k-mooney | vsaienko: sorry that a bit of an info dump but you can see why this is non trivial to do corectly and why the documenation of this really matters | |
| 14:53:10 | vsaienko | I haven't check packed ring format | |
| 14:53:51 | vsaienko | is what happen if you have more cpu then the VF that backs the vdpa device supprot in hardware queue - probably this is always a case when we have no multiqueue, VM has always more CPU and queues. | |
| 14:53:58 | sean-k-mooney | ack you change enabls opting in to that | |
| 14:54:36 | sean-k-mooney | vsaienko: in nova we set the queu count equal to the number of cpus when you enabel multi queue | |
| 14:54:55 | sean-k-mooney | so the question is fi the vdpa device only supprot 4 queue and you set it to 16 in qemu/libvirt what happens | |
| 14:55:14 | vsaienko | if the queue count in xml is higher than on VDPA device VM will see number that is configured on vdpa side | |
| 14:55:21 | vsaienko | I checked this | |
| 14:55:59 | sean-k-mooney | ack so we woudl need to 1 agree thats ok, it shoudl be and 2 docuemnt that so OPs are not surpised by that | |
| 14:56:45 | sean-k-mooney | we have a config option to clamp the number of queue so it is already posibel today for the vcpu count and queu count to be diffent even with muliqueueu enebled | |
| 14:57:29 | sean-k-mooney | the reason i asked about packed form at is you can now set it here https://review.opendev.org/c/openstack/nova/+/966855/2/nova/virt/libvirt/vif.py#206 | |
| 14:58:19 | sean-k-mooney | by the way did you intenilaly remove this https://review.opendev.org/c/openstack/nova/+/966855/1/nova/virt/libvirt/config.py form your v2? | |
| 15:06:51 | vsaienko | I remove because driver_name is set not to viritio there, and that config.py change is not needed | |
| 15:07:20 | sean-k-mooney | ack | |
| 15:10:40 | sean-k-mooney | vsaienko: so it looks like the libviert supprot was added by https://lists.libvirt.org/archives/list/devel@lists.libvirt.org/thread/D5QUZT2ZCUKF2526EOLGCB5JWCHHTIRV/?sort=thread and qemu by https://patchew.org/QEMU/20210903091031.47303-1-jasowang@redhat.com/ | |
| 15:13:02 | sean-k-mooney | https://github.com/qemu/qemu/commit/402378407dbdce79ce745a13f5c84815f929cfdd | |
| 15:13:24 | sean-k-mooney | so that is qemu v6.2.0 and later | |
| 15:14:06 | vsaienko | yes, looks like its supported now | |
| 15:14:49 | vsaienko | and need to check mtu, it seems we need to set it on xml as well | |
| 15:15:03 | vsaienko | because right now I can't use more than 1500 inside guest | |
| 15:15:16 | sean-k-mooney | we do suport settign the mtu i think | |
| 15:15:29 | sean-k-mooney | alhtogh that is one fo the thing that will conflict with the vdpa tool | |
| 15:16:18 | sean-k-mooney | our min qemu is currently 6.2.0 https://github.com/openstack/nova/blob/master/nova/virt/libvirt/driver.py#L221 | |
| 15:16:28 | sean-k-mooney | so we dont have to check if that supprot vdpa multi queue | |
| 15:16:46 | sean-k-mooney | but we do have to confirm whne the libvirt supprot actully hiped and if its in 8.0.0 | |
| 15:17:53 | vsaienko | libvirt change is in 8x version https://bugzilla.redhat.com/show_bug.cgi?id=2024406 | |
| 15:18:12 | sean-k-mooney | ya i was looking at that but that 8.2.0 | |
| 15:19:38 | vsaienko | we do suport settign the mtu i think - does it mean that we can't use jumboframes with vdpa? | |
| 15:20:42 | vsaienko | on 8 libvirt indeed setting mtu for vdpa is not possible https://paste.openstack.org/show/bi1VqvW3ScWE5obNpaRu/ | |
| 15:20:45 | sean-k-mooney | so the mtu will only be set if the mtu on the neutorn netowrk is not 1500 | |
| 15:21:10 | sean-k-mooney | i tought https://issues.redhat.com/browse/RHEL-7298 refence both mac and mtu but apprently not | |
| 15:22:11 | vsaienko | yes they closed https://bugzilla.redhat.com/show_bug.cgi?id=2057748 as duplicate but forgot to implement? | |
| 15:22:36 | vsaienko | on nova side mtu is not configured for vdpa device at all | |
| 15:22:50 | vsaienko | even when neutron network has > 1500 | |
| 15:22:51 | sean-k-mooney | ack the its beacuse of that orgial limiation | |
| 15:23:12 | sean-k-mooney | right so i dont know if that was also fixed in libvirt/qemu or not | |
| 15:25:29 | sean-k-mooney | vsaienko: so ya https://github.com/libvirt/libvirt/commit/a5e659f071ae5f5fc9aadb46ad7c31736425f8cf | |
| 15:25:42 | sean-k-mooney | libvirt supprot for mutlqueue came in 8.2.0 | |
| 15:26:04 | sean-k-mooney | that means we need to either rais our min version to libvirt 10 which is declared as or next min version | |
| 15:26:24 | sean-k-mooney | or you woudl have to support libvirt 8.0.0 by checkign the local version before enabling it | |
| 15:26:54 | sean-k-mooney | bumping the min version is the better chocies btu mroe work | |
| 15:27:18 | sean-k-mooney | we now know the minium requiremts at lesat qemu:6.2.0 and libvirt:8.2.0 | |
| 15:27:37 | sean-k-mooney | that woudl need to be captured in the release notes | |
| 15:27:52 | vsaienko | ack | |
| 15:29:02 | vsaienko | do we need additional check for libvirt version, because right now min libvirt is 8.0.0 and it is supported starting from 8.2.0 | |
| 15:29:34 | sean-k-mooney | yep i sasid that while you disconencted | |
| 15:29:51 | vsaienko | ack | |
| 15:29:57 | sean-k-mooney | we either need to supprot 8.0.0 via a version check ro bump our min version | |
| 15:30:06 | sean-k-mooney | whcih we can also do this cycle | |
| 15:30:31 | sean-k-mooney | its a littel more work but we have 1 or 2 feature that it woudl simplfy | |
| 15:30:34 | bauzas | sean-k-mooney: gibi: I can join the eventlet meeting but honestly I don't have the context | |
| 15:34:17 | vsaienko | should I add limitations section to and add mtu limitation here? https://docs.openstack.org/nova/latest/admin/vdpa.html | |
| 15:39:43 | sean-k-mooney | https://bugs.launchpad.net/nova/+bug/2119114 | |
| 16:58:01 | gibi | sean-k-mooney: dansmith: bauzas: Why does it make sense to use a named lock instance in the provide tree? https://github.com/openstack/nova/blob/b7d50570c7a79a38b0db6476ccb3c662b237f69b/nova/compute/provider_tree.py#L252 It was there from the beginning added by Jay but it does not make sense to me. This means that any instances of the ProviderTree class will share the same named lock instead of have | |
| 16:58:07 | gibi | its own lock. | |
| 16:59:12 | dansmith | the providertree is always for a single compute provider root in placement (per process) right? | |
| 17:00:16 | gibi | we are deep.copying ProviderTree objects around within the same process | |
| 17:00:34 | dansmith | right but they represent the same provider (tree) in placement right? | |
| 17:00:38 | gibi | but a single ProviderTree represents a single compute | |
| 17:01:21 | gibi | hm | |
| 17:01:23 | gibi | maybe ironic | |
| 17:01:31 | gibi | working with multiple nodes per compute service | |
| 17:01:56 | gibi | has multiple trees one for each node | |
| 17:02:39 | gibi | the ProviderTree class allows multiplre roots for some reason :/ | |
| 17:04:11 | bauzas | yeah I remember sometimes before we said we could have multiple roots | |
| 17:04:19 | bauzas | at least jay said it :p | |
| 17:04:59 | bauzas | I can't unfortunately remember the usecase where we could have multiple roots per compute | |
| 17:05:17 | bauzas | probably shared storage or something like that | |
| 17:06:22 | gibi | OK multiple roots is because the a single ProviderTree will have all the root providers that are connected to our compute root via placement aggregates | |