Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2018-03-07
20:12:31 bauzas mriedem: just did
20:12:54 bauzas mriedem: I was looking at today's discussions
20:13:56 mriedem dansmith: can you +W https://review.openstack.org/#/c/550259/
20:14:00 mriedem that backport
20:14:01 artom mriedem, to be fair, maybe there just isn't that much to do in Nova for trusted VFs (if doing it without placement), in which case maybe it would have needed to be a Neutron spec?
20:14:15 mriedem artom: well, look at 'work items'
20:14:22 artom In any case, I still agree that we're better off tracking them in placement
20:14:23 mriedem "something something pci request spec, and libvirt driver, done!"
20:14:30 openstackgerrit Surya Seetharaman proposed openstack/nova-specs master: Support disabling a cell https://review.openstack.org/546684
20:14:41 mriedem don't look behind the curtain
20:14:54 artom But it's cloud!
20:15:07 artom You're not *supposed* to look behind the curtain
20:15:15 mriedem ew https://review.openstack.org/#/c/458513/11/nova/network/linux_net.py
20:15:36 mriedem ^ should happen in os-vif i'm pretty sure
20:15:39 artom The hell, we did that in nova-network?!
20:15:40 dansmith mriedem: yes
20:15:50 mriedem artom: that's common 'networking' code
20:15:57 artom Ah, sorry
20:16:31 mikal So are we deleting nova-network this release? Can I ignore that stuff for the purposes of privsep?
20:16:50 artom I think we said we're getting rid of cellsv1 first?
20:17:01 mriedem mikal: yes
20:17:07 mriedem artom: other way around
20:17:09 dansmith artom: other way around
20:17:10 dansmith heh
20:17:13 mriedem artom: you were there, in the room!
20:17:28 artom Oh so other way round?
20:17:31 mriedem so i think i get what the trusted VFs thing is now
20:17:31 mikal I'd like to be core for just the one review to +2 the removal of nova-net please. For old times sake.
20:17:43 mriedem 1. create port in neutron with a trusted attribute
20:17:56 mriedem 2. nova takes that off the port binding profile and shoves into the pci request spec,
20:18:19 artom (Oh right, your action item was to make cellsv1 work with Neutron)
20:18:26 mriedem 3. libvirt driver reads that off the pci request spec and shells out to set the vf to trusted on the host
20:18:42 mriedem why do we need the virt driver doing this?
20:18:49 dansmith mriedem: but in reality, it should be "2. nova takes that and shoves it into traits"
20:19:01 dansmith mriedem: then placement weeds everything we can't land on for us
20:19:18 mriedem dansmith: yeah that's what i was thinking the spec (and asked) - how to get go through scheduling to find a host that handles this type of port
20:19:23 mriedem since we don't have network-aware scheduling
20:19:33 mriedem jesus words
20:20:40 bauzas is efried still on PTO ?
20:21:10 bauzas I was considering to work on a spec for providing NUMA affinity using nested RPs, and AFAIR he told me there was already a spec
20:21:21 bauzas am I wrong?
20:21:40 mriedem he is
20:21:49 artom So this might be a question for neutron, but... In the case of trusted VFs (and passthrough of whole PFs), what the actual cable going from the card is connected to at the other end is kinda important
20:22:01 artom How is that expressed when creating the neutron port?
20:22:22 dansmith mriedem: as bauzas would say... "le woot": http://logs.openstack.org/94/550194/6/check/nova-next/0842c76/job-output.txt.gz#_2018-03-07_20_06_24_926495
20:22:38 artom Or put differently, how does the port become associated with the right hardware?
20:23:08 artom Maybe a question for sean tomorrow morning
20:23:36 bauzas dansmith: \o/
20:25:53 mriedem le woff
20:25:55 mriedem *woof
20:26:08 mriedem bauzas: artom: ok i dumped some placement thoughts in https://review.openstack.org/#/c/485522/
20:26:28 mriedem i'm sure the response will be, "well this 'works' with code we have up for review today and placement isn't done yet"
20:26:49 mriedem it doesn't, at all, address the scheduling issue of finding a host that allows nova to configure VFs as trusted
20:26:55 artom Which is *technically* true :/
20:27:02 dansmith I assume the pci filter can do that kind of thing,
20:27:10 dansmith if we extend the data structure to have a trusted flag or something
20:27:14 dansmith but that's the thing I think we should avoid
20:27:22 artom I thought that's what he did?
20:27:27 dansmith building more cruft on that crap is not the way we should go, IMHO
20:27:28 mriedem i don't really know anything about the pci filter
20:27:40 artom dansmith, https://review.openstack.org/#/c/458820/ ?
20:27:41 mriedem does the pci filter get information off the pci request spec?
20:27:56 dansmith artom: right, but that's not merged
20:27:58 bauzas dansmith: le woof is victory https://scontent-cdg2-1.xx.fbcdn.net/v/t1.0-9/28577622_10213176224494032_6122884744312777644_n.jpg?oh=e91e488a77ace375c978611b0964e159&oe=5B39D007
20:28:07 artom dansmith, right, but it's what's being proposed
20:28:12 dansmith bauzas: I said "le woot" not "le woof"
20:28:18 dansmith artom: right, that's what I was saying
20:28:26 artom dansmith, oh you were saying that? right
20:28:29 dansmith artom: mriedem sounded like he didn't see the connection for the scheduler
20:28:35 cdent that's an awesome dog
20:28:58 dansmith bauzas: my understanding is I can make any word french by putting "le" in front...
20:29:12 mriedem dansmith: artom: ok i didn't realize the connection with the pci filter
20:29:16 bauzas it can work with "le café"
20:29:22 bauzas "le budget"
20:29:25 artom dansmith, ah, and your response was "there's the connection, by having the PCI filter work with this new trusted thing in pci requests"
20:29:30 dansmith yes
20:29:58 bauzas yeah the spec was unclear, I asked for details
20:30:02 mriedem so the api calls create_pci_requests_for_sriov_ports which gets the binding profile off the port and gets the pci request spec with the new trusted attribute
20:31:15 mriedem ok and then the pci filter magically picks a host based on the pci request spec for a trusted vf, tied via flavor extra spec?
20:31:33 dansmith not just that, it picks an _actual_ pci device I think
20:31:35 dansmith it's AWESOME
20:33:34 mriedem PciPassthroughFilter right?
20:34:04 mriedem man i'm confused
20:34:20 mriedem good thing the release note would have documented how all of this works for a deployer that wants to use it
20:34:30 mriedem artom: i guess?
20:34:48 mriedem the spec just says, "operators will have to use host aggregates"
20:35:19 artom I sort of assumed it would be in the same way that we have pinned CPUs now
20:35:31 artom Aggregates for trusted VFs and aggregated for normal VFs
20:35:46 artom So that any VF in a 'trusted' aggregate is, err, trusted
20:38:27 dansmith that would kindof suck
20:39:40 mriedem the PciPassthroughFilter is looking at HostState.pci_stats, but i don't know what goes into that thing saying "this host reported it supports trusted VFs"
20:40:09 openstackgerrit Merged openstack/nova master: Only attempt a rebuild claim for an evacuation to a new host https://review.openstack.org/546268
20:42:13 mriedem i guess https://review.openstack.org/#/c/458514/13/releasenotes/notes/trusted-vfs-abee6dff7c9b6940.yaml ?
20:42:27 mriedem so the pci passthrough_whitelist has to report the devices that are trusted,
20:42:37 mriedem and the PciPassthroughFilter links those to the ports in the pci request spec?
20:43:25 mriedem and pci passthrough_whitelist is our poor man's placement inventory system yes?
20:43:26 mriedem jaypipes: ^?
20:43:47 openstackgerrit Dan Smith proposed openstack/nova master: Make nova-manage db purge take --all-cells https://review.openstack.org/550502
20:44:52 dansmith mriedem: this ^ passed, so I converted the hook to run --all-cells on the purge
20:45:20 dansmith all the other patches have a clean swath of green from zuul too

Earlier   Later