| Posted | Nick | Remark | |
|---|---|---|---|
| #openstack-nova - 2023-04-17 | |||
| 09:27:52 | bauzas | https://github.com/openstack/nova/blob/master/nova/virt/libvirt/volume/fs.py#L113 | |
| 09:27:55 | bauzas | found it | |
| 09:28:51 | bauzas | https://github.com/openstack/nova/blob/master/nova/virt/libvirt/volume/mount.py#L262 this is what's eventually called | |
| 09:29:20 | bauzas | which indeed calls a prevsep helper https://github.com/openstack/nova/blob/master/nova/virt/libvirt/volume/mount.py#L308 | |
| 09:29:53 | bauzas | and yeah look https://github.com/openstack/nova/blob/master/nova/privsep/fs.py#L30 | |
| 09:33:01 | bauzas | oh wait | |
| 09:33:25 | Uggla | bauzas, yes and as gibi said there is no link with the context. That's the reason why I don't understand the issue. | |
| 09:33:37 | bauzas | Uggla: as a reminder, the privsep helper doesn't run as root | |
| 09:34:49 | bauzas | it runs as the nova user with sudo rights, which is different | |
| 09:36:43 | bauzas | iirc, rootwrap only runs privsep with https://github.com/openstack/nova/blob/master/etc/nova/rootwrap.d/compute.filters | |
| 09:37:14 | bauzas | and then the privsep client connects to the privsep socket | |
| 09:37:58 | bauzas | that's why I'm suspecting you miss some privsep capability | |
| 09:39:37 | Uggla | bauzas, hum you lost me. As privsep is a replacement for rootwrap in my mind. | |
| 09:41:31 | bauzas | Uggla: privsep is a deamon https://github.com/openstack/oslo.privsep/blob/master/oslo_privsep/daemon.py | |
| 09:41:54 | bauzas | and iirc, we start it thru rootwrap https://github.com/openstack/oslo.privsep/blob/master/oslo_privsep/daemon.py#L32 | |
| 09:42:23 | bauzas | then, the privsep client connects to the deamon which has escalated rights | |
| 09:42:24 | gibi | bauzas: if the same sequence works with and admin user token but not with the demo user token then I don't think this is a privsep capability issue but I have no better idea :/ | |
| 09:42:45 | bauzas | gibi: I haven't understand the same | |
| 09:43:10 | bauzas | I thought Uggla was saying that the mount wasn't working with a nova user, but a mount command was working fine with sudo righrs | |
| 09:43:35 | bauzas | hence me thinking this was about a missing cap | |
| 09:44:12 | bauzas | but agreed, if an admin API call can mount thru nova, then yeah, this is a keystone context problem and not a privsep one | |
| 09:46:37 | Uggla | bauzas, the sudo mount command was just to prove that it is not an issue with missing directory and that mount.nfs is possible on the host. And it works fine as an os admin and not as a os user. | |
| 09:46:50 | bauzas | ohok | |
| 09:46:54 | bauzas | my bad then | |
| 09:47:03 | bauzas | soooo, pdb it | |
| 09:48:24 | stephenfin | sqlalchemy-migrate is finally gone. Whoop! 🥳 | |
| 09:48:24 | stephenfin | sqlalchemy-migrate is finally gone. Whoop! 🥳 | |
| 09:48:31 | bauzas | Uggla: have you checked that all the attributes for the privsep.mount call are the same, between a standard user and an admin user? | |
| 09:48:32 | stephenfin | Now to get the other projects to do the same | |
| 09:48:40 | stephenfin | Thanks for the reviews :) | |
| 09:48:43 | bauzas | stephenfin: kudos for the win. | |
| 09:49:11 | stephenfin | I hope nova jobs will start passing here shortly https://review.opendev.org/c/openstack/requirements/+/879743 | |
| 09:49:12 | Uggla | bauzas, unless if I'm wrong we are calling the same. | |
| 09:50:09 | Uggla | bauzas, if you wish I can show you a demo in the beginning of the afternoon. | |
| 09:50:28 | bauzas | Uggla: I've seen a couple of log debug messages, should be quick to verify | |
| 09:50:31 | bauzas | Uggla: sure | |
| 10:26:32 | songwenping | bauzas, sean-k-mooney: hi guys, i have attach nvidia gpu(v100/A100) to the vm with 'virsh attach-device --persistent --live' , i can find the gpu in the vm with lspci, but nvidia-smi cnanot find the gpu, with the error 'NVRM: GPU 0000:05:00.0: RmInitAdapter failed! ', do you have any advices? | |
| 11:29:52 | opendevreview | Rajesh Tailor proposed openstack/nova master: Fix case-sensitivity for metadata keys https://review.opendev.org/c/openstack/nova/+/873901 | |
| 12:07:51 | bauzas | songwenping: this is unfortunately not a nova issue, IMHO | |
| 12:08:18 | bauzas | if you can list the vgpu by lcpci, it looks to me a nvidia driver issue | |
| 12:14:12 | songwenping | bauzas: right, but where can we find some doc to prove it's nvidia driver issue? | |
| 12:24:08 | opendevreview | Stephen Finucane proposed openstack/placement master: db: Replace use of deprecated API https://review.opendev.org/c/openstack/placement/+/880623 | |
| 12:24:09 | opendevreview | Stephen Finucane proposed openstack/placement master: tests: Use base class for all functional tests https://review.opendev.org/c/openstack/placement/+/880624 | |
| 12:24:09 | opendevreview | Stephen Finucane proposed openstack/placement master: tests: Warn on *any* SAWarning warning https://review.opendev.org/c/openstack/placement/+/880625 | |
| 12:31:57 | Uggla | gibi, bauzas I have progressed a little. In fact I was fooled. The mount issue is not linked to user vs admin. But the first to the api is failing, but the second one is working, and this is not linked to timing. | |
| 12:32:22 | Uggla | s/first/first call/ | |
| 12:46:58 | bauzas | Uggla: let's discuss this directly by gmeet if you want ;) | |
| 12:47:08 | Uggla | ok | |
| 12:47:17 | Uggla | cleaning my env and calling you | |
| 12:48:13 | bauzas | songwenping: honestly, nova just adds a mdev on the libvirt XML https://libvirt.org/drvnodedev.html#mediated-devices-mdevs | |
| 12:48:32 | bauzas | Uggla: ok | |
| 13:22:03 | artom | Uggla, oh, the privsep mount thing | |
| 13:22:09 | artom | Yeah, it's super weird, I am indeed a witness | |
| 13:41:30 | opendevreview | Dan Smith proposed openstack/nova master: Remove silent failure to find a node on rebuild https://review.opendev.org/c/openstack/nova/+/880632 | |
| 13:41:30 | opendevreview | Dan Smith proposed openstack/nova master: Stop ignoring missing compute nodes in claims https://review.opendev.org/c/openstack/nova/+/880633 | |
| 13:41:41 | dansmith | bauzas: I pulled the RT stuff out into a separate set^ | |
| 13:42:02 | bauzas | dansmith: on a long call with Uggla but okay, I'll try | |
| 13:42:07 | dansmith | and fixed another silent failure we ignore in evacuate | |
| 14:21:10 | bauzas | gibi: dansmith: sean-k-mooney: fyi, we discussed with Uggla about his series and we discovered a potential large leak for ACLs in Manila access rights | |
| 14:21:32 | gibi | ack | |
| 14:21:39 | bauzas | gibi: dansmith: sean-k-mooney: when adding an access-allow for the share, we pass the compute IP address to Manila | |
| 14:21:55 | dansmith | which is visible by the user? | |
| 14:21:57 | bauzas | then the IP address can be seen by any user in the same project, and also any user can delete this ACL | |
| 14:22:19 | dansmith | that'd be bad | |
| 14:22:42 | bauzas | so we need to tell the Manila folks to somehow hide those details | |
| 14:22:57 | dansmith | yeah | |
| 14:23:22 | bauzas | if the ACL is done by a service, it shouldn't be seen by an enduser, neither be able to delete it | |
| 14:23:46 | bauzas | if we ask Manila to lock a share, that's a different concern | |
| 14:23:55 | sean-k-mooney | bauzas: this is partly why i wanted to use the cert auth method not the ip one | |
| 14:23:57 | bauzas | because users can delete ACLs without unlocking | |
| 14:24:09 | dansmith | ...yup | |
| 14:24:38 | bauzas | sean-k-mooney: it should also leak the certificate, right? | |
| 14:24:49 | bauzas | and users could also delete the ACL | |
| 14:25:23 | bauzas | there are two problems to resolve : 1/ we need to hide the ACL, 2/ we need to make sure it's not possible to delete it by an user | |
| 14:25:30 | Uggla | sean-k-mooney, also today you can do a manila show share that will reveal IP + export location | |
| 14:25:32 | sean-k-mooney | the cert would be a per vm cert generted for that insntace | |
| 14:25:33 | bauzas | we == Manila API | |
| 14:25:40 | sean-k-mooney | and it woudl only be the public key | |
| 14:25:50 | sean-k-mooney | so i dont think we care if that is visable | |
| 14:26:31 | dansmith | does nova use only the user's token to talk to mania? | |
| 14:26:32 | bauzas | sean-k-mooney: then we would need to extend the lock mechanism to deny any ACL modification | |
| 14:26:35 | dansmith | *manila | |
| 14:26:53 | sean-k-mooney | https://docs.openstack.org/api-ref/shared-file-system/?expanded=grant-access-detail#grant-access | |
| 14:26:58 | bauzas | dansmith: we tested this also with admin | |
| 14:27:02 | dansmith | I'm asking | |
| 14:27:08 | sean-k-mooney | i would prefer to use cert or user for auth then ip | |
| 14:27:20 | bauzas | dansmith: if you're an admin, you can generate a share and add an ACL | |
| 14:27:36 | bauzas | dansmith: but any user from the same project can both see the share and delete the ACL you created | |
| 14:27:40 | dansmith | bauzas: can you answer my question? | |
| 14:28:01 | dansmith | does nova use the user's token (only) to talk to manila? | |
| 14:28:13 | sean-k-mooney | dansmith: yes i belive we use only the user token in the current proposal | |
| 14:28:18 | bauzas | dansmith: IIRC today yes but Uggla knows better about it | |
| 14:28:24 | dansmith | ack | |
| 14:28:30 | sean-k-mooney | but we could add an admin token if we needed too. i dont think we use any admin api today | |
| 14:28:38 | dansmith | we have a lot of different places where we've done that in the past and it has come back to bite us | |
| 14:28:39 | sean-k-mooney | but we might for the lock api | |
| 14:28:42 | bauzas | we discussed this at the PTG, we could use a service token if Manila adds it | |
| 14:28:51 | bauzas | but again, we missed at the PTG the ACL problem | |
| 14:28:59 | dansmith | cinder, glance, etc.. so perhaps we should not replicate the same thing here | |
| 14:29:21 | sean-k-mooney | bauzas: the probelm is leakign the ip/subnet of the nova comptue host | |