Earlier  
Posted Nick Remark
#openstack-nova - 2022-02-08
12:22:52 dmitriis sean-k-mooney: looks like WhitelistPciAddress.match treats the presence of pci_phys_addr as an indicator that the caller is trying to match a VF https://review.opendev.org/c/openstack/nova/+/824834/10/nova/pci/devspec.py#252 And then PciDeviceSpec.match https://review.opendev.org/c/openstack/nova/+/824834/10/nova/pci/devspec.py#377 does a
12:22:52 dmitriis comparison of its vendor_id/product_id against both ANY and the device dict passed in.
12:30:30 dmitriis IIUC, something like this `passthrough_whitelist = {"vendor_id":"*","product_id":"*", "address":"0000:0a:00.0"}` would match both a PF with address 0000:0a:00.0 and any vendor/device ID, and VFs of any PF with a PCI address 0000:0a:00.0
12:33:02 dmitriis so, in summary: I could block this explicitly if remote_managed: "true" is present because it allows both PFs and VFs to be matched but we want to prohibit PFs.
12:35:47 sean-k-mooney dmitriis: sorry was on a meeting reading back
12:35:52 dmitriis np
12:36:51 sean-k-mooney you mean block "vendor_id":"*","product_id":"*", explcitly if remote managed
12:37:24 sean-k-mooney i think you need to do this outside of the whitelist/devspec parseing
12:38:17 sean-k-mooney when we are constucting the device pools in the pci tracker we could block it as we woudl know if its a pf
12:46:41 sean-k-mooney dmitriis: im just reviewing you latest patches by the way lets continue this conversation after that
12:47:27 dmitriis sean-k-mooney: ack, trying to figure out what to move to PciDevTracker. There's some remove-managed-related logic already in the PciDeviceSpec in the current patches
12:49:45 sean-k-mooney one important thign to rememebr is the nova/pci fould has to be driver independent so we need to ensure that we do not do any checks there that would be driver dependent. those need to live in the virt driver. in this case i think you can do checks pci manager possible in update_devices_from_hypervisor_resources
12:50:34 sean-k-mooney similar to https://github.com/openstack/nova/blob/master/nova/pci/manager.py#L134-L136
12:51:11 sean-k-mooney that is currently delegating back to the whitelist class https://github.com/openstack/nova/blob/master/nova/pci/whitelist.py#L85-L93
12:51:49 sean-k-mooney but its really just validting the whitelist data is correct and that the device being tested is allowed
12:52:50 sean-k-mooney dmitriis: i could maybe live with a check here https://github.com/openstack/nova/blob/master/nova/pci/whitelist.py#L92
12:53:17 sean-k-mooney assert that dev is not a pf if spec has remote managed
12:55:28 sean-k-mooney dmitriis: perhaps it would make sense to change device_assignable to return the spec that allowed it so that you can use it to do other validation
12:55:37 dmitriis sean-k-mooney: ok, I think I see where you are going. While the PF and VF-related checks for the remote-managed tag are not themselves driver-dependent, whether we need to do them or not is driver-depenent.
12:56:14 dmitriis s/depenent/dependent/
12:57:02 sean-k-mooney well yes and no the data required to do the check is driver depenent since we need to get the list of device form the driver
12:57:14 sean-k-mooney i guess technially you could get that from sysfs.
12:57:32 sean-k-mooney im concerned that might break powervm
12:57:53 sean-k-mooney or hyperv but im not sure it would
12:58:36 dmitriis yeah, I added a couple of utils to do sysfs lookup in the last upload https://review.opendev.org/c/openstack/nova/+/824834/10/nova/pci/devspec.py#324 https://review.opendev.org/c/openstack/nova/+/824834/10/nova/pci/utils.py#217
12:59:05 sean-k-mooney so https://github.com/openstack/nova/blob/master/nova/pci/utils.py#L132-L144 could be called by the whitelist
12:59:24 sean-k-mooney in device_assignable
12:59:35 sean-k-mooney to check if the dev is a pf
12:59:52 sean-k-mooney and you could see if it has remote managed and return true or false
13:00:02 sean-k-mooney that would filter out the pfs
13:01:03 sean-k-mooney may main concern with tha tis currently device_assinable does not make casles to the underlying system
13:01:25 sean-k-mooney it used the in memory specs list and the data passed in
13:02:57 dmitriis sean-k-mooney: there's one code-path where it may do it https://github.com/openstack/nova/blob/master/nova/pci/devspec.py#L290
13:03:16 dmitriis get_function_by_ifname uses sysfs for the netdev lookup
13:03:43 sean-k-mooney yes indriectly by the match function
13:03:47 sean-k-mooney with dev name
13:04:12 sean-k-mooney so ya if it works for you i woudl move the check to device_assignable
13:04:32 sean-k-mooney although
13:04:41 sean-k-mooney you curently have it in _init_dev_details
13:05:45 sean-k-mooney dmitriis: im just looking at https://review.opendev.org/c/openstack/nova/+/824834/9..10/nova/pci/devspec.py#287
13:06:57 dmitriis ack
13:08:09 opendevreview Stephen Finucane proposed openstack/placement master: tox: Enable SQLAlchemy 2.0 warnings https://review.opendev.org/c/openstack/placement/+/801108
13:08:09 opendevreview Stephen Finucane proposed openstack/placement master: db: Use Row, not LegacyRow https://review.opendev.org/c/openstack/placement/+/828305
13:17:13 dmitriis sean-k-mooney: the goal of moving the checks closer to parsing was to avoid devices from getting into the pools in the first place (to avoid filtering PFs in PciDeviceStats) and for early config error detection. During the earlier reviews I went from (1) filtering in PciDeviceStats -> (2) returning false while matching in PciDeviceSpec -> (3)
13:17:13 dmitriis raising exceptions in _init_dev_details
13:17:30 sean-k-mooney dmitriis: while i dont nessisaly love doing these remote managed check in _init_dev_details i think im ok with it for now
13:18:22 sean-k-mooney dmitriis: your current code avoids needeign to pass allow remote managed allover the place which was the main code smell
13:18:36 sean-k-mooney we have other checks like
13:18:38 sean-k-mooney if address and self.dev_name:
13:18:41 sean-k-mooney raise exception.PciDeviceInvalidDeviceName()
13:18:46 sean-k-mooney in _init_dev_details
13:18:52 sean-k-mooney so it kind of fits there
13:19:55 dmitriis sean-k-mooney: well, the other thing still needs addressing I think: https://review.opendev.org/c/openstack/nova/+/827839/2/nova/pci/devspec.py#318. Do we want to check for the use of `remote_managed: "true"` in case a driver doesn't support it?
13:20:22 dmitriis perhaps not in the devspec.py code but I could try to expose a property on the whitelist and check for this elsewhere
13:20:24 sean-k-mooney no i dont think so
13:20:34 sean-k-mooney we could
13:20:38 dmitriis sean-k-mooney: ok, so I can then just drop this altogether
13:20:42 sean-k-mooney but the request filter woudl block it
13:20:50 sean-k-mooney with the capablty traits
13:21:35 sean-k-mooney https://review.opendev.org/c/openstack/nova/+/812111/22/nova/scheduler/request_filter.py
13:21:40 sean-k-mooney that is unconditionally enabled
13:21:57 sean-k-mooney so as long as the driver does not incorrectly report the trait
13:22:03 dmitriis so, we are essentially allowing `remote_managed: "true"` to be specified in the config even if a particular driver doesn't support it. It will not be surfaced to an operator early but other parts of the system will work as they should
13:22:06 sean-k-mooney it the host will not be included in the set form placemnt
13:22:48 sean-k-mooney i guess we could see how gibi feels but i think documenation fo "this only work with libvirt" + the prefileter is enough
13:24:28 dmitriis Works for me. I'll just drop the extra check in the other patch for now and we can conclude there if we are ok with the result
13:24:53 dmitriis since we document it for deployers/operators it should be ok in my view
13:25:16 sean-k-mooney cool
13:25:54 sean-k-mooney so assuming the neutorn lib stuff gets sorted we hopefully can get this all merged in the next week or two
13:26:03 sean-k-mooney dmitriis: what is the state of the neutron code
13:27:25 dmitriis sean-k-mooney: just got updated to include the new VNIC_TYPE_REMOTE_MANAGED https://review.opendev.org/c/openstack/neutron/+/808961/. Otherwise fnordahl promoted it from the WIP status and is waiting for reviews
13:28:10 dmitriis there's also a second patch around extra validation https://review.opendev.org/c/openstack/neutron/+/818420
13:28:56 dmitriis the test failure on the first is about VNIC_TYPE_REMOTE_MANAGED so we're just waiting for it to land
13:30:14 sean-k-mooney ok we likely need a depends on link agaisnt the neutorn series at some point in the nova one
13:31:17 sean-k-mooney https://review.opendev.org/c/openstack/nova/+/812111/22 should also be before https://review.opendev.org/c/openstack/nova/+/824835/13 i think
13:32:31 sean-k-mooney dmitriis: when you respin those for the vnic type change can you put the filter before the final patch that enables the feature
13:33:03 sean-k-mooney if that is a lot of work we can proably keep the order as is and just merge tehm together
13:33:29 sean-k-mooney /together/at the same time/
14:07:35 dmitriis sean-k-mooney: ack, the filter one needs `is_smartnic_port` from the previous patch which I am going to replace to is_remote_managed_port
14:08:18 dmitriis so I guess I could add the is_remote_managed_port in the filter patch instead and reorder
14:08:54 sean-k-mooney if that is not a lot of touble i think it woudl better yes but ill leave that to you
14:09:40 sean-k-mooney it just chagnes if we merge the patch all at one or together.
14:14:20 dmitriis sean-k-mooney: ack, started looking into it now (had to attend a call)
14:14:33 opendevreview Jonathan Race proposed openstack/nova master: Adds Pick guest CPU architecture based on host arch in libvirt driver support https://review.opendev.org/c/openstack/nova/+/822053
14:43:11 sean-k-mooney stephenfin: if you rebase https://review.opendev.org/c/openstack/nova/+/705792/12 it might help
14:43:22 sean-k-mooney stephenfin: we are now skiping the failing test https://review.opendev.org/c/openstack/nova/+/705792/12
14:43:31 stephenfin zuul should do that for me
14:43:38 stephenfin so long as there are no conflicts
14:43:48 sean-k-mooney it should but i think the last run did not have it
14:43:51 sean-k-mooney perhaps
14:43:53 sean-k-mooney it was before that merged
14:44:25 sean-k-mooney oh i ment https://review.opendev.org/c/openstack/nova/+/827851
14:44:35 sean-k-mooney ah it was
14:44:41 sean-k-mooney the last run was on the 3rd
14:44:50 sean-k-mooney the skip merged on the 5th
14:45:57 sean-k-mooney stephenfin: after those two patches merge care to respin your mock replacement patch?
14:46:56 sean-k-mooney oh looks like there is a func tst failure...
14:48:40 opendevreview Balazs Gibizer proposed openstack/placement master: Extend the RP db query to support any-traits https://review.opendev.org/c/openstack/placement/+/825848
14:48:41 opendevreview Balazs Gibizer proposed openstack/placement master: DB layer should only depend on trait id not names https://review.opendev.org/c/openstack/placement/+/826490
14:48:51 opendevreview Balazs Gibizer proposed openstack/placement master: Enhance doc of _get_trees_with_traits https://review.opendev.org/c/openstack/placement/+/825780

Earlier   Later