Earlier  
Later
New Search
Posted Nick Remark
#openstack-nova - 2020-06-19
15:57:14 dansmith it is, the question mark was rhetorical :)
15:57:24 stephenfin oh
15:57:33 artom dansmith, I'm trying to understand what you mean by reading `man 2 select`
15:57:55 stephenfin we'll come fish you out of that rabbit hole later
15:57:58 artom There's FD_SET(), but that's a macro
15:58:26 dansmith artom: the readfds, for example is a pointer to a word, with bits set nonzero if the FD represented by that bit should be checked for readability
15:58:31 artom (Also, `man 2 select` is the perfect title for a geek version of The Bachelorette)
16:02:18 artom dansmith, there's nothing about masks in the readfds para...
16:02:27 artom It's a set of file descriptors
16:02:49 dansmith artom: no, it's not, you're just not meant to know how it works under the covers
16:03:27 dansmith FD_SET is just readfds|=(1 << fd)
17:05:57 openstackgerrit Stephen Finucane proposed openstack/nova master: Guard against missing image cache directory https://review.opendev.org/736964
17:07:44 stephenfin dansmith: ^
17:09:25 dansmith stephenfin: I think you got that logic wrong
17:09:32 stephenfin crap
17:09:48 dansmith if the directory does not exist, then True, else if it does, True if st_dev matches
17:11:12 stephenfin At first glance it reads correctly to me. Bear with me
17:11:35 dansmith it's okay it's confusing, so make sure I'm right here
17:12:15 dansmith if the directory does exist, then we need to check to see if the st_dev matches.. if the directory does not exist, then we return True because when we create it, it will be on the same device
17:12:24 stephenfin right, yeah, makes sense
17:12:34 stephenfin yay, Fridays
17:12:39 stephenfin I'll respin again
17:18:53 openstackgerrit Stephen Finucane proposed openstack/nova master: Guard against missing image cache directory https://review.opendev.org/736964
17:23:14 dansmith stephenfin: got it
17:24:07 stephenfin I've done enough on that so I'll defer to melwitt or someone to approve
17:24:28 melwitt yeah I'll review it
17:25:57 dansmith ack, I think that's a good idea
17:47:09 melwitt dansmith: added some comments, am I missing something? because to me it seems this is doing the opposite of what is wanted in the commit message https://review.opendev.org/736964
17:48:29 melwitt I see that you've said the first time an image is cached, the dir will be created, but update_provider_tree is a periodic task right? what will ensure the cache dir is created by the time we ask for get_disk_usage()?
17:49:18 dansmith nothing will ensure that it's created
17:49:32 dansmith the bug is that if it's not created and this runs, it doesn't handle the case where it's not created
17:49:56 dansmith however, we don't just want to do nothing if it's not created, we want to take the path that assumes that when it does get created, it will necessarily be on the same st_dev, and thus not report any cache
17:50:35 melwitt yeah, I got that part but from what I was looking, if the cache dir doesn't exist it will fall through and start checking usage of the cache dir and if it doesn't exist won't that be bad?
17:51:09 dansmith where will it fall through?
17:51:14 melwitt cause we'll end up here if the dir doesn't exist https://github.com/openstack/nova/blob/f5f7c2540150c7ee7640c834d5caec31b3f5a7ab/nova/virt/libvirt/imagecache.py#L366
17:52:11 dansmith er wait, maybe I'm inverting that in my head
17:52:47 melwitt it will return 0 disk usage if self.cache_dir_is_on_same_dev_as_instances_dir is False, right?
17:52:49 dansmith okay right right, I forgot how this ended up
17:52:58 dansmith we initially talked about reporting inventory for the cache disk itself,
17:53:03 melwitt else it will do the math on the cache dir
17:53:28 dansmith and this is the opposite where we reserve only if the cache is the _same_ disk, which I think is why we endded up checking the reverse condition instead of what we're looking for
17:54:05 dansmith so yeah, we need to bail if the directory doesn't exist because we're checking the cache_dir which might not exist yet
17:54:25 dansmith I still think the property shouldnt' explode if the directory doesn't exist,
17:54:44 dansmith but you're right, it'll still hit the same problem but a little later on
17:55:19 dansmith the logic is right, it just doesn't prevent us from hitting the same problem in the sum()
17:55:45 melwitt oh, ok, so the logic on exists() should not be inverted to remove the 'not'
17:56:44 dansmith no, because if it does not exist, the not will make that true, which will short-circuit and return true without evaluating the st_dev part
17:57:03 melwitt oh, right
17:58:29 dansmith I'll pull this down and make that change, just a sec
17:59:20 melwitt ok, cool
17:59:26 dansmith eff, just realized I didn't commit something I did earlier today unrelated
17:59:40 openstackgerrit Dan Smith proposed openstack/nova master: DNM: Try to make a glance multistore job https://review.opendev.org/734184
17:59:53 dansmith so another few days in the queue for me
18:00:33 melwitt heh :(
18:10:56 openstackgerrit Dan Smith proposed openstack/nova master: Guard against missing image cache directory https://review.opendev.org/736964
18:10:58 dansmith melwitt: ^
18:12:45 dansmith added comments to hopefully make it clearer why each thing is happening
18:14:14 melwitt looking ... ah yep, that is helpful
18:19:48 dansmith thanks. I was clearly focused on getting that property to return the right thing all the time and not realizing we'd be introducing a new vector to the error :)
20:43:10 mnaser just wanna put eyes on https://review.opendev.org/#/c/736370/ if/when possible :)
21:04:42 sean-k-mooney mnaser: form the title that makes sense
21:05:05 sean-k-mooney ill take a look on monday
21:05:21 sean-k-mooney actully its pretty short
21:07:53 mnaser sean-k-mooney: thank you!
21:18:08 sean-k-mooney melwitt: just finished reviewing https://review.opendev.org/#/c/736370/2 want to take a look?
21:18:39 sean-k-mooney mnaser: the only thing im not sure is do we want to have a release note or not.
21:19:43 sean-k-mooney mnaser: this would have failed anyway later so the only behavior change is that by failing in _validate_flavor_image_nostatus we now fail before creating a server record
21:20:08 sean-k-mooney so instead of ending up with a server in error state we just get an excption form spawn
21:21:37 sean-k-mooney which is better but as an operator would you want that called out in the release notes give the excption you are raising inherits form FlavorDiskTooSmall which can already casue the same behvoiur if the volume is less the min_gb
21:21:42 sean-k-mooney mnaser: ^
21:31:22 arjunbaindur hey, have a question about configuring PCI passthrough. We have a device that has 2 functions. On the hypervisor, the PCI address has it on same bus and same slot. When it shows up inside the VM, the PCI address has it under different slots
21:31:31 arjunbaindur for example, suppose device has two functions at 83:00.0 and 83:00.1
21:31:39 arjunbaindur Under the VM, the devices show up under different slots, for example 0:5.0 and 0:6.0
21:32:34 sean-k-mooney im guessing its a gpu with a second function for the audio encoder on the gpu?
21:32:38 arjunbaindur any idea how to get them to show up as two functions under the same slot
21:32:39 sean-k-mooney or a similar device
21:32:56 sean-k-mooney unfortunetly we do not have a way to do that currently
21:33:12 sean-k-mooney we did dicuss it breifly about 3-6 months ago
21:33:46 sean-k-mooney efffectivly we would need to modify the pci alisia to support multifuntion devices
21:34:46 sean-k-mooney such as allowing one pci aliase to refrence another and adding a multifuntion=true tag
21:35:17 sean-k-mooney so the parent alias would refence the child aliase and but or would have the multifunction=ture tag
21:36:13 sean-k-mooney arjunbaindur: i think on windows this can break the audio encoder on nvidga gpus since the driver is expecting it to be a second fuction on the same bus/slot
21:36:18 sean-k-mooney but on linux it works
21:36:34 sean-k-mooney im sure it would be an issue for other device that had the same toplogy
21:39:01 arjunbaindur to be specific, this is a Xilinx card
21:39:12 arjunbaindur https://www.xilinx.com/support/documentation/boards_and_kits/accelerator-cards/1_5/ug1301-getting-started-guide-alveo-accelerator-cards.pdf
21:39:57 arjunbaindur since the two devices show up under diff PCI slots, we arent able to install drivers
21:40:35 sean-k-mooney ah ok well its the same issue unfortunetly
21:40:48 sean-k-mooney are you trying to reproguam the aclerat using 1 fucntion
21:40:56 sean-k-mooney and consume it using the second
21:42:54 sean-k-mooney 32-64GB of ddr4 + and fpga and some qsfp ports. that looks like a pretty interesting card
21:43:14 sean-k-mooney arjunbaindur: are you trying to use this via cyborg or just novas pci passhtough
21:44:08 arjunbaindur just Nova PCI passthrough: followed this doc basically: https://docs.openstack.org/nova/rocky/admin/pci-passthrough.html
21:44:41 arjunbaindur Added the two PCI addresses for each function in the whitelist
21:44:42 sean-k-mooney ya so as i said we are missing the ablity to group multiple fucntion together form teh same device that are schulded togehter
21:45:07 arjunbaindur is there an upstream bug or anything to track this?
21:45:28 sean-k-mooney i was trying to find one. its not really a bug since it was never part of the orginial design
21:46:03 sean-k-mooney i suspcet this would need a short spec or maybe a specless blueprint
21:48:16 sean-k-mooney arjunbaindur: looking at the open blueprints no
21:48:37 sean-k-mooney so the feature you need is support for "multifunction pci device passthough"

Earlier   Later