On 11/09/15 10:49 -0400, Nikhil Komawar wrote:
On 9/11/15 10:40 AM, Flavio Percoco wrote:On 11/09/15 07:58 -0400, Nikhil Komawar wrote:You are right in the sense that's the ideal scenario. (Impl-wise) However, even today we do not guarantee that behavior. If someone were to propose a new driver or a change driver capability or any thing of such order, images in status killed won't be guaranteed to have removed the garbage data. The driver may not choose to be resilient enough or would not take the responsibility of data removal synchronously on failures.I think it's glance's responsibility to make sure the driver deletes the image data. If the API is not strong enough to guarantee this, then we should change that.Sounds like a good direction but I still think we cannot commit to it. It really depends on the driver and introduction of new or upgrades to the existing drivers (if not their capabilities) is a not a good idea. This will result in backward incompatibility in many cases and there can be enough corner cases when we think about v1, v2 and tasks to address them at Glance API level.
mmh, I guess my reply was confusing again. I meant the glance_store API and not Glance's API. What I'm trying to say is that just depending on the scrubber (which not everyone deploys) won't be enough to fix this issue entirely.
I personally prefer the image-data delete to be asynchronous and fault tolerant (on DELETE of image record). However, even that cannot match the performance of a scrubber like service.
Yes, I also prefer the data to be deleted asynchrounously when it comes to not blocking glance-api threads. However, that doesn't make this step completely fault-tolerant. What I'm saying is that depending on a asynchronous deletion only when there's also the chance to possibility of having a synchrounous one is not good.
So, in certain cases we need to advice running the scrubber if the data has the tendency to be left behind.
This is where we difer in thoughts. I understand where you're coming from with this but I think we shouldn't have those cases. The fact that there's data being left behind is the real issue. We shouldn't have those cases. Having async deletions is fine and I agree they are useful. Cheers, Flavio
Taking that fact in account, I have thought of Brianna's patch to be okay.Oh sure, I'm not trying to say it was a wrong choice. Sorry if it sounded like that. I was replying to the thought of extending scrubber (unless there's a patch that does this that I might have missed). Cheers, FlavioOn 9/11/15 4:42 AM, Flavio Percoco wrote: On 10/09/15 15:36 -0400, Nikhil Komawar wrote: The solution to this problem is to improve the scrubber to clean up the garbage data left behind in the backend store during such failed uploads. Currently, scrubber cleans up images in pending_delete and extending that to images in killed status would avoid such a situation. While the above would certainly help, I think it's not the right solution. Images in status "killed" should not have data to begin with. I'd rather find a way to clean that data as soon as the image is moved to a "killed" state instead of extending the scrubber. Cheers, Flavio On 9/10/15 3:28 PM, Poulos, Brianna L. wrote: Malini, Thank you for bringing up the ³killed² state as it relates to quota. We opted to move the image to a killed state since that is what occurs when an upload fails, and the signature verification failure would occur during an upload. But we should keep in mind the potential to take up space and yet not take up quota when signature verification fails. Regarding the MD5 hash, there is currently a glance spec [1] to allow the hash method used for the checksum to be configurable‹currently it is hardcoded in glance. After making it configurable, the default would transition from MD5 to something more secure (like SHA-256). [1] https://review.openstack.org/#/c/191542/ Thanks, ~Brianna On 9/10/15, 5:10 , "Bhandaru, Malini K" <malini.k.bhand...@intel.com> wrote: Brianna, I can imagine a denial of service attack by uploading images whose signature is invalid if we allow them to reside in Glance In a "killed" state. This would be less of an issue "killed" images still consume storage quota until actually deleted. Also given MD-5 less secure, why not have the default hash be SHA-1 or 2? Regards Malini -----Original Message----- From: Poulos, Brianna L. [mailto:brianna.pou...@jhuapl.edu] Sent: Wednesday, September 09, 2015 9:54 AM To: OpenStack Development Mailing List (not for usage questions) Cc: stuart.mcla...@hp.com Subject: Re: [openstack-dev] [glance] [nova] Verification of glance images before boot Stuart is right about what will currently happen in Nova when an image is downloaded, which protects against unintentional modifications to the image data. What is currently being worked on is adding the ability to verify a signature of the checksum. The flow of this is as follows: 1. The user creates a signature of the "checksum hash" (currently MD5) of the image data offline. 2. The user uploads a public key certificate, which can be used to verify the signature to a key manager (currently Barbican). 3. The user creates an image in glance, with signature metadata properties. 4. The user uploads the image data to glance. 5. If the signature metadata properties exist, glance verifies the signature of the "checksum hash", including retrieving the certificate >from the key manager. 6. If the signature verification fails, glance moves the image to a killed state, and returns an error message to the user. 7. If the signature verification succeeds, a log message indicates that it succeeded, and the image upload finishes successfully. 8. Nova requests the image from glance, along with the image properties, in order to boot it. 9. Nova uses the signature metadata properties to verify the signature (if a configuration option is set). 10. If the signature verification fails, nova does not boot the image, but errors out. 11. If the signature verification succeeds, nova boots the image, and a log message notes that the verification succeeded. Regarding what is currently in Liberty, the blueprint mentioned [1] has merged, and code [2] has also been merged in glance, which handles steps 1-7 of the flow above. For steps 7-11, there is currently a nova blueprint [3], along with code [4], which are proposed for Mitaka. Note that we are in the process of adding official documentation, with examples of creating the signature as well as the properties that need to be added for the image before upload. In the meantime, there's an etherpad that describes how to test the signature verification functionality in Glance [5]. Also note that this is the initial approach, and there are some limitations. For example, ideally the signature would be based on a cryptographically secure (i.e. not MD5) hash of the image. There is a spec in glance to allow this hash to be configurable [6]. [1] https://blueprints.launchpad.net/glance/+spec/ image-signing-and-verificati o n-support [2] https://github.com/openstack/glance/commit/ 484ef1b40b738c87adb203bba6107dd b 4b04ff6e [3] https://review.openstack.org/#/c/188874/ [4] https://review.openstack.org/#/c/189843/ [5] https://etherpad.openstack.org/p/ liberty-glance-image-signing-instructions [6] https://review.openstack.org/#/c/191542/ Thanks, ~Brianna On 9/9/15, 12:16 , "Nikhil Komawar" <nik.koma...@gmail.com> wrote: That's correct. The size and the checksum are to be verified outside of Glance, in this case Nova. However, you may want to note that it's not necessary that all Nova virt drivers would use py-glanceclient so you would want to check the download specific code in the virt driver your Nova deployment is using. Having said that, essentially the flow seems appropriate. Error must be raise on mismatch. The signing BP was to help prevent the compromised Glance from changing the checksum and image blob at the same time. Using a digital signature, you can prevent download of compromised data. However, the feature has just been implemented in Glance; Glance users may take time to adopt. On 9/9/15 11:15 AM, stuart.mcla...@hp.com wrote: The glance client (running 'inside' the Nova server) will re-calculate the checksum as it downloads the image and then compare it against the expected value. If they don't match an error will be raised. How can I know that the image that a new instance is spawned from - is actually the image that was originally registered in glance - and has not been maliciously tampered with in some way? Is there some kind of verification that is performed against the md5sum of the registered image in glance before a new instance is spawned? Is that done by Nova? Glance? Both? Neither? The reason I ask is some 'paranoid' security (that is their job I suppose) people have raised these questions. I know there is a glance BP already merged for L [1] - but I would like to understand the actual flow in a bit more detail. Thanks. [1] https://blueprints.launchpad.net/glance/+spec/ image-signing-and-verif ica tion-support -- Best Regards, Maish Saidel-Keesing ------------------------------ _______________________________________________ OpenStack-dev mailing list OpenStack-dev@lists.openstack.org http://lists.openstack.org/cgi-bin/mailman/listinfo /openstack-dev End of OpenStack-dev Digest, Vol 41, Issue 22 ********************************************* ______________________________________________________________________ ___ _ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org? subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/ openstack-dev -- Thanks, Nikhil _______________________________________________________________________ ___ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org? subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/ openstack-dev __________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org? subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/ openstack-dev __________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org? subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/ openstack-dev __________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org? subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev -- Thanks, Nikhil __________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org? subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev __________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev -- Thanks, Nikhil-- Thanks, Nikhil
-- @flaper87 Flavio Percoco
pgpdUt2eK82j4.pgp
Description: PGP signature
__________________________________________________________________________ OpenStack Development Mailing List (not for usage questions) Unsubscribe: openstack-dev-requ...@lists.openstack.org?subject:unsubscribe http://lists.openstack.org/cgi-bin/mailman/listinfo/openstack-dev