[ceph-users] radosgw, Keystone integration, and the S3 API
florian at citynetwork.eu
Mon Nov 19 07:23:06 PST 2018
I've recently started a documentation patch to better explain Swift
compatibility and OpenStack integration for radosgw; a WIP PR is at
https://github.com/ceph/ceph/pull/25056/. I have, however, run into an
issue that I would really *like* to document, except I don't know
whether what I'm seeing is how things are supposed to work. :)
This is about multi-tenancy in radosgw, in combination with S3
authentication via Keystone (and EC2-compatible credentials generated
from OpenStack, as explained in my doc patch). Now, when I enable
rgw_s3_use_keystone_auth and rgw_keystone_implicit_tenants, then, if I
create an S3 bucket in radosgw for the first time, naming that bucket
"foo", the following things happen:
* I see a user that has been created, and that I can query with
"radosgw-admin user info", that is named
(that is, the Keystone tenant/project UUID twice, separated by a $
character). Its display_name is the name of my tenant.
* With "radosgw-admin bucket list
I see a bucket that has been created, and that has been named "foo".
So far, all is well. If I do this, then I can see an bucket named
"foo" if I use an S3 client, and I can see a container named "foo",
with identical content, if I use the Swift API.
Now, if I enable rgw_swift_account_in_url, and update my Keystone
object store endpoint to include AUTH_%(tenant_id)s, then using the
Swift API I can also use public ACLs and temp URLs.
However, I am stumped trying to to understand how exactly this is meant
to work with the S3 API.
So I have two questions:
(1) What do I have to do to get publicly-readable buckets to work in
the Keystone-authenticated scenario? Moreover, what is the correct
path to use, for a non-S3 client like curl or a browser, to access
an object? It seems that using
for S3 objects with a public ACL set, but if I try to use the same
approach with a signed object, I get a 403 with
SignatureDoesNotMatch. It seems like what I have to use for a
signed object is, instead,
However, if I do *ask* for a signed object that includes the
tenant name, as in "s3cmd signurl
s3://5ed51981f4a8468292bf2c578806ebf:foo/bar +120", then I *can*
use the same URL format as for public ACL objects. Is this the
intended behavior? If so, does that mean that an application
using the S3 API, and access/secret keys from OpenStack-backed
EC2, should configure always itself to use the "<tenant_id>:"
prefix to precede the bucket name?
(2) Do I understand the documentation
correctly in that whenever one uses multitenancy of any kind in
radosgw, S3 bucket hostnames can't ever be used? Thus, is it correct
to say that if a radosgw instance is meant to *only* ever
authenticate its users against Keystone, where there is always a
radosgw tenant that is being created, then it's pointless to set
If anyone could shed a light on the above, I can write up the answer and
amend the doc patch. Thanks!
 This would be an additional question: why is the project UUID in
there *twice*? Surely there's a good cause for that, but it presently
escapes me. http://docs.ceph.com/docs/master/radosgw/multitenancy/ says
"TBD – don’t forget to explain the function of rgw keystone implicit
tenants = true" here, which isn't very helpful. :)
More information about the ceph-users