On Fri, Jan 15, 2021 at 12:51:20PM -0500, Theodore Ts'o wrote: > On Fri, Jan 15, 2021 at 04:24:23PM +0000, Christoph Hellwig wrote: > > > > That is what the capabilities are designed for and we already check > > for them. > > So perhaps I'm confused, but my understanding is that in the > containers world, capabilities are a lot more complicated. There is: > > 1) The initial namespace capability set > > 2) The container's user-namespace capability set > > 3) The namespace in which the file system is mounted --- which is > "usually, but not necessarily the initial namespace" and > presumably could potentially not necessarily be the current > container's user name space, is namespaces can be hierarchically > arranged. > > Is that correct? If so, how does this patch set change things (if > any), and and how does this interact with quota administration > operations? The cases you listed are correct. The patchset doesn't change them. Simply put, the patchset doesn't alter capability checking in any way. > > On a related note, ext4 specifies a "reserved user" or "reserved > group" which can access the reserved blocks. If we have a file system > which is mounted in a namespace running a container which is running > RHEL or SLES, and in that container, we have a file system mounted (so > it was not mounted in the initial namespace), with id-mapping --- and > then there is a further sub-container created with its own user > sub-namespace further mapping uids/gids --- will the right thing > happen? For that matter, how *is* the "right thing" defined? In short, nothing changes. Whatever happened before happens now. Specifically s_resuid/s_resgid are superblock mount options and so never change on a per-mount basis and thus also aren't affected by idmapped mounts. > > Sorry if this is a potentially stupid question, but I find user > namespaces and id and capability mapping to be hopefully confusing for > my tiny brain. :-) No, I really appreciate the questions. :) My brain can most likely handle less. :) Christian
