Quoting Serge E. Hallyn (serge@xxxxxxxxxx): > Quoting Eric W. Biederman (ebiederm@xxxxxxxxxxxx): > > "Serge E. Hallyn" <serge@xxxxxxxxxx> writes: > > > > > Quoting Andrew G. Morgan (morgan@xxxxxxxxxx): > > >> > > >> I guess I'm confused how we have strayed so far that this isn't an obvious > > >> requirement. Uid=0 as being the root of privilege was the basic problem > > >> that capabilities were designed to change. > > > > > > The task executing the file can be any uid mapped into the namespace. The > > > file only has to be owned by the root of the user_ns. Which I agree is > > > unfortunate. We can work around it by putting the root uid into the xattr > > > itself (which still isn't orthogonal but allows the file to at least by > > > owned by non-root), but the problem then is that a task needs to know its > > > global root k_uid just to write the xattr. > > > > The root kuid is just make_kuids(user_ns, 0) so it is easy to find. > > > > It might be a hair better to use the userns->owner instead of the root > > uid. That would allow user namespaces without a mapped root to still > > use file capabilities. > > That's all fine if the kernel does it for us magically. Which is what we're > talking about below. Above I was talking about userspace putting it into > the xattr. > > > >> Uid is an acl concept. Capabilities are supposed to be independent of that. > > >> > > >> If we want to support NS file capabilities I would look at replacing the > > >> xattr syscall with a dedicated file capabilities modification syscall. Then > > > > > > That was one ofthe possibilities I'd mentioned in my earlier proposal, > > > fwiw. The problem is if we want tar to still work unmodified then > > > simple xattr operations still have to work. > > > > > > Maybe there's workable semantics there though. Worth thinking about. > > > > If the problem is compatibilty please look at > > posix_acl_fix_xattr_from_user. With something similar for the > > All right. Excellent. I simply didn't think something like that would > be acceptable. I tend to think of xattrs as just out of band file contents, > but generally under user control. I guess that's not right. > > > security.capability attribute we can perform whatever transformation > > makes sense. I admit adding 4 bytes is a bit of a pain in that context > > but not a big one. > > If we can do all the magic in the kernel behind the scenes, then I > absolutely do not mind adding a new security.capability version with 4 > more bytes. Userspace can just write the old xattr format with the new > version number, kernel fills in the userns owner kuid. It's what I > originally wanted to do, but didn't think was acceptable. > > Sounds great! So I'm still mulling this over and still undecided as to whether we want to 1. leave the xattr as is and use a new pair of syscalls for setting/unsetting filecaps. This would truly let us hide the implementation detail of the file having to be owned by root (apart from returning a perhaps-unexpected EPERM when file isn't owned by uid 0, and documenting that as something that can be changed later) 2. hide the magic in get/setxattr of security.capability. And if we do that, then whether to hide the security.nscapability (or newer-version security.capbility if that's what we do). probably not hide it... -serge -- To unsubscribe from this list: send the line "unsubscribe linux-api" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html