On Fri, 1 Nov 2013 18:29:43 +-0000 +ACI-Myklebust, Trond+ACI +ADw-Trond.Myklebust+AEA-netapp.com+AD4 wrote: +AD4 On Fri, 2013-11-01 at 13:58 -0400, Jeff Layton wrote: +AD4 +AD4 On Fri, 1 Nov 2013 17:47:46 +-0000 +AD4 +AD4 +ACI-Myklebust, Trond+ACI +ADw-Trond.Myklebust+AEA-netapp.com+AD4 wrote: +AD4 +AD4 +AD4 +AD4 +AD4 On Fri, 2013-11-01 at 13:18 -0400, Jeff Layton wrote: +AD4 +AD4 +AD4 +AD4 On Fri, 1 Nov 2013 17:05:08 +-0000 +AD4 +AD4 +AD4 +AD4 +ACI-Myklebust, Trond+ACI +ADw-Trond.Myklebust+AEA-netapp.com+AD4 wrote: +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 On Nov 1, 2013, at 12:57, Jeff Layton +ADw-jlayton+AEA-redhat.com+AD4 wrote: +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 On Fri, 1 Nov 2013 16:50:00 +-0000 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +ACI-Myklebust, Trond+ACI +ADw-Trond.Myklebust+AEA-netapp.com+AD4 wrote: +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APg On Fri, 2013-11-01 at 12:02 -0400, Jeff Layton wrote: +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APgA+ It looks like +AF8-nfs4+AF8-get+AF8-security+AF8-label() has the same problem, but I've +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APgA+ so far been unable to get it to be called, so I didn't patch it. It +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APgA+ seems like getxattr does some special stuff for SELinux labels that +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APgA+ cause them only to ever be fetched once. +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APgA+ +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APgA+ Is there some trick to it? +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APgA+ +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APg +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APg Doesn't 'ls -Z' cause them to security label to be read again? +AD4 +AD4 +AD4 +AD4 +AD4 +AD4APg +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 As best I can tell, security labels are set on the inode when the inode +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 is instantiated, and then are reset on changes (i.e. setxattr). If +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +ICY-and on getxattr, afaics. +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 I don't see that. The call chain is something like this: +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 vfs+AF8-getxattr +AD4 +AD4 +AD4 +AD4 xattr+AF8-getsecurity +AD4 +AD4 +AD4 +AD4 security+AF8-inode+AF8-getsecurity +AD4 +AD4 +AD4 +AD4 selinux+AF8-inode+AF8-getsecurity +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 ...and that function looks like it just converts the current security +AD4 +AD4 +AD4 +AD4 context on the inode to text and plops that into the buffer. +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 Ah, you're right. You have to turn off SELinux in order to hit +AD4 +AD4 +AD4 nfs4+AF8-xattr+AF8-get+AF8-nfs4+AF8-label. +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 another client changes the label though, it's not clear to me how your +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 client would ever notice it until the inode is dropped from the cache. +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 ISTR Eric Paris explaining to me that they do that for performance +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 reasons but it seems like something that needs to be reconsidered in +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 light of labeled NFS. Not picking up a security label change seems like +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 a bug, IMO... +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 To be effective, the security label should normally be set at file creation time. It should rarely, if ever, change. Why would you need to change it from a different client? +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 At least in Fedora, there are SELinux policy changes all the time. +AD4 +AD4 +AD4 +AD4 Sometimes that involves changing how files are labeled. I don't think +AD4 +AD4 +AD4 +AD4 it's reasonable to assume that they only get set at creation time. +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 That doesn't answer the question. Again, why would you need to do this +AD4 +AD4 +AD4 from another client? If you don't have a real-life use case, then it's +AD4 +AD4 +AD4 just another 'doctor it hurts' problem. Stop doing it... +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 +AD4 Ok, how about this then: +ADs) +AD4 +AD4 +AD4 +AD4 NFS doesn't have O+AF8-TMPFILE, so you really +ACo-can't+ACo set the context at +AD4 +AD4 creation time, at least with normal syscalls. There will always be a +AD4 +AD4 race window between creating a file and setting the SELinux context on +AD4 +AD4 it. +AD4 +AD4 Ummm. This is supposed to be a security feature. How can you tolerate race windows? +AD4 +AD4 IOW: If I get the security label +ACI-wrong+ACI on lookup due to such a race, +AD4 then how am I supposed to know not to act on it? +AD4 +AD4 +AD4 FWIW, I just started playing with this stuff and this behavior just +AD4 +AD4 gave me pause. This is the sort of thing that will give people fits +AD4 +AD4 since it's unexpected behavior. +AD4 +AD4 +AD4 +AD4 When I change something on the server, I typically expect the client to +AD4 +AD4 see that change in a timely fashion. As it stands now, it won't -- at +AD4 +AD4 least until the inode gets purged from the cache. +AD4 +AD4 See my previous answer: don't do that... If you want a different answer, +AD4 then feel free to propose a caching model, but note that it's going to +AD4 be very hard to deal sensibly with your race condition above. That's the +AD4 reason why we went with the current caching model. +AD4 I think we may be misunderstanding one another. I'm not suggesting that we need to always fetch the label before any use of an inode. The problem I was seeing is similar to what Dave mentioned a month or so ago. Even in the case where we are fetching a security label, we aren't always applying it to the inode. The red flag for me is that we are calling nfs+AF8-setsecurity on existing inodes in nfs+AF8-prime+AF8-dcache, but aren't calling it in +AF8AXw-nfs+AF8-revalidate+AF8-inode. That's why 'ls -Z' works to make the client pick up the label change when the parent directory changes, but 'stat' doesn't. I've just sent a patch that seems to fix the behavior for me in most cases. Does it look reasonable? -- Jeff Layton +ADw-jlayton+AEA-redhat.com+AD4 -- To unsubscribe from this list: send the line "unsubscribe linux-nfs" in the body of a message to majordomo@xxxxxxxxxxxxxxx More majordomo info at http://vger.kernel.org/majordomo-info.html
