On Mon, 7 Nov 2016, Jeff Layton wrote:
> On Mon, 2016-11-07 at 14:36 +0000, Sage Weil wrote:
> > On Mon, 7 Nov 2016, Jeff Layton wrote:
> > > On Mon, 2016-11-07 at 14:05 +0000, Sage Weil wrote:
> > > > On Mon, 7 Nov 2016, Jeff Layton wrote:
> > > > > On Mon, 2016-11-07 at 16:43 +0800, Yan, Zheng wrote:
> > > > > > On Fri, Nov 4, 2016 at 8:57 PM, Jeff Layton <jlayton@xxxxxxxxxx> wrote:
> > > > > > >
> > > > > > > On Fri, 2016-11-04 at 07:34 -0400, Jeff Layton wrote:
> > > > > > > >
> > > > > > > > The userland ceph has MClientCaps at struct version 9. This brings the
> > > > > > > > kernel up the same version.
> > > > > > > >
> > > > > > > > With this change, we have to start tracking the btime and change_attr,
> > > > > > > > so that the client can pass back sane values in cap messages. The
> > > > > > > > client doesn't care about the btime at all, so this is just passed
> > > > > > > > around, but the change_attr is used when ceph is exported via NFS.
> > > > > > > >
> > > > > > > > For now, the new "sync" parm is left at 0, to preserve the existing
> > > > > > > > behavior of the client.
> > > > > > > >
> > > > > > > > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
> > > > > > > > ---
> > > > > > > > fs/ceph/caps.c | 33 +++++++++++++++++++++++++--------
> > > > > > > > 1 file changed, 25 insertions(+), 8 deletions(-)
> > > > > > > >
> > > > > > > > diff --git a/fs/ceph/caps.c b/fs/ceph/caps.c
> > > > > > > > index 6e99866b1946..452f5024589f 100644
> > > > > > > > --- a/fs/ceph/caps.c
> > > > > > > > +++ b/fs/ceph/caps.c
> > > > > > > > @@ -991,9 +991,9 @@ struct cap_msg_args {
> > > > > > > > struct ceph_mds_session *session;
> > > > > > > > u64 ino, cid, follows;
> > > > > > > > u64 flush_tid, oldest_flush_tid, size, max_size;
> > > > > > > > - u64 xattr_version;
> > > > > > > > + u64 xattr_version, change_attr;
> > > > > > > > struct ceph_buffer *xattr_buf;
> > > > > > > > - struct timespec atime, mtime, ctime;
> > > > > > > > + struct timespec atime, mtime, ctime, btime;
> > > > > > > > int op, caps, wanted, dirty;
> > > > > > > > u32 seq, issue_seq, mseq, time_warp_seq;
> > > > > > > > kuid_t uid;
> > > > > > > > @@ -1026,13 +1026,13 @@ static int send_cap_msg(struct cap_msg_args *arg)
> > > > > > > >
> > > > > > > > /* flock buffer size + inline version + inline data size +
> > > > > > > > * osd_epoch_barrier + oldest_flush_tid */
> > > > > > > > - extra_len = 4 + 8 + 4 + 4 + 8;
> > > > > > > > + extra_len = 4 + 8 + 4 + 4 + 8 + 4 + 4 + 4 + 8 + 8 + 1;
> > > > > > > > msg = ceph_msg_new(CEPH_MSG_CLIENT_CAPS, sizeof(*fc) + extra_len,
> > > > > > > > GFP_NOFS, false);
> > > > > > > > if (!msg)
> > > > > > > > return -ENOMEM;
> > > > > > > >
> > > > > > > > - msg->hdr.version = cpu_to_le16(6);
> > > > > > > > + msg->hdr.version = cpu_to_le16(9);
> > > > > > > > msg->hdr.tid = cpu_to_le64(arg->flush_tid);
> > > > > > > >
> > > > > > > > fc = msg->front.iov_base;
> > > > > > > > @@ -1068,17 +1068,30 @@ static int send_cap_msg(struct cap_msg_args *arg)
> > > > > > > > }
> > > > > > > >
> > > > > > > > p = fc + 1;
> > > > > > > > - /* flock buffer size */
> > > > > > > > + /* flock buffer size (version 2) */
> > > > > > > > ceph_encode_32(&p, 0);
> > > > > > > > - /* inline version */
> > > > > > > > + /* inline version (version 4) */
> > > > > > > > ceph_encode_64(&p, arg->inline_data ? 0 : CEPH_INLINE_NONE);
> > > > > > > > /* inline data size */
> > > > > > > > ceph_encode_32(&p, 0);
> > > > > > > > - /* osd_epoch_barrier */
> > > > > > > > + /* osd_epoch_barrier (version 5) */
> > > > > > > > ceph_encode_32(&p, 0);
> > > > > > > > - /* oldest_flush_tid */
> > > > > > > > + /* oldest_flush_tid (version 6) */
> > > > > > > > ceph_encode_64(&p, arg->oldest_flush_tid);
> > > > > > > >
> > > > > > > > + /* caller_uid/caller_gid (version 7) */
> > > > > > > > + ceph_encode_32(&p, (u32)-1);
> > > > > > > > + ceph_encode_32(&p, (u32)-1);
> > > > > > >
> > > > > > > A bit of self-review...
> > > > > > >
> > > > > > > Not sure if we want to set the above to something else -- maybe 0 or to
> > > > > > > current's creds? That may not always make sense though (during e.g.
> > > > > > > writeback).
> > > > > > >
> > > > >
> > > > > Looking further, I'm not quite sure I understand why we send creds at
> > > > > all in cap messages. Can you clarify where that matters?
> > > > >
> > > > > The way I look at it, would be to consider caps to be something like a
> > > > > more granular NFS delegation or SMB oplock.
> > > > >
> > > > > In that light, a cap flush is just the client sending updated attrs for
> > > > > the exclusive caps that it has already been granted. Is there a
> > > > > situation where we would ever want to refuse that update?
> > > >
> > > > A chmod or chown can be done locally if you have excl caps and flushed
> > > > back to the MDS via a caps message. We need to verify the user has
> > > > permission to make the change.
> > > >
> > >
> > > My take is that once the MDS has delegated Ax to the client, then it's
> > > effectively trusting the client to handle permissions enforcement
> > > correctly. I don't see why we should second guess that.
> > >
> > > > > Note that nothing ever checks the return code for _do_cap_update in the
> > > > > userland code. If the permissions check fails, then we'll end up
> > > > > silently dropping the updated attrs on the floor.
> > > >
> > > > Yeah. This was mainly for expediency... the protocol assumes that flushes
> > > > don't fail. Given that the client does it's own permissions check, I
> > > > think the way to improve this is to have it prevent the flush in the first
> > > > place, so that it's only nefarious clients that are effected (and who
> > > > cares if they get confused). I don't think we have a particularly good
> > > > way to tell the client it can't, say, sudo chmod 0:0 a file, though.
> > > >
> > >
> > > Sorry, I don't quite follow. How would we prevent the flush from a
> > > nefarious client (which is not something we can really control)?
> > >
> > > In any case...ISTM that the permissions check in _do_cap_update ought to
> > > be replaced by a cephx key check. IOW, what we really want to know is
> > > whether the client is truly the one to which we delegated the caps. If
> > > so, then we sort of have to trust that it's doing the right thing with
> > > respect to permissions checking here.
> >
> > The capability can say "you are allowed to be uid 1000 or uid 1020." We
> > want to delegate the EXCL caps to the client so that a create + chmod +
> > chown + write can all happen efficiently, but we still need to ensure that
> > the values they set are legal (a permitted uid/gid combo).
> >
> > A common example would be user workstations that are allowed access to
> > /home/user and restricted via their mds caps to their uid/gid. We need to
> > prevent them from doing a 'sudo chown 0:0 foo'...
> >
> >
>
>
> On what basis do you make such a decision though? For instance, NFS does
> root-squashing which is (generally) a per-export+per-client thing.
> It sounds like you're saying that ceph has different semantics here?
I don't remember all the specifics :(, but I do remember a long discussion
about whether root_squash made sense and ultimately deciding that the
semantics were weak enough to not bother implementing. Instead, we
explicitly enumerate the uids/gids that are allowed and just match against
that.
> (cc'ing Greg here)
>
> Also, chown (at least under POSIX) is reserved for superuser only, and
> now that I look, I think this check in MDSAuthCaps::is_capable may be
> wrong:
>
> // chown/chgrp
> if (mask & MAY_CHOWN) {
> if (new_uid != caller_uid || // you can't chown to someone else
> inode_uid != caller_uid) { // you can't chown from someone else
> continue;
> }
> }
>
> Shouldn't this just be a check for whether the caller_uid is 0 (or
> whatever the correct check for the equivalent to the kernel's CAP_CHOWN
> would be)?
In the single-user case root_squash would be sufficient, provided that we
also somehow ensured that whatever uid the new file was assigned was set
to the "correct" value. It would have been a kludgey and/or limited
solution, though. And that only addresses the UID portion...
Instead, we validate whatever values come back against what is in the cap
(uid *and* gid) to ensure it is allowed. Note that the capability
can take a list of gids, e.g.,
"allow rw path=/foo uid=1 gids=1,2,3"
This is also laying some of the groundwork for the future world in which
the client *host* gets a capability saying something like "I trust you do
be an honest broker of kerberos identities" and to pass per-user tickets
to the MDS as users come and go from the host, such that a given request
will be validated against some specific user's session (kerberos
identity). Of course, in such a situation a malicious host could allow
one user to impersonate some other user that is also logged into the host,
but that is a still a much stronger model than simply trustly the
host/client completely (ala AUTH_UNIX in NFS-land). The idea is that
eventually the check_caps() would validate against a dynamic "user"
session as well as the client session capability which would have per-user
context pulled from kerberos or LDAP or whatever.
In any case, the core issue is that it wouldn't be sufficient to, say,
refuse to issue EXCL caps on a file to the client unless we are prepare to
trust any values they send back. Or I guess it would, but it would rule
out lots of common scenarios where there are huge benefits
(performance-wise) to issueing the EXCL caps. That's why we have
to validate the contents of the metadata in the cap flush messages...
Does that make sense?
sage
