Message ID | 20230607152038.469739-4-aleksandr.mikhalitsyn@canonical.com (mailing list archive) |
---|---|
State | New, archived |
Headers | show |
Series | ceph: support idmapped mounts | expand |
On Wed, Jun 7, 2023 at 5:21 PM Alexander Mikhalitsyn <aleksandr.mikhalitsyn@canonical.com> wrote: > > From: Christian Brauner <christian.brauner@ubuntu.com> > > Inode operations that create a new filesystem object such as ->mknod, > ->create, ->mkdir() and others don't take a {g,u}id argument explicitly. > Instead the caller's fs{g,u}id is used for the {g,u}id of the new > filesystem object. > > Cephfs mds creation request argument structures mirror this filesystem > behavior. They don't encode a {g,u}id explicitly. Instead the caller's > fs{g,u}id that is always sent as part of any mds request is used by the > servers to set the {g,u}id of the new filesystem object. > > In order to ensure that the correct {g,u}id is used map the caller's > fs{g,u}id for creation requests. This doesn't require complex changes. > It suffices to pass in the relevant idmapping recorded in the request > message. If this request message was triggered from an inode operation > that creates filesystem objects it will have passed down the relevant > idmaping. If this is a request message that was triggered from an inode > operation that doens't need to take idmappings into account the initial > idmapping is passed down which is an identity mapping and thus is > guaranteed to leave the caller's fs{g,u}id unchanged.,u}id is sent. > > The last few weeks before Christmas 2021 I have spent time not just > reading and poking the cephfs kernel code but also took a look at the > ceph mds server userspace to ensure I didn't miss some subtlety. > > This made me aware of one complication to solve. All requests send the > caller's fs{g,u}id over the wire. The caller's fs{g,u}id matters for the > server in exactly two cases: > > 1. to set the ownership for creation requests > 2. to determine whether this client is allowed access on this server > > Case 1. we already covered and explained. Case 2. is only relevant for > servers where an explicit uid access restriction has been set. That is > to say the mds server restricts access to requests coming from a > specific uid. Servers without uid restrictions will grant access to > requests from any uid by setting MDS_AUTH_UID_ANY. > > Case 2. introduces the complication because the caller's fs{g,u}id is > not just used to record ownership but also serves as the {g,u}id used > when checking access to the server. > > Consider a user mounting a cephfs client and creating an idmapped mount > from it that maps files owned by uid 1000 to be owned uid 0: > > mount -t cephfs -o [...] /unmapped > mount-idmapped --map-mount 1000:0:1 /idmapped > > That is to say if the mounted cephfs filesystem contains a file "file1" > which is owned by uid 1000: > > - looking at it via /unmapped/file1 will report it as owned by uid 1000 > (One can think of this as the on-disk value.) > - looking at it via /idmapped/file1 will report it as owned by uid 0 > > Now, consider creating new files via the idmapped mount at /idmapped. > When a caller with fs{g,u}id 1000 creates a file "file2" by going > through the idmapped mount mounted at /idmapped it will create a file > that is owned by uid 1000 on-disk, i.e.: > > - looking at it via /unmapped/file2 will report it as owned by uid 1000 > - looking at it via /idmapped/file2 will report it as owned by uid 0 > > Now consider an mds server that has a uid access restriction set and > only grants access to requests from uid 0. > > If the client sends a creation request for a file e.g. /idmapped/file2 > it will send the caller's fs{g,u}id idmapped according to the idmapped > mount. So if the caller has fs{g,u}id 1000 it will be mapped to {g,u}id > 0 in the idmapped mount and will be sent over the wire allowing the > caller access to the mds server. > > However, if the caller is not issuing a creation request the caller's > fs{g,u}id will be send without the mount's idmapping applied. So if the > caller that just successfully created a new file on the restricted mds > server sends a request as fs{g,u}id 1000 access will be refused. This > however is inconsistent. > > From my perspective the root of the problem lies in the fact that > creation requests implicitly infer the ownership from the {g,u}id that > gets sent along with every mds request. > > I have thought of multiple ways of addressing this problem but the one I > prefer is to give all mds requests that create a filesystem object a > proper, separate {g,u}id field entry in the argument struct. This is, > for example how ->setattr mds requests work. > > This way the caller's fs{g,u}id can be used consistenly for server > access checks and is separated from the ownership for new filesystem > objects. > > Servers could then be updated to refuse creation requests whenever the > {g,u}id used for access checking doesn't match the {g,u}id used for > creating the filesystem object just as is done for setattr requests on a > uid restricted server. But I am, of course, open to other suggestions. > > Cc: Xiubo Li <xiubli@redhat.com> > Cc: Jeff Layton <jlayton@kernel.org> > Cc: Ilya Dryomov <idryomov@gmail.com> > Cc: ceph-devel@vger.kernel.org > Signed-off-by: Christian Brauner <christian.brauner@ubuntu.com> > Signed-off-by: Alexander Mikhalitsyn <aleksandr.mikhalitsyn@canonical.com> > --- > fs/ceph/mds_client.c | 22 ++++++++++++++++++---- > 1 file changed, 18 insertions(+), 4 deletions(-) > > diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c > index 810c3db2e369..e4265843b838 100644 > --- a/fs/ceph/mds_client.c > +++ b/fs/ceph/mds_client.c > @@ -2583,6 +2583,8 @@ static struct ceph_msg *create_request_message(struct ceph_mds_session *session, > void *p, *end; > int ret; > bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME); > + kuid_t caller_fsuid; > + kgid_t caller_fsgid; > > ret = set_request_path_attr(req->r_inode, req->r_dentry, > req->r_parent, req->r_path1, req->r_ino1.ino, > @@ -2651,10 +2653,22 @@ static struct ceph_msg *create_request_message(struct ceph_mds_session *session, > > head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch); > head->op = cpu_to_le32(req->r_op); > - head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, > - req->r_cred->fsuid)); > - head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, > - req->r_cred->fsgid)); > + /* > + * Inode operations that create filesystem objects based on the > + * caller's fs{g,u}id like ->mknod(), ->create(), ->mkdir() etc. don't > + * have separate {g,u}id fields in their respective structs in the > + * ceph_mds_request_args union. Instead the caller_{g,u}id field is > + * used to set ownership of the newly created inode by the mds server. > + * For these inode operations we need to send the mapped fs{g,u}id over > + * the wire. For other cases we simple set req->r_mnt_idmap to the > + * initial idmapping meaning the unmapped fs{g,u}id is sent. > + */ > + caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns, > + VFSUIDT_INIT(req->r_cred->fsuid)); > + caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns, > + VFSGIDT_INIT(req->r_cred->fsgid)); > + head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, caller_fsuid)); > + head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, caller_fsgid)); > head->ino = cpu_to_le64(req->r_deleg_ino); > head->args = req->r_args; > > -- > 2.34.1 > Probably it's worth adding to a commit message or cover letter, but let it be there for now. Explanation/demonstration from this thread: https://lore.kernel.org/lkml/CAEivzxefBRPozUPQxYgVh0gOpjsovtBuJ3w9BoqSizpST_YxTA@mail.gmail.com/#t 1. Mount cephfs mount.ceph admin@XYZ.cephfs=/ /mnt/ceph -o mon_addr=127.0.0.1:6789,secret=very_secret_key 2. Make 1000:1000 a root dentry owner (it will be convenient because we want to use mapping 1000:0:1 for simplicity) chown 1000:1000 /mnt/ceph 3. create an idmapped mount based on a regular /mnt/ceph mount using a mount-idmapped tool that was written by Christian. [ taken from https://raw.githubusercontent.com/brauner/mount-idmapped/master/mount-idmapped.c ] ./mount-idmapped --map-mount b:1000:0:1 /mnt/ceph /mnt/ceph_idmapped "b" stands for "both", so we are creating a mapping of length 1 for both UID and GID. 1000 is a UID/GID "on-disk", 0 is a mapped UID/GID. 4. Just to be precise, let's look at which UID/GID we have now. root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph total 4 drwxrwxrwx 2 1000 1000 0 Jun 1 17:51 . drwxr-xr-x 4 0 0 4096 Jun 1 16:55 .. root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph_idmapped total 4 drwxrwxrwx 2 0 0 0 Jun 1 17:51 . drwxr-xr-x 4 0 0 4096 Jun 1 16:55 .. 5. Now let's create a bunch of files with different owners and through different mounts (idmapped/non-idmapped). 5.1. Create a file from 0:0 through the idmapped mount (it should appear as 1000:1000 on disk) root@ubuntu:/home/ubuntu# sudo -u#0 -g#0 touch /mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid0 5.2. Create a file from 1000:1000 through the idmapped mount (should fail because 1000:1000 is not a valid UID/GID as it can't be mapped back to the "on-disk" UID/GID set). root@ubuntu:/home/ubuntu# sudo -u#1000 -g#1000 touch /mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid1000 touch: cannot touch '/mnt/ceph_idmapped/created_through_idmapped_mnt_with_uid1000': Value too large for defined data type ... and we've got EOVERFLOW. That's correct! 5.3. Create a file from 0:0 but through the regular mount. (it should appear as overflowuid(=65534) in idmapped mount, because 0:0 on-disk is not mapped to the UID/GID set). root@ubuntu:/home/ubuntu# sudo -u#0 -g#0 touch /mnt/ceph/created_directly_with_uid0 5.4. Create a file from 1000:1000 but through the regular mount. (it should appear as 0:0 in idmapped mount, because 1000 (on-disk) mapped to 0). root@ubuntu:/home/ubuntu# sudo -u#1000 -g#1000 touch /mnt/ceph/created_directly_with_uid1000 6. Now let's look on the result: root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph total 4 drwxrwxrwx 2 1000 1000 3 Jun 1 17:54 . drwxr-xr-x 4 0 0 4096 Jun 1 16:55 .. -rw-r--r-- 1 0 0 0 Jun 1 17:54 created_directly_with_uid0 -rw-rw-r-- 1 1000 1000 0 Jun 1 17:54 created_directly_with_uid1000 -rw-r--r-- 1 1000 1000 0 Jun 1 17:53 created_through_idmapped_mnt_with_uid0 root@ubuntu:/home/ubuntu# ls -lan /mnt/ceph_idmapped total 4 drwxrwxrwx 2 0 0 3 Jun 1 17:54 . drwxr-xr-x 4 0 0 4096 Jun 1 16:55 .. -rw-r--r-- 1 65534 65534 0 Jun 1 17:54 created_directly_with_uid0 -rw-rw-r-- 1 0 0 0 Jun 1 17:54 created_directly_with_uid1000 -rw-r--r-- 1 0 0 0 Jun 1 17:53 created_through_idmapped_mnt_with_uid0
diff --git a/fs/ceph/mds_client.c b/fs/ceph/mds_client.c index 810c3db2e369..e4265843b838 100644 --- a/fs/ceph/mds_client.c +++ b/fs/ceph/mds_client.c @@ -2583,6 +2583,8 @@ static struct ceph_msg *create_request_message(struct ceph_mds_session *session, void *p, *end; int ret; bool legacy = !(session->s_con.peer_features & CEPH_FEATURE_FS_BTIME); + kuid_t caller_fsuid; + kgid_t caller_fsgid; ret = set_request_path_attr(req->r_inode, req->r_dentry, req->r_parent, req->r_path1, req->r_ino1.ino, @@ -2651,10 +2653,22 @@ static struct ceph_msg *create_request_message(struct ceph_mds_session *session, head->mdsmap_epoch = cpu_to_le32(mdsc->mdsmap->m_epoch); head->op = cpu_to_le32(req->r_op); - head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, - req->r_cred->fsuid)); - head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, - req->r_cred->fsgid)); + /* + * Inode operations that create filesystem objects based on the + * caller's fs{g,u}id like ->mknod(), ->create(), ->mkdir() etc. don't + * have separate {g,u}id fields in their respective structs in the + * ceph_mds_request_args union. Instead the caller_{g,u}id field is + * used to set ownership of the newly created inode by the mds server. + * For these inode operations we need to send the mapped fs{g,u}id over + * the wire. For other cases we simple set req->r_mnt_idmap to the + * initial idmapping meaning the unmapped fs{g,u}id is sent. + */ + caller_fsuid = from_vfsuid(req->r_mnt_idmap, &init_user_ns, + VFSUIDT_INIT(req->r_cred->fsuid)); + caller_fsgid = from_vfsgid(req->r_mnt_idmap, &init_user_ns, + VFSGIDT_INIT(req->r_cred->fsgid)); + head->caller_uid = cpu_to_le32(from_kuid(&init_user_ns, caller_fsuid)); + head->caller_gid = cpu_to_le32(from_kgid(&init_user_ns, caller_fsgid)); head->ino = cpu_to_le64(req->r_deleg_ino); head->args = req->r_args;