inode_change_ok
Why are there no calls to inode_change_ok in nfs (on the client), but there are in most other filesytsems? Seems like there are some cases in nfs in which a local permission check is done via a call to nfs_permission which calls generic_permission ... if that is the case why not do a call to inode_change_ok in similar cases?
For the case of cifs vfs, which is also missing this call - I was thinking of adding something like:
if (!cifs_sb->mnt_cifs_flags & CIFS_MOUNT_NO_PERM)
inode_change_ok(direntry->d_inode, attrs);to fs/cifs/inode.c near the beginning of cifs_setattr. Although the permissions (actually ACLs) are checked on the server during setattr from cifs to Samba or cifs servers such as Windows, it is common for convenience for users to mount with one id, rather than authenticating each user (so that there are multiple smb uids) in which case the permission check on setattr on the client can be important since apparently the ".permission" entry point does not seem to get invoked in the chown/chmod path.
Quoted 5 lines · Steve French
Why are there no calls to
inode_change_okin nfs (on the client), but
there are in most other filesytsems? Seems like there are some cases
in nfs in which a local permission check is done via a call tonfs_permissionwhich callsgeneric_permission... if that is the case why not do a call toinode_change_okin similar cases?
Under the NFS model, the server manages the permissions, not the client.
The purpose of inode_change_ok() is to perform a load of local checks which are simply alien to that model:
a) your capabilities don't mean anything to the server. Its decision to
grant the ability to change owner of a file is based on your credentials, not your capabilities.
b) Even the uid/gid checks don't take into account the fact that the
server may be mapping you into different users/groups (c.f. root squashing etc.).
All, in all, a call to inode_change_ok() would at best be redundant, and at worst, be plain incorrect.
Cheers,
Trond
Quoted 16 lines · Trond Myklebust
Quoted 5 lines · Steve French
Why are there no calls to
inode_change_okin nfs (on the client), but
there are in most other filesytsems? Seems like there are some cases
in nfs in which a local permission check is done via a call tonfs_permissionwhich callsgeneric_permission... if that is the case why not do a call toinode_change_okin similar cases?Under the NFS model, the server manages the permissions, not the client.
The purpose of
inode_change_ok()is to perform a load of local checks which are simply alien to that model:a) your capabilities don't mean anything to the server. Its decision to
grant the ability to change owner of a file is based on your credentials, not your capabilities.
Quoted 6 lines
b) Even the uid/gid checks don't take into account the fact that the
server may be mapping you into different users/groups (c.f. root squashing etc.).All, in all, a call to
inode_change_ok()would at best be redundant, and at worst, be plain incorrect.
BTW: The call to permission() is thoroughly redundant too, and really wants to be optimised away. Permissions checking is done by the server on the actual SETATTR RPC call...
Cheers,
Trond
Quoted 11 lines · Trond Myklebust
CAP_CHOWNwould therefore only be useful in the corner case where the server is allowing full root privileges to a client, and where the client is using capabilities to limit the root account's (or a setuid processes') ability to exercise those full privileges.Cheers,
Trond
It sounds like I do need to make the change to cifs to add the optional (based on the mount parm the admin can choose which model) call to inode_change_ok to make the permissions check consistent when "noperm" is not enabled on that mount , but I was trying to figure out how important that change to the cifs code was and whether it was a priority to get in. I realize that this corner case is far less common in the wild for nfs. The corner case you describe above did seem to come up today for cifs though (where the admin had in effect, trusted the client, and allowed the client to mount as root with full root priv on the server - I can imagine this being valid in their particular case, with a trusted client os, over a trusted lan, but obviously would not be the most common scenario). Obviously they did not turn on multiusermount in /proc/fs/cifs (ie the ability to send different uids on the wire depending on the uid of the client calling process) so the only perm check that would be useful that was occuring at the client the way they had chosen to configure/mount was happening in generic_permission in the client. In their particular configuration open and other paths in the vfs that call permission behaved fine and the permission checks worked as expected - but chown always worked (which is of course wrong) since the vfs does not call permission but expects the vfs (if at all) to call inode_change_ok - so open failing, but chown working was confusing to them.
Of course it does not make it much easier trying to compare with other filesystems - ncpfs does seem to call both (generic_permission inode_change_ok) and but the afs model (openafs) is far more complicated and somewhat confusing (and interestingly afs does not call generic_permission but does call inode_change_ok) so it is not really possible to draw too many conclusions for how afs approached the same two usage models (apparently afs also had a third usage model based on machine address)