new_inode_smack() bogosities

new_inode_smack() bogosities

[Al Viro]

1) in smack_sb_kern_mount()
        isp = inode->i_security;
        if (isp == NULL)
                inode->i_security = new_inode_smack(sp->smk_root);
        else   
                isp->smk_inode = sp->smk_root;
looks very fishy.  How in hell had that inode managed to get created
in the first place without going through smack_inode_alloc_security()?
Is that about mounting an fs instance with in-core superblock that
is older than the call of smack_init()?  That shouldn't be possible due
to initcall ordering, but if that's the case, just what would happen when
we step on *other* inodes on the same fs?  There's a bunch of places where
smack assumes that ->i_security is never NULL...  And if that's really
impossible, WTF is that new_inode_smack() doing there?

2) again in smack_sb_kern_mount()
        spin_lock(&sp->smk_sblock);
        if (sp->smk_initialized != 0) {
                spin_unlock(&sp->smk_sblock);
                return 0;
        }
        sp->smk_initialized = 1;
        spin_unlock(&sp->smk_sblock);
For one thing, security_sb_kern_mount() is serialized by sb->s_umount.
For another, if it wouldn't be... this code wouldn't be able to prevent
a race.  Sure, only one of them will proceed to initialize the sucker.
And another may cheerfully return before the work is actually done...
In any case, it *is* serialized on per-superblock basis.

3) in smk_fill_super(), we have
        root_inode = sb->s_root->d_inode;
        root_inode->i_security = new_inode_smack(smack_known_floor.smk_known);
Again, huh?  This should be called after smack_init() had been done; hell, 
you explicitly say so in the comments:
 * Do not register smackfs if Smack wasn't enabled
 * on boot. We can not put this method normally under the
 * smack_init() code path since the security subsystem get
 * initialized before the vfs caches.
Sounds like a misspelled assignment to ->smk_inode to me (and a leak, while
we are at it)...