There's a bunch of failure exits in ffs_fs_mount() with
seriously broken recovery logics. Most of that appears to stem
from misunderstanding of the ->kill_sb() semantics; unlike
->put_super() it is called for *all* superblocks of given type,
no matter how (in)complete the setup had been. ->put_super()
is called only if ->s_root is not NULL; any failure prior to
setting ->s_root will have the call of ->put_super() skipped.
->kill_sb(), OTOH, awaits every superblock that has come from
sget().
Current behaviour of ffs_fs_mount():
We have struct ffs_sb_fill_data data on stack there. We do
ffs_dev = functionfs_acquire_dev_callback(dev_name);
and store that in data.private_data. Then we call mount_nodev(),
passing it ffs_sb_fill() as a callback. That will either fail
outright, or manage to call ffs_sb_fill(). There we allocate an
instance of struct ffs_data, slap the value of ffs_dev (picked
from data.private_data) into ffs->private_data and overwrite
data.private_data by storing ffs into an overlapping member
(data.ffs_data). Then we store ffs into sb->s_fs_info and attempt
to set the rest of the things up (root inode, root dentry, then
create /ep0 there). Any of those might fail. Should that
happen, we get ffs_fs_kill_sb() called before mount_nodev()
returns. If mount_nodev() fails for any reason whatsoever,
we proceed to
functionfs_release_dev_callback(data.ffs_data);
That's broken in a lot of ways. Suppose the thing has failed in
allocation of e.g. root inode or dentry. We have
functionfs_release_dev_callback(ffs);
ffs_data_put(ffs);
done by ffs_fs_kill_sb() (ffs accessed via sb->s_fs_info), followed by
functionfs_release_dev_callback(ffs);
from ffs_fs_mount() (via data.ffs_data). Note that the second
functionfs_release_dev_callback() has every chance to be done to freed memory.
Suppose we fail *before* root inode allocation. What happens then?
ffs_fs_kill_sb() doesn't do anything to ffs (it's either not called at all,
or it doesn't have a pointer to ffs stored in sb->s_fs_info). And
functionfs_release_dev_callback(data.ffs_data);
is called by ffs_fs_mount(), but here we are in nasal daemon country - we
are reading from a member of union we'd never stored into. In practice,
we'll get what we used to store into the overlapping field, i.e. ffs_dev.
And then we get screwed, since we treat it (struct gfs_ffs_obj * in
disguise, returned by functionfs_acquire_dev_callback()) as struct
ffs_data *, pick what would've been ffs_data ->private_data from it
(*well* past the actual end of the struct gfs_ffs_obj - struct ffs_data
is much bigger) and poke in whatever it points to.
FWIW, there's a minor leak on top of all that in case if ffs_sb_fill()
fails on kstrdup() - ffs is obviously forgotten.
The thing is, there is no point in playing all those games with union.
Just allocate and initialize ffs_data *before* calling mount_nodev() and
pass a pointer to it via data.ffs_data. And once it's stored in
sb->s_fs_info, clear data.ffs_data, so that ffs_fs_mount() knows that
it doesn't need to kill the sucker manually - from that point on
we'll have it done by ->kill_sb().
Signed-off-by: Al Viro <viro@zeniv.linux.org.uk>
Acked-by: Michal Nazarewicz <mina86@mina86.com>
Cc: stable <stable@vger.kernel.org> # 3.3+
Signed-off-by: Greg Kroah-Hartman <gregkh@linuxfoundation.org>
2606b28aab