[RFC PATCH 3/3] fs: detect that the i_rwsem has already been taken exclusively

[Dave Chinner]

On Sun, Oct 01, 2017 at 07:42:42PM -0400, Mimi Zohar wrote:
> On Mon, 2017-10-02 at 09:34 +1100, Dave Chinner wrote:
> > On Sun, Oct 01, 2017 at 11:41:48AM -0700, Linus Torvalds wrote:
> > > On Sun, Oct 1, 2017 at 5:08 AM, Mimi Zohar <zohar at linux.vnet.ibm.com> wrote:
> > > >
> > > > Right, re-introducing the iint->mutex and a new i_generation field in
> > > > the iint struct with a separate set of locks should work.  It will be
> > > > reset if the file metadata changes (eg. setxattr, chown, chmod).
> > > 
> > > Note that the "inner lock" could possibly be omitted if the
> > > invalidation can be just a single atomic instruction.
> > > 
> > > So particularly if invalidation could be just an atomic_inc() on the
> > > generation count, there might not need to be any inner lock at all.
> > > 
> > > You'd have to serialize the actual measurement with the "read
> > > generation count", but that should be as simple as just doing a
> > > smp_rmb() between the "read generation count" and "do measurement on
> > > file contents".
> > 
> > We already have a change counter on the inode, which is modified on
> > any data or metadata write (i_version) under filesystem locks.  The
> > i_version counter has well defined semantics - it's required by
> > NFSv4 to increment on any metadata or data change - so we should be
> > able to rely on it's behaviour to implement IMA as well. Filesystems
> > that support i_version are marked with [SB|MS]_I_VERSION in the
> > superblock (IS_I_VERSION(inode)) so it should be easy to tell if IMA
> > can be supported on a specific filesystem (btrfs, ext4, fuse and xfs
> > ATM).
> 
> Recently I received a patch to replace i_version with mtime/atime.

mtime is not guaranteed to change on data writes - the resolution of
the filesystem timestamps may mean mtime only changes once a second
regardless of the number of writes performed to that file. That's
why NFS can't use it as a change attribute, and hence we have
i_version....

>  Now, even more recently, I received a patch that claims that
> i_version is just a performance improvement.

Did you ask them to explain/quantify the performance improvement? 

e.g. Using i_version on XFS slows down performance on small
writes by 2-3% because i_version because all data writes log a
version change rather than only logging a change when mtime updates.
We take that penalty because NFS requires specific change attribute
behaviour, otherwise we wouldn't have implemented it at all in
XFS...

>  For file systems that
> don't support i_version, assume that the file has changed.
> 
> For file systems that don't support i_version, instead of assuming
> that the file has changed, we can at least use i_generation.

I'm not sure what you mean here - the struct inode already has a
i_generation variable. It's a lifecycle indicator used to
discriminate between alloc/free cycles on the same inode number.
i.e. It only changes at inode allocation time, not whenever the data
in the inode changes...

> With Linus' suggested changes, I think this will work nicely.
> 
> > The IMA code should be able to sample that at measurement time and
> > either fail or be retried if i_version changes during measurement.
> > We can then simply make the IMA xattr write conditional on the
> > i_version value being unchanged from the sample the IMA code passes
> > into the filesystem once the filesystem holds all the locks it needs
> > to write the xattr...
> 
> > I note that IMA already grabs the i_version in
> > ima_collect_measurement(), so this shouldn't be too hard to do.
> > Perhaps we don't need any new locks or counterst all, maybe just
> > the ability to feed a version cookie to the set_xattr method?
> 
> The security.ima xattr is normally written out in
> ima_check_last_writer(), not in ima_collect_measurement().

Which, if IIUC, does this to measure and update the xattr:

ima_check_last_writer
  -> ima_update_xattr
    -> ima_collect_measurement
    -> ima_fix_xattr

>  ima_collect_measurement() calculates the file hash for storing in the
> measurement list (IMA-measurement), verifying the hash/signature (IMA-
> appraisal) already stored in the xattr, and auditing (IMA-audit).

Yup, and it samples the i_version before it calculates the hash and
stores it in the iint, which then gets passed to ima_fix_xattr().
Looks like all that is needed is to pass the i_version back to the
filesystem through the xattr call....

IOWs, sample the i_version early while we hold the inode lock and
check the writer count, then if it is the last writer drop the inode
lock and call ima_update_xattr(). The sampled i_version then tells
us if the file has changed before we write the updated xattr...

> The only time that ima_collect_measurement() writes the file xattr is
> in "fix" mode.  Writing the xattr will need to be deferred until after
> the iint->mutex is released.

ima_collect_measurement() doesn't write an xattr at all - it just
reads the file data and calculates the hash.

> There should be no open writers in ima_check_last_writer(), so the
> file shouldn't be changing.

If that code is not holding the inode i_rwsem across
ima_update_xattr(), then the writer check is racy as hell.  We're
trying to get rid of the need for this code to hold the inode lock
to stabilise the writer count for the entire operation, and it looks
to me like everything is there to use the i_version to ensure the
the IMA code doesn't need to hold the inode lock across
ima_collect_measurement() and ima_fix_xattr()...

Cheers,

Dave.
-- 
Dave Chinner
david at fromorbit.com
--
To unsubscribe from this list: send the line "unsubscribe linux-security-module" in
the body of a message to majordomo at vger.kernel.org
More majordomo info at  http://vger.kernel.org/majordomo-info.html