Re: [RFC 2/2] ima: Fix detection of read/write violations on stacked filesystems

From: Mimi Zohar
Date: Wed May 01 2024 - 17:14:01 EST


On Tue, 2024-04-23 at 13:06 +0200, Miklos Szeredi wrote:
> On Tue, 16 Apr 2024 at 21:06, Mimi Zohar <zohar@xxxxxxxxxxxxx> wrote:
> > On Tue, 2024-04-16 at 16:46 +0200, Miklos Szeredi wrote:
> > > On Tue, 16 Apr 2024 at 14:18, Mimi Zohar <zohar@xxxxxxxxxxxxx> wrote:
> > > > Originally there was a single measureent unless the filesystem was mounted with
> > > > SB_I_VERSION. With commit a2a2c3c8580a ("ima: Use i_version only when
> > > > filesystem supports it") this changed to always re-measure the file if the
> > > > filesystem wasn't mounted with SB_I_VERSION.
> > >
> > > Does the i_version get stored and compared only while the inode is in memory?
> > >
> > > In that case I think it should be possible to support a version number
> > > for the overlay inode.
> >
> > i_version was insufficient to detect a file change for overlay. Commit
> > b836c4d29f27 ("ima: detect changes to the backing overlay") also compares the
> > i_ino and s_dev as well. Refer to
> > https://lore.kernel.org/lkml/20231025143906.133218-1-zohar@xxxxxxxxxxxxx/.
>
> Which is rather ad-hoc.
>
> I'm talking about returning something in overlay i_version, which
> really indicates the version of the overlay file calculated from the
> i_version of the underlying files. The only issue is making this
> i_version persistent, AFAICS. If that's not needed than the overlayfs
> specific logic in IMA could be moved into overlayfs, where it belongs.

IMA saves the i_version in order to detect whether or not the file has changed.
between one access and another. The i_version value, itself, does not need to be
persistent but needs to be consistent.

>
> > Here in this patch set we need to detect IMA read/write violations, based on the
> > i_readcount/i_writecount. If an overlay file is opened for read, but the
> > backing file is already opened for write, the file measurement is
> > meaningless. An "open-writers" violation needs to be generated; and the IMA
> > measurement list needs to be invalidated.
>
> If there's no other way, then let's implement an API to query the
> writecount that can take overlayfs into account. This is for the VFS
> and/or overlayfs to calculate, not for IMA.

Thanks, that will definitely simplify IMA.

Mimi