Re: [PATCH] overlayfs: Trigger file re-evaluation by IMA / EVM after writes

[Christian Brauner]

On Tue, Apr 11, 2023 at 05:32:11AM -0400, Jeff Layton wrote:
> On Tue, 2023-04-11 at 10:38 +0200, Christian Brauner wrote:
> > On Sun, Apr 09, 2023 at 06:12:09PM -0400, Jeff Layton wrote:
> > > On Sun, 2023-04-09 at 17:22 +0200, Christian Brauner wrote:
> > > > On Fri, Apr 07, 2023 at 09:29:29AM -0400, Jeff Layton wrote:
> > > > > > > > > 
> > > > > > > > > I would ditch the original proposal in favor of this 2-line patch shown here:
> > > > > > > > > 
> > > > > > > > > https://lore.kernel.org/linux-integrity/a95f62ed-8b8a-38e5-e468-ecbde3b221af@xxxxxxxxxxxxx/T/#m3bd047c6e5c8200df1d273c0ad551c645dd43232
> > > > > > 
> > > > > > We should cool it with the quick hacks to fix things. :)
> > > > > > 
> > > > > 
> > > > > Yeah. It might fix this specific testcase, but I think the way it uses
> > > > > the i_version is "gameable" in other situations. Then again, I don't
> > > > > know a lot about IMA in this regard.
> > > > > 
> > > > > When is it expected to remeasure? If it's only expected to remeasure on
> > > > > a close(), then that's one thing. That would be a weird design though.
> > > > > 
> > > > > > > > > 
> > > > > > > > > 
> > > > > > > > 
> > > > > > > > Ok, I think I get it. IMA is trying to use the i_version from the
> > > > > > > > overlayfs inode.
> > > > > > > > 
> > > > > > > > I suspect that the real problem here is that IMA is just doing a bare
> > > > > > > > inode_query_iversion. Really, we ought to make IMA call
> > > > > > > > vfs_getattr_nosec (or something like it) to query the getattr routine in
> > > > > > > > the upper layer. Then overlayfs could just propagate the results from
> > > > > > > > the upper layer in its response.
> > > > > > > > 
> > > > > > > > That sort of design may also eventually help IMA work properly with more
> > > > > > > > exotic filesystems, like NFS or Ceph.
> > > > > > > > 
> > > > > > > > 
> > > > > > > > 
> > > > > > > 
> > > > > > > Maybe something like this? It builds for me but I haven't tested it. It
> > > > > > > looks like overlayfs already should report the upper layer's i_version
> > > > > > > in getattr, though I haven't tested that either:
> > > > > > > 
> > > > > > > -----------------------8<---------------------------
> > > > > > > 
> > > > > > > [PATCH] IMA: use vfs_getattr_nosec to get the i_version
> > > > > > > 
> > > > > > > IMA currently accesses the i_version out of the inode directly when it
> > > > > > > does a measurement. This is fine for most simple filesystems, but can be
> > > > > > > problematic with more complex setups (e.g. overlayfs).
> > > > > > > 
> > > > > > > Make IMA instead call vfs_getattr_nosec to get this info. This allows
> > > > > > > the filesystem to determine whether and how to report the i_version, and
> > > > > > > should allow IMA to work properly with a broader class of filesystems in
> > > > > > > the future.
> > > > > > > 
> > > > > > > Reported-by: Stefan Berger <stefanb@xxxxxxxxxxxxx>
> > > > > > > Signed-off-by: Jeff Layton <jlayton@xxxxxxxxxx>
> > > > > > > ---
> > > > > > 
> > > > > > So, I think we want both; we want the ovl_copyattr() and the
> > > > > > vfs_getattr_nosec() change:
> > > > > > 
> > > > > > (1) overlayfs should copy up the inode version in ovl_copyattr(). That
> > > > > >     is in line what we do with all other inode attributes. IOW, the
> > > > > >     overlayfs inode's i_version counter should aim to mirror the
> > > > > >     relevant layer's i_version counter. I wouldn't know why that
> > > > > >     shouldn't be the case. Asking the other way around there doesn't
> > > > > >     seem to be any use for overlayfs inodes to have an i_version that
> > > > > >     isn't just mirroring the relevant layer's i_version.
> > > > > 
> > > > > It's less than ideal to do this IMO, particularly with an IS_I_VERSION
> > > > > inode.
> > > > > 
> > > > > You can't just copy up the value from the upper. You'll need to call
> > > > > inode_query_iversion(upper_inode), which will flag the upper inode for a
> > > > > logged i_version update on the next write. IOW, this could create some
> > > > > (probably minor) metadata write amplification in the upper layer inode
> > > > > with IS_I_VERSION inodes.
> > > > 
> > > > I'm likely just missing context and am curious about this so bear with me. Why
> > > > do we need to flag the upper inode for a logged i_version update? Any required
> > > > i_version interactions should've already happened when overlayfs called into
> > > > the upper layer. So all that's left to do is for overlayfs' to mirror the
> > > > i_version value after the upper operation has returned.
> > > 
> > > > ovl_copyattr() - which copies the inode attributes - is always called after the
> > > > operation on the upper inode has finished. So the additional query seems odd at
> > > > first glance. But there might well be a good reason for it. In my naive
> > > > approach I would've thought that sm along the lines of:
> > > > 
> > > > diff --git a/fs/overlayfs/util.c b/fs/overlayfs/util.c
> > > > index 923d66d131c1..8b089035b9b3 100644
> > > > --- a/fs/overlayfs/util.c
> > > > +++ b/fs/overlayfs/util.c
> > > > @@ -1119,4 +1119,5 @@ void ovl_copyattr(struct inode *inode)
> > > >         inode->i_mtime = realinode->i_mtime;
> > > >         inode->i_ctime = realinode->i_ctime;
> > > >         i_size_write(inode, i_size_read(realinode));
> > > > +       inode_set_iversion_raw(inode, inode_peek_iversion_raw(realinode));
> > > >  }
> > > > 
> > > > would've been sufficient.
> > > > 
> > > 
> > > Nope, because then you wouldn't get any updates to i_version after that
> > > point.
> > > 
> > > Note that with an IS_I_VERSION inode we only update the i_version when
> > > there has been a query since the last update. What you're doing above is
> > > circumventing that mechanism. You'll get the i_version at the time of of
> > > the ovl_copyattr, but there won't be any updates of it after that point
> > > because the QUERIED bit won't end up being set on realinode.
> > 
> > I get all that.
> > But my understanding had been that the i_version value at the time of
> > ovl_copyattr() would be correct. Because when ovl_copyattr() is called
> > the expected i_version change will have been done in the relevant layer
> > includig raising the QUERIED bit. Since the layers are not allowed to be
> > changed outside of the overlayfs mount any change to them can only
> > originate from overlayfs which would necessarily call ovl_copyattr()
> > again. IOW, overlayfs would by virtue of its implementation keep the
> > i_version value in sync.
> >
> > Overlayfs wouldn't even raise SB_I_VERSION. It would indeed just be a
> > cache of i_version of the relevant layer.
> > 
> > > 
> > > 
> > > > Since overlayfs' does explicitly disallow changes to the upper and lower trees
> > > > while overlayfs is mounted it seems intuitive that it should just mirror the
> > > > relevant layer's i_version.
> > > > 
> > > > 
> > > > If we don't do this, then we should probably document that i_version doesn't
> > > > have a meaning yet for the inodes of stacking filesystems.
> > > > 
> > > 
> > > Trying to cache the i_version is counterproductive, IMO, at least with
> > > an IS_I_VERSION inode.
> > > 
> > > The problem is that a query against the i_version has a side-effect. It
> > > has to (atomically) mark the inode for an update on the next change.
> > > 
> > > If you try to cache that value, you'll likely end up doing more queries
> > > than you really need to (because you'll need to keep the cache up to
> > > date) and you'll have an i_version that will necessarily lag the one in
> > > the upper layer inode.
> > > 
> > > The whole point of the change attribute is to get the value as it is at
> > > this very moment so we can check whether there have been changes. A
> > > laggy value is not terribly useful.
> > > 
> > > Overlayfs should just always call the upper layer's ->getattr to get the
> > > version. I wouldn't even bother copying it up in the first place. Doing
> > > so is just encouraging someone to try use the value in the overlayfs
> > > inode, when they really need to go through ->getattr and get the one
> > > from the upper layer.
> > 
> > That seems reasonable to me. I read this as an agreeing with my earlier
> > suggestion to document that i_version doesn't have a meaning for the
> > inodes of stacking filesystems and that we should spell out that
> > vfs_getattr()/->getattr() needs to be used to interact with i_version.
> > 
> 
> It really has no meaning in the stacked filesystem's _inode_. The only
> i_version that has any meaning in a (simple) stacking setup is the upper
> layer inode.

Ok, we're on the same page then.

> 
> > We need to explain to subsystems such as IMA somwhere what the correct
> > way to query i_version agnostically is; independent of filesystem
> > implementation details.
> > 
> > Looking at IMA, it queries the i_version directly without checking
> > whether it's an IS_I_VERSION() inode first. This might make a
> > difference.
> > 
> 
> IMA should just use getattr. That allows the filesystem to present the
> i_version to the caller as it sees fit. Fetching i_version directly
> without testing for IS_I_VERSION is wrong, because you don't know what
> that field contains, or whether the fs supports it at all.

Yep, same page again.

> 
> > Afaict, filesystems that persist i_version to disk automatically raise
> > SB_I_VERSION. I would guess that it be considered a bug if a filesystem
> > would persist i_version to disk and not raise SB_I_VERSION. If so IMA
> > should probably be made to check for IS_I_VERSION() and it will probably
> > get that by switching to vfs_getattr_nosec().
> 
> Not quite. SB_I_VERSION tells the vfs that the filesystem wants the
> kernel to manage the increment of the i_version for it. The filesystem
> is still responsible for persisting that value to disk (if appropriate).

Yes, sure it's the filesystems responsibility to persist it to disk or
not. What I tried to ask was that when a filesystem does persist
i_version to disk then would it be legal to mount it without
SB_I_VERSION (because ext2/ext3 did use to have that mount option)? If
it would then the filesystem would probably need to take care to leave
the i_version field in struct inode uninitialized to avoid confusion or
would that just work? (Mere curiosity, don't feel obligated to go into
detail here. I don't want to hog your time.)

> 
> Switching to vfs_getattr_nosec should make it so IMA doesn't need to
> worry about the gory details of all of this. If STATX_CHANGE_COOKIE is
> set in the response, then it can trust that value. Otherwise, it's no
> good.

Yep, same page again.