Re: [PATCH v5 23/23] integrity: Switch from rbtree to LSM-managed blob for integrity_iint_cache

From: Roberto Sassu
Date: Wed Nov 29 2023 - 13:47:07 EST


On 11/29/2023 6:22 PM, Paul Moore wrote:
On Wed, Nov 29, 2023 at 7:28 AM Roberto Sassu
<roberto.sassu@xxxxxxxxxxxxxxx> wrote:

On Mon, 2023-11-20 at 16:06 -0500, Paul Moore wrote:
On Mon, Nov 20, 2023 at 3:16 AM Roberto Sassu
<roberto.sassu@xxxxxxxxxxxxxxx> wrote:
On Fri, 2023-11-17 at 15:57 -0500, Paul Moore wrote:
On Nov 7, 2023 Roberto Sassu <roberto.sassu@xxxxxxxxxxxxxxx> wrote:

Before the security field of kernel objects could be shared among LSMs with
the LSM stacking feature, IMA and EVM had to rely on an alternative storage
of inode metadata. The association between inode metadata and inode is
maintained through an rbtree.

Because of this alternative storage mechanism, there was no need to use
disjoint inode metadata, so IMA and EVM today still share them.

With the reservation mechanism offered by the LSM infrastructure, the
rbtree is no longer necessary, as each LSM could reserve a space in the
security blob for each inode. However, since IMA and EVM share the
inode metadata, they cannot directly reserve the space for them.

Instead, request from the 'integrity' LSM a space in the security blob for
the pointer of inode metadata (integrity_iint_cache structure). The other
reason for keeping the 'integrity' LSM is to preserve the original ordering
of IMA and EVM functions as when they were hardcoded.

Prefer reserving space for a pointer to allocating the integrity_iint_cache
structure directly, as IMA would require it only for a subset of inodes.
Always allocating it would cause a waste of memory.

Introduce two primitives for getting and setting the pointer of
integrity_iint_cache in the security blob, respectively
integrity_inode_get_iint() and integrity_inode_set_iint(). This would make
the code more understandable, as they directly replace rbtree operations.

Locking is not needed, as access to inode metadata is not shared, it is per
inode.

Signed-off-by: Roberto Sassu <roberto.sassu@xxxxxxxxxx>
Reviewed-by: Casey Schaufler <casey@xxxxxxxxxxxxxxxx>
Reviewed-by: Mimi Zohar <zohar@xxxxxxxxxxxxx>
---
security/integrity/iint.c | 71 +++++-----------------------------
security/integrity/integrity.h | 20 +++++++++-
2 files changed, 29 insertions(+), 62 deletions(-)

diff --git a/security/integrity/iint.c b/security/integrity/iint.c
index 882fde2a2607..a5edd3c70784 100644
--- a/security/integrity/iint.c
+++ b/security/integrity/iint.c
@@ -231,6 +175,10 @@ static int __init integrity_lsm_init(void)
return 0;
}

+struct lsm_blob_sizes integrity_blob_sizes __ro_after_init = {
+ .lbs_inode = sizeof(struct integrity_iint_cache *),
+};

I'll admit that I'm likely missing an important detail, but is there
a reason why you couldn't stash the integrity_iint_cache struct
directly in the inode's security blob instead of the pointer? For
example:

struct lsm_blob_sizes ... = {
.lbs_inode = sizeof(struct integrity_iint_cache),
};

struct integrity_iint_cache *integrity_inode_get(inode)
{
if (unlikely(!inode->isecurity))
return NULL;
return inode->i_security + integrity_blob_sizes.lbs_inode;
}

It would increase memory occupation. Sometimes the IMA policy
encompasses a small subset of the inodes. Allocating the full
integrity_iint_cache would be a waste of memory, I guess?

Perhaps, but if it allows us to remove another layer of dynamic memory
I would argue that it may be worth the cost. It's also worth
considering the size of integrity_iint_cache, while it isn't small, it
isn't exactly huge either.

On the other hand... (did not think fully about that) if we embed the
full structure in the security blob, we already have a mutex available
to use, and we don't need to take the inode lock (?).

That would be excellent, getting rid of a layer of locking would be significant.

I'm fully convinced that we can improve the implementation
significantly. I just was really hoping to go step by step and not
accumulating improvements as dependency for moving IMA and EVM to the
LSM infrastructure.

I understand, and I agree that an iterative approach is a good idea, I
just want to make sure we keep things tidy from a user perspective,
i.e. not exposing the "integrity" LSM when it isn't required.

Ok, I went back to it again.

I think trying to separate integrity metadata is premature now, too
many things at the same time.

I'm not bothered by the size of the patchset, it is more important
that we do The Right Thing. I would like to hear in more detail why
you don't think this will work, I'm not interested in hearing about
difficult it may be, I'm interested in hearing about what challenges
we need to solve to do this properly.

The right thing in my opinion is to achieve the goal with the minimal set of changes, in the most intuitive way.

Until now, there was no solution that could achieve the primary goal of this patch set (moving IMA and EVM to the LSM infrastructure) and, at the same time, achieve the additional goal you set of removing the 'integrity' LSM.

If you see the diff, the changes compared to v5 that was already accepted by Mimi are very straightforward. If the assumption I made that in the end the 'ima' LSM could take over the role of the 'integrity' LSM, that for me is the preferable option.

Given that the patch set is not doing any design change, but merely moving calls and storing pointers elsewhere, that leaves us with the option of thinking better what to do next, including like you suggested to make IMA and EVM use disjoint metadata.

I started to think, does EVM really need integrity metadata or it can
work without?

The fact is that CONFIG_IMA=n and CONFIG_EVM=y is allowed, so we have
the same problem now. What if we make IMA the one that manages
integrity metadata, so that we can remove the 'integrity' LSM?

I guess we should probably revisit the basic idea of if it even makes
sense to enable EVM without IMA? Should we update the Kconfig to
require IMA when EVM is enabled?

That would be up to Mimi. Also this does not seem the main focus of the patch set.

Regarding the LSM order, I would take Casey's suggestion of introducing
LSM_ORDER_REALLY_LAST, for EVM.

Please understand that I really dislike that we have imposed ordering
constraints at the LSM layer, but I do understand the necessity (the
BPF LSM ordering upsets me the most). I really don't want to see us
make things worse by adding yet another ordering bucket, I would
rather that we document it well and leave it alone ... basically treat
it like the BPF LSM (grrrrrr).

Uhm, that would not be possible right away (the BPF LSM is mutable), remember that we defined LSM_ORDER_LAST so that an LSM can be always enable and placed as last (requested by Mimi)?

Thanks

Roberto