On Wed, 2023-11-29 at 19:46 +0100, Roberto Sassu wrote:
> On 11/29/2023 6:22 PM, Paul Moore wrote:
> > On Wed, Nov 29, 2023 at 7:28 AM Roberto Sassu
> > <[email protected]> wrote:
> >>
> >> On Mon, 2023-11-20 at 16:06 -0500, Paul Moore wrote:
> >>> On Mon, Nov 20, 2023 at 3:16 AM Roberto Sassu
> >>> <[email protected]> wrote:
> >>>> On Fri, 2023-11-17 at 15:57 -0500, Paul Moore wrote:
> >>>>> On Nov 7, 2023 Roberto Sassu <[email protected]> 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 <[email protected]>
> >>>>>> Reviewed-by: Casey Schaufler <[email protected]>
> >>>>>> Reviewed-by: Mimi Zohar <[email protected]>
> >>>>>> ---
> >>>>>> 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.
First you suggested lumping IMA and EVM together, dropping EVM
entirely. Now you're suggesting making EVM dependent on IMA. Please
stop. EVM and IMA should remain independent of each other. The first
user of EVM is IMA.
> >> 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)?
Making EVM a full fledged LSM was contingent on two things - EVM always
being enabled if configured and being the last LSM. Using capability
as a precedent for ordering requirement, Mickaël suggested defining
LSM_ORDER_LAST, which you agreed to. It sounds like you're
backtracking on an agreement.
Mimi