On 2/5/2024 3:10 PM, Paul Moore wrote:
> On Mon, Feb 5, 2024 at 6:01 PM Fan Wu <[email protected]> wrote:
>> On 2/3/2024 2:25 PM, Paul Moore wrote:
>>> On Jan 30, 2024 Fan Wu <[email protected]> wrote:
>>>>
>>>> As is typical with LSMs, IPE uses securityfs as its interface with
>>>> userspace. for a complete list of the interfaces and the respective
>>>> inputs/outputs, please see the documentation under
>>>> admin-guide/LSM/ipe.rst
>>>>
>>>> Signed-off-by: Deven Bowers <[email protected]>
>>>> Signed-off-by: Fan Wu <[email protected]>
>>>> ---
>>>> v2:
>>>> + Split evaluation loop, access control hooks,
>>>> and evaluation loop from policy parser and userspace
>>>> interface to pass mailing list character limit
>>>>
>>>> v3:
>>>> + Move policy load and activation audit event to 03/12
>>>> + Fix a potential panic when a policy failed to load.
>>>> + use pr_warn for a failure to parse instead of an
>>>> audit record
>>>> + Remove comments from headers
>>>> + Add lockdep assertions to ipe_update_active_policy and
>>>> ipe_activate_policy
>>>> + Fix up warnings with checkpatch --strict
>>>> + Use file_ns_capable for CAP_MAC_ADMIN for securityfs
>>>> nodes.
>>>> + Use memdup_user instead of kzalloc+simple_write_to_buffer.
>>>> + Remove strict_parse command line parameter, as it is added
>>>> by the sysctl command line.
>>>> + Prefix extern variables with ipe_
>>>>
>>>> v4:
>>>> + Remove securityfs to reverse-dependency
>>>> + Add SHA1 reverse dependency.
>>>> + Add versioning scheme for IPE properties, and associated
>>>> interface to query the versioning scheme.
>>>> + Cause a parser to always return an error on unknown syntax.
>>>> + Remove strict_parse option
>>>> + Change active_policy interface from sysctl, to securityfs,
>>>> and change scheme.
>>>>
>>>> v5:
>>>> + Cause an error if a default action is not defined for each
>>>> operation.
>>>> + Minor function renames
>>>>
>>>> v6:
>>>> + No changes
>>>>
>>>> v7:
>>>> + Propagating changes to support the new ipe_context structure in the
>>>> evaluation loop.
>>>>
>>>> + Further split the parser and userspace interface changes into
>>>> separate commits.
>>>>
>>>> + "raw" was renamed to "pkcs7" and made read only
>>>> + "raw"'s write functionality (update a policy) moved to "update"
>>>> + introduced "version", "policy_name" nodes.
>>>> + "content" renamed to "policy"
>>>> + changes to allow the compiled-in policy to be treated
>>>> identical to deployed-after-the-fact policies.
>>>>
>>>> v8:
>>>> + Prevent securityfs initialization if the LSM is disabled
>>>>
>>>> v9:
>>>> + Switch to securityfs_recursive_remove for policy folder deletion
>>>>
>>>> v10:
>>>> + Simplify and correct concurrency
>>>> + Fix typos
>>>>
>>>> v11:
>>>> + Correct code comments
>>>>
>>>> v12:
>>>> + Correct locking and remove redundant code
>>>> ---
>>>> security/ipe/Makefile | 2 +
>>>> security/ipe/fs.c | 101 +++++++++
>>>> security/ipe/fs.h | 16 ++
>>>> security/ipe/ipe.c | 3 +
>>>> security/ipe/ipe.h | 2 +
>>>> security/ipe/policy.c | 123 ++++++++++
>>>> security/ipe/policy.h | 9 +
>>>> security/ipe/policy_fs.c | 469 +++++++++++++++++++++++++++++++++++++++
>>>> 8 files changed, 725 insertions(+)
>>>> create mode 100644 security/ipe/fs.c
>>>> create mode 100644 security/ipe/fs.h
>>>> create mode 100644 security/ipe/policy_fs.c
>>>
>>> ...
>>>
>>>> diff --git a/security/ipe/policy.c b/security/ipe/policy.c
>>>> index f22a576a6d68..61fea3e38e11 100644
>>>> --- a/security/ipe/policy.c
>>>> +++ b/security/ipe/policy.c
>>>> @@ -43,6 +71,68 @@ static int set_pkcs7_data(void *ctx, const void *data, size_t len,
>>>> return 0;
>>>> }
>>>>
>>>> +/**
>>>> + * ipe_update_policy - parse a new policy and replace old with it.
>>>> + * @root: Supplies a pointer to the securityfs inode saved the policy.
>>>> + * @text: Supplies a pointer to the plain text policy.
>>>> + * @textlen: Supplies the length of @text.
>>>> + * @pkcs7: Supplies a pointer to a buffer containing a pkcs7 message.
>>>> + * @pkcs7len: Supplies the length of @pkcs7len.
>>>> + *
>>>> + * @text/@textlen is mutually exclusive with @pkcs7/@pkcs7len - see
>>>> + * ipe_new_policy.
>>>> + *
>>>> + * Context: Requires root->i_rwsem to be held.
>>>> + * Return:
>>>> + * * !IS_ERR - The existing policy saved in the inode before update
>>>> + * * -ENOENT - Policy doesn't exist
>>>> + * * -EINVAL - New policy is invalid
>>>> + */
>>>> +struct ipe_policy *ipe_update_policy(struct inode *root,
>>>> + const char *text, size_t textlen,
>>>> + const char *pkcs7, size_t pkcs7len)
>>>> +{
>>>> + int rc = 0;
>>>> + struct ipe_policy *old, *ap, *new = NULL;
>>>> +
>>>> + old = (struct ipe_policy *)root->i_private;
>>>> + if (!old)
>>>> + return ERR_PTR(-ENOENT);
>>>> +
>>>> + new = ipe_new_policy(text, textlen, pkcs7, pkcs7len);
>>>> + if (IS_ERR(new))
>>>> + return new;
>>>> +
>>>> + if (strcmp(new->parsed->name, old->parsed->name)) {
>>>> + rc = -EINVAL;
>>>> + goto err;
>>>> + }
>>>> +
>>>> + if (ver_to_u64(old) > ver_to_u64(new)) {
>>>> + rc = -EINVAL;
>>>> + goto err;
>>>> + }
>>>> +
>>>> + root->i_private = new;
>>>> + swap(new->policyfs, old->policyfs);
>>>
>>> Should the swap() take place with @ipe_policy_lock held?
>>>
>> I think we are safe here because root->i_rwsem is held. Other two
>> operations set_active and delete are also depending on the inode lock.
>>>> + mutex_lock(&ipe_policy_lock);
>>>> + ap = rcu_dereference_protected(ipe_active_policy,
>>>> + lockdep_is_held(&ipe_policy_lock));
>>>> + if (old == ap) {
>>>> + rcu_assign_pointer(ipe_active_policy, new);
>>>> + mutex_unlock(&ipe_policy_lock);
>>>> + synchronize_rcu();
>>>
>>> I'm guessing you are forcing a synchronize_rcu() here because you are
>>> free()'ing @old in the caller, yes? Looking at the code, I only see
>>> one caller, update_policy(). With only one caller, why not free @old
>>> directly in ipe_update_policy()? Do you see others callers that would
>>> do something different?
>>>
>> The call of synchronize_rcu() is because we are updating the current
>> active policy so we need to set the new policy as active.
>
> Unless I'm mistaken, a syncronize_rcu() call only ensures that the
> current task will see the updated value by waiting until all current
> RCU critical sections have finished. Given the mutex involved here I
> don't believe this is necessary, but please correct me if I'm wrong.
>
Sorry for the confusion. I think your previous comment was right, the
call of synchronize_rcu() is to free the old one. And I should put the
free of old just after the synchronize_rcu() call.
Thanks,
Fan