| Message ID | 20210421171446.785507-1-omosnace@redhat.com (mailing list archive) |
|---|---|
| Headers | show |
| Series | selinux,anon_inodes: Use a separate SELinux class for each type of anon inode | expand |
On Wed, Apr 21, 2021 at 1:14 PM Ondrej Mosnacek <omosnace@redhat.com> wrote: > > This series aims to correct a design flaw in the original anon_inode > SELinux support that would make it hard to write policies for anonymous > inodes once more types of them are supported (currently only userfaultfd > inodes are). A more detailed rationale is provided in the second patch. > > The first patch extends the anon_inode_getfd_secure() function to accept > an additional numeric identifier that represents the type of the > anonymous inode being created, which is passed to the LSMs via > security_inode_init_security_anon(). > > The second patch then introduces a new SELinux policy capability that > allow policies to opt-in to have a separate class used for each type of > anon inode. That means that the "old way" will still ... will what? :) I think it would be a very good idea if you could provide some concrete examples of actual policy problems encountered using the current approach. I haven't looked at these patches very seriously yet, but my initial reaction is not "oh yes, we definitely need this". > I wish I had realized the practical consequences earlier, while the > patches were still under review, but it only started to sink in after > the authors themselves later raised the issue in an off-list > conversation. Even then, I still hoped it wouldn't be that bad, but the > more I thought about how to apply this in an actual policy, the more I > realized how much pain it would be to work with the current design, so > I decided to propose these changes. > > I hope this will be an acceptable solution. > > A selinux-testsuite patch that adapts the userfaultfd test to work also > with the new policy capability enabled will follow. > > Ondrej Mosnacek (2): > LSM,anon_inodes: explicitly distinguish anon inode types > selinux: add capability to map anon inode types to separate classes > > fs/anon_inodes.c | 42 +++++++++++++--------- > fs/userfaultfd.c | 6 ++-- > include/linux/anon_inodes.h | 4 ++- > include/linux/lsm_hook_defs.h | 3 +- > include/linux/security.h | 19 ++++++++++ > security/security.c | 3 +- > security/selinux/hooks.c | 28 ++++++++++++++- > security/selinux/include/classmap.h | 2 ++ > security/selinux/include/policycap.h | 1 + > security/selinux/include/policycap_names.h | 3 +- > security/selinux/include/security.h | 7 ++++ > 11 files changed, 95 insertions(+), 23 deletions(-) > > -- > 2.30.2
On Wed, Apr 21, 2021 at 10:38 PM Paul Moore <paul@paul-moore.com> wrote: > On Wed, Apr 21, 2021 at 1:14 PM Ondrej Mosnacek <omosnace@redhat.com> wrote: > > > > This series aims to correct a design flaw in the original anon_inode > > SELinux support that would make it hard to write policies for anonymous > > inodes once more types of them are supported (currently only userfaultfd > > inodes are). A more detailed rationale is provided in the second patch. > > > > The first patch extends the anon_inode_getfd_secure() function to accept > > an additional numeric identifier that represents the type of the > > anonymous inode being created, which is passed to the LSMs via > > security_inode_init_security_anon(). > > > > The second patch then introduces a new SELinux policy capability that > > allow policies to opt-in to have a separate class used for each type of > > anon inode. That means that the "old way" will still > > ... will what? :) Whoops, I thought I had gone over all the text enough times, but apparently not :) It should have said something along the lines of: ...will still work and will be used by default. > > I think it would be a very good idea if you could provide some > concrete examples of actual policy problems encountered using the > current approach. I haven't looked at these patches very seriously > yet, but my initial reaction is not "oh yes, we definitely need this". An example is provided in patch 2. It is a generalized problem that we would eventually run into in Fedora policy (at least) with the unconfined_domain_type attribute and so far only hypothetical future types of anon inodes.
On Thu, Apr 22, 2021 at 7:40 AM Ondrej Mosnacek <omosnace@redhat.com> wrote: > On Wed, Apr 21, 2021 at 10:38 PM Paul Moore <paul@paul-moore.com> wrote: > > On Wed, Apr 21, 2021 at 1:14 PM Ondrej Mosnacek <omosnace@redhat.com> wrote: > > > > > > This series aims to correct a design flaw in the original anon_inode > > > SELinux support that would make it hard to write policies for anonymous > > > inodes once more types of them are supported (currently only userfaultfd > > > inodes are). A more detailed rationale is provided in the second patch. > > > > > > The first patch extends the anon_inode_getfd_secure() function to accept > > > an additional numeric identifier that represents the type of the > > > anonymous inode being created, which is passed to the LSMs via > > > security_inode_init_security_anon(). > > > > > > The second patch then introduces a new SELinux policy capability that > > > allow policies to opt-in to have a separate class used for each type of > > > anon inode. That means that the "old way" will still > > > > ... will what? :) > > Whoops, I thought I had gone over all the text enough times, but > apparently not :) It should have said something along the lines of: > > ...will still work and will be used by default. That's what I figured from my quick glance at the code, but I wanted to make sure. > > I think it would be a very good idea if you could provide some > > concrete examples of actual policy problems encountered using the > > current approach. I haven't looked at these patches very seriously > > yet, but my initial reaction is not "oh yes, we definitely need this". > > An example is provided in patch 2. It is a generalized problem that we > would eventually run into in Fedora policy (at least) with the > unconfined_domain_type attribute and so far only hypothetical future > types of anon inodes. Yes, I read the example you provided in patch 2, but it was still a little too abstract for my liking. I have the same concern that Stephen mentioned, I was just giving you an opportunity to show that in this case the additional object classes were warranted.