Re: [PATCH bpf-next 0/7] Add new way to add BPF LSM hooks

[Anton Protopopov <a.s.protopopov@gmail.com>]

On 26/08/31 03:34PM, Jakub Kicinski wrote:
> On Mon, 31 Aug 2026 11:09:25 +0000 Anton Protopopov wrote:
> > The BPF LSM programs are allowed to attach to LSM hooks. This enables
> > operators to mitigate known bugs without a need to reboot or livepatch
> > machines.  BPF has shown very useful to create such runtime policies.
> > However, many APIs and parts of kernel aren't covered by existing LSM
> > hooks and this would be beneficial to extend the coverage.
> 
> Dunno. Do you have any reason to believe that any of the CVEs your LLM
> gathered for you here are actually getting exploited? Spot checking 
> a few they seem to be mostly driver bugs. What security model do you
> have in mind? Untrusted/malicious users with physical NIC access?

For the untrusted part, here are some existing examples:

  * old untrusted bugs: CVE-2022-50651, CVE-2025-40255
  * "namespace CAP_NET_ADMIN" bugs: CVE-2024-43836, CVE-2025-21921

Also, the recent copy-fail is a stronger example (though not generic
netlink-related).

The general idea is that we gate the common de-multiplexors such that
not only known bugs, but mainly those which will appear in future are
covered. Rough numbers for coverage: around 5% of known cves are
covered with existing LSM hooks. Another ~5-7% can be covered if we
add netlink-related hooks [this series + net/sched, nftables,
rtnetlink, others smaller]. So, statistically, we know where
bugs had appeared in the past, so we can "predict" where new
will appear. Some of them might be severe, so this would be
good to have hooks in place to be able to "mitigate" them.

Just in case, to test this patch locally, I've found around 10 new
ethtool-related bug candidates. I've sent a fix to one, d09c98a6da21
("virtio_net: Fix resize of the RX ring"), which was easy to
reproduce in a VM. Others require specific hardware, though popular,
so I can try to reproduce some, and was planning to do this later.
Of those findings one is unprivileged, it triggers a OOB read.