On Thu, May 24, 2018 at 09:41:08AM +0200, Jesper Dangaard Brouer wrote:

On Wed, 23 May 2018 15:02:45 -0700
Alexei Starovoitov [off-list ref] wrote:

quoted

On Wed, May 23, 2018 at 02:18:19PM +0200, Eugene Syromiatnikov wrote:

quoted

Some BPF sysctl knobs affect the loading of BPF programs, and during
system boot/init stages these sysctls are not yet configured.
A concrete example is systemd, that has implemented loading of BPF
programs.

Thus, to allow controlling these setting at early boot, this patch set
adds the ability to change the default setting of these sysctl knobs
as well as option to override them via a boot-time kernel parameter
(in order to avoid rebuilding kernel each time a need of changing these
defaults arises).

The sysctl knobs in question are kernel.unprivileged_bpf_disable,
net.core.bpf_jit_harden, and net.core.bpf_jit_kallsyms.  

- systemd is root. today it only uses cgroup-bpf progs which require root,
  so disabling unpriv during boot time makes no difference to systemd.
  what is the actual reason to present time?

- say in the future systemd wants to use so_reuseport+bpf for faster
  networking. With unpriv disable during boot, it will force systemd
  to do such networking from root, which will lower its security barrier.
  How that make sense?

- bpf_jit_kallsyms sysctl has immediate effect on loaded programs.
  Flipping it during the boot or right after or any time after
  is the same thing. Why add such boot flag then?

- jit_harden can be turned on by systemd. so turning it during the boot
  will make systemd progs to be constant blinded.
  Constant blinding protects kernel from unprivileged JIT spraying.
  Are you worried that systemd will attack the kernel with JIT spraying?

I think you are missing that, we want the ability to change these
defaults in-order to avoid depending on /etc/sysctl.conf settings, and
that the these sysctl.conf setting happen too late.

What does it mean 'happens too late' ?
Too late for what?
sysctl.conf has plenty of system critical knobs like
kernel.perf_event_paranoid, kernel.core_pattern, etc
The behavior of the host is drastically different after sysctl config
is applied.

For example with jit_harden, there will be a difference between the
loaded BPF program that got loaded at boot-time with systemd (no
constant blinding) and when someone reloads that systemd service after
/etc/sysctl.conf have been evaluated and setting bpf_jit_harden (now
slower due to constant blinding).   This is inconsistent behavior.

net.core.bpf_jit_harden can be flipped back and forth at run-time,
so bpf progs before and after will be either blinded or not.
I don't see any inconsistency.
In general I think bootparams should be used only for things
like kpti=on/off that cannot be set by sysctl.