Re: [PATCH v6 4/9] coredump: add coredump socket

[Christian Brauner <brauner@kernel.org>]

On Tue, May 13, 2025 at 10:56:03AM +0200, Lennart Poettering wrote:
> On Mo, 12.05.25 19:14, Kuniyuki Iwashima (kuniyu@amazon.com) wrote:
> 
> > > > Note this version does not use prefix.  Now it requires users to
> > > > just pass the socket cookie via core_pattern so that the kernel
> > > > can verify the peer.
> > >
> > > Exactly - this means the pattern cannot be static in a sysctl.d early
> > > on boot anymore, and has to be set dynamically by <something>.
> >
> > You missed the socket has to be created dynamically by <something>.
> 
> systemd implements socket activation: the generic code in PID 1 can
> bind a socket, and then generically forks off a process (or instances
> of processes for connection-based sockets) once traffic is seen on
> that socket. On a typical, current systemd system, PID 1 does this for
> ~40 sockets by default. The code to bind AF_UNIX or AF_INET/AF_INET6
> sockets is entirely generic.
> 
> Currently, in the existing systemd codebase coredumping is implemented
> via socket activation: the core_pattern handler binary quickly hands
> off the coredump fds to an AF_UNIX socket bound that way, and the
> service behind that does the heavy lifting. Our hope is that with
> Christian's work we can make the kernel deliver the coredumps directly
> to the socket PID1 generically binds, getting rid of one middle man.
> 
> By requiring userspace to echo the SO_COOKIE value into the
> core_pattern sysctl in a special formatting, you define a bespoke
> protocol: it's not just enough to bind a socket (for which the generic
> code in PID1 is good enough), and to write a fixed
> string into a sysctl (for which the generic code in the current
> /etc/sysctl.d/ manager, i.e. systemd-sysctl, works fine). But you
> suddenly are asking from userspace, that some specific tool runs at
> early boot, extracts the socket cookie from PID1 somehow, and writes
> that into sysctl. We'd have to come up with a new tool for that, we
> can no longer use generic tools. And that's the part that Luca doesn't
> like.
> 
> To a large degree I agree with Luca about this. I would much prefer
> Christian's earlier proposal (i.e. to simply define some prefix of
> AF_UNIX abstract namespace addresses as requiring privs to bind),
> because that would enable us to do generic handling in userspace: the
> existing socket binding logic in PID 1, and the existing sysctl.d
> handling in the systemd suite would be good enough to set up
> everything for the coredump handling.
> 
> That said, I'd take what we can get. If enforcing privs on some
> abstract namespace socket address prefix is not acceptable, then we
> can probably make the SO_COOKIE proposal work (Luca: we'd just hook
> some small tool into ExecStartPost= of the .socket unit, and make PID1
> pass the cookie in some env var or so to it; the tool would then just
> echo that env var into the sysctl with the fixed prefix). In my eyes,
> it's not ideal though: it would mean the sysctl data on every instance
> of the system system image would necessarily deviate (because the
> socket cookie is going to be different), which mgmt tools won't like
> (as you cannot compare sysctl state anymore), and we'd have a weak
> conflict of ownership: right now most sysctl settings are managed by
> /etc/sysctl.d/, but the core_pattern suddenly wouldn't be
> anymore. This will create conflicts because suddenly two components
> write to the thing, and will start fighting.
> 
> Hence: I'd *much* prefer Christian's original approach as it does not
> have these issues. But I'll take what I can get, we can make the
> cookie thing work, but it's much uglier.
> 
> I am not sure I understand why enforcing privs on some abstract
> namespace socke address prefix is such an unacceptable idea though.

I prefer the prefix approach as well. It's clean, simple and is safe by
itself and elegant. And it fits into the generic socket activation and
system administration models. I mainly show-cased the cookie model as an
elaborate workaround. It can be done but it's ugly and more difficult to
use.

I do have one more idea how to solve this problem cleanly using regular
socket paths that hopefully pleases everyone.