On Tue, Aug 26, 2025 at 3:13 PM Jason Gunthorpe <jgg@xxxxxxxxxx> wrote:
>
> On Tue, Aug 26, 2025 at 03:02:13PM +0000, Pasha Tatashin wrote:
> > I'm trying to understand the drawbacks of the PID-based approach.
> > Could you elaborate on why passing a PID in the RESTORE_FD ioctl is
> > not a good idea?
>
> It will be a major invasive change all over the place in the kernel
> to change things that assume current to do something else. We should
> try to avoid this.
>
> > In this flow, the client isn't providing an arbitrary PID; the trusted
> > luod agent is providing the PID of a process it has an active
> > connection with.
>
> PIDs are wobbly thing, you can never really trust them unless they are
> in a pidfd.
Makes, sense, using a PID by value is fragile due to reuse. Luod would
acquire a pidfd for the client process from its socket connection and
pass that pidfd to the kernel in the RESTORE_FD ioctl. The kernel
would then be operating on a stable, secure handle to the target
process.
> > The idea was to let luod handle the session/security story, and the
> > kernel handle the core preservation mechanism. Adding sessions to the
> > kernel, delegates the management and part of the security model into
> > the kernel. I am not sure if it is necessary, what can be cleanly
> > managed in userspace should stay in userspace.
>
> session fds were an update imagined to allow the kernel to partition
> things the session FD it self could be shared with other processes.
I understand the model you're proposing: luod acts as a factory,
issuing session FDs that are then passed to clients, allowing them to
perform restore operations within their own context. While we can
certainly extend the design to support that, I am still trying to
determine if it's strictly necessary, especially if the same outcome
(correct resource attribution) can be achieved with less kernel
complexity. My primary concern is that functionality that can be
cleanly managed in userspace should remain there.
> I think in the calls the idea was it was reasonable to start without
> sessions fds at all, but in this case we shouldn't be mucking with
> pids or current.
The existing interface, with the addition of passing a pidfd, provides
the necessary flexibility without being invasive. The change would be
localized to the new code that performs the FD retrieval and wouldn't
involve spoofing current or making widespread changes.
For example, to handle cgroup charging for a memfd, the flow inside
memfd_luo_retrieve() would look something like this:
task = get_pid_task(target_pid, PIDTYPE_PID);
mm = get_task_mm(task);
// ...
folio = kho_restore_folio(phys);
// Charge to the target mm, not 'current->mm'
mem_cgroup_charge(folio, mm, ...);
mmput(mm);
put_task_struct(task);
This approach seems quite contained, and does not modify the existing
interfaces. It avoids the need for the kernel to manage the entire
session state and its associated security model.
Pasha
