Re: [PATCH v3 17/30] liveupdate: luo_files: luo_ioctl: Unregister all FDs on device close

[Pasha Tatashin <pasha.tatashin@soleen.com>]

On Wed, Aug 27, 2025 at 11:34 AM Pratyush Yadav <pratyush@kernel.org> wrote:
>
> Hi Pasha,
>
> On Thu, Aug 07 2025, Pasha Tatashin wrote:
>
> > Currently, a file descriptor registered for preservation via the remains
> > globally registered with LUO until it is explicitly unregistered. This
> > creates a potential for resource leaks into the next kernel if the
> > userspace agent crashes or exits without proper cleanup before a live
> > update is fully initiated.
> >
> > This patch ties the lifetime of FD preservation requests to the lifetime
> > of the open file descriptor for /dev/liveupdate, creating an implicit
> > "session".
> >
> > When the /dev/liveupdate file descriptor is closed (either explicitly
> > via close() or implicitly on process exit/crash), the .release
> > handler, luo_release(), is now called. This handler invokes the new
> > function luo_unregister_all_files(), which iterates through all FDs
> > that were preserved through that session and unregisters them.
>
> Why special case files here? Shouldn't you undo all the serialization
> done for all the subsystems?

Good point, subsystems should also be cancelled, and system should be
brought back to normal state. However, with session support, we will
be dropping only FDs that belong to a specific session when its FD is
closed, or all FDs+subsystems when closing /dev/liveupdate.

> Anyway, this is buggy. I found this when testing the memfd patches. If
> you preserve a memfd and close the /dev/liveupdate FD before reboot,
> luo_unregister_all_files() calls the cancel callback, which calls
> kho_unpreserve_folio(). But kho_unpreserve_folio() fails because KHO is
> still in finalized state. This doesn't happen when cancelling explicitly
> because luo_cancel() calls kho_abort().

Yes, KHO still has its states, that break the LUO logic. I think,
there is going to be some limitations until "stateless" kho patches
land.

> I think you should just make the release go through the cancel flow,
> since the operation is essentially a cancel anyway. There are subtle
> differences here though, since the release might be called before
> prepare, so we need to be careful of that.

Makes sense.

Thank you,
Pasha