On Fri, Jul 17, 2026 at 10:07 AM Pasha Tatashin <[email protected]> wrote: > > On 07-16 17:41, Pratyush Yadav wrote: > > On Thu, Jul 16 2026, Pasha Tatashin wrote: > > > > > On 07-15 14:23, Jason Gunthorpe wrote: > > >> On Wed, Jul 15, 2026 at 03:50:33PM +0200, Pratyush Yadav wrote: > > >> > Hi David, > > >> > > > >> > On Tue, Jul 14 2026, David Matlack wrote: > > >> > > > >> > > Remove the single-opener restriction for /dev/liveupdate by removing > > >> > > the > > >> > > atomic in_use tracking and the exclusive open check in luo_open() > > >> > > that > > >> > > returned -EBUSY. Protect luo_session_deserialize() with a mutex > > >> > > guard so > > >> > > that concurrent open attempts by multiple processes safely executes > > >> > > deserialization only once. Update liveupdate selftest to verify that > > >> > > multiple concurrent openers succeed. > > >> > > > > >> > > LUO does not inherently require a single opener. There is some > > >> > > documentation about it simplifying state management, but the only > > >> > > thing > > >> > > it actually protects is the session deserialization during first > > >> > > open, > > >> > > which can be easily handled with a mutex. > > >> > > > > >> > > Relaxing the single-opener requirement avoids the kernel forcing a > > >> > > design pattern on userspace that it itself does not require, e.g. > > >> > > allowing multiple userspace processes to create and manage sessions. > > >> > > > >> > Agreed. When the kernel had a global state machine in the early > > >> > versions > > >> > of LUO, this might have been more relevant. With sessions, even if we > > >> > later add a state machine, it likely will be per-session instead of > > >> > being global. So I think letting userspace open /dev/liveupdate > > >> > multiple > > >> > times makes a lot of sense. > > >> > > > >> > Also, today's systemd only supports preserving individual files, and > > >> > does not hand out sessions. To get sessions, userspace must open > > > > > > It should in the future, because of permissions issue, see below. > > > > > >> > /dev/liveupdate and create a session. This opens up room for one bad > > >> > process to block every other process from creating sessions. It also > > >> > imposes a need for userspace to add a polling/retry logic for getting > > >> > sessions and serializes their execution around this point. > > >> > > >> Shouldn't systemd open and own /dev/liveupdate? That was at least what > > >> I originally expected here, you'd talk to it and get a session FD > > >> through dbus. > > >> > > >> Moving to multi-opening /dev/liveupdate and removing visibility of > > >> what sessions are open from systemd is a different model > > >> > > >> Not saying this patch is wrong or anything, but that I don't really > > >> understand what kind of model you are going for now. > > > > > > CC ca for systemd's take. > > > > > > /dev/liveupdate should only be accessed by a privileged process, and > > > sessions should be accessed by whoever originally created them. While > > > this patch does not change the permissions, it paves the road for us to > > > move in the wrong direction: instead of having a privileged userspace > > > manager that distributes the sessions to their rightful owners, it > > > encourages userspace to work around the permissions so that VMMs access > > > /dev/liveupdate to retrieve or store their sessions directly. This would > > > also allow them to access sessions belonging to any other participant of > > > the live update. > > > > But you can still do all this. In fact, systemd's release notes suggest > > doing so [0]: > > > > Units can also create their own LUO Sessions by talking to the > > kernel directly, and store them in their FD Stores, and those will > > also be preserved and passed down to the unit after kexec > > They should not. Ideally, systemd should be the sole entity > communicating directly with /dev/liveupdate. > > > > > Having a single opener does not in practice prevent userspace from > > creating sessions directly. All it does is to force them to turn the > > open into a polling loop. So I don't think single open achieves the goal > > That is a fragile design, and they should avoid it. To be clear, I do > not see any functional bugs in this patch, but I still do not understand > the underlying use case. > > LUO should rely on a single manager agent that interacts with clients to > create, preserve across reboots, and restore session IDs on their behalf. > > In fact, systemd is the only userspace process that AFAIK survives > across the reboot() boundary; all other processes are terminated before > reboot() is invoked. Relying on multiple independent LUO managers is > therefore architecturally, wrong as they will close sessionfds. > > While the kernel can technically allow multiple openers for > /dev/liveupdate, I do not see what problem this actually solves. > Instead, it seems to encourage bad architectural decisions in userspace.
> > you think it should. Userspace already works around this restriction. > > > > So as long as we keep access to /dev/liveupdate restricted to privileged > > processes, I don't see why single open is any better. > > What is the concrete use case for this patch, once we live in the world > where systemd is fully capable of creating, storing, and retrieving > sessions? I guess I would flip the question around: What problem does the kernel requiring a single opener of /dev/liveupdate solve? If systemd wants to have a single opener model it can certainly enforce that itself without the kernel's help. I'm not sure why the kernel needs to impose a restriction on userspace that it doesn't require itself. > > > > > [0] https://github.com/systemd/systemd/releases#release-v261 > > > > > > > > Instead, sessions should be created and retrieved by a privileged > > > process that knows to send them back to their rightful owner after > > > retrieval. > > > > > > Also, as a minor concern, each userspace LUO manager needs its own > > > session to maintain state. This means that by allowing multiple > > > managers, they may run into naming conflicts for those state sessions. > > > > -- > > Regards, > > Pratyush Yadav
