On Wed, Feb 25, 2026 at 3:55 PM Samiullah Khawaja <[email protected]> wrote: > > On Wed, Feb 25, 2026 at 3:15 PM David Matlack <[email protected]> wrote: > > > > On 2026-02-25 02:33 PM, Alex Williamson wrote: > > > On Thu, 29 Jan 2026 21:24:51 +0000 > > > David Matlack <[email protected]> wrote: > > > > > > +int __init vfio_pci_liveupdate_init(void) > > > > +{ > > > > + if (!liveupdate_enabled()) > > > > + return 0; > > > > + > > > > + return liveupdate_register_file_handler(&vfio_pci_liveupdate_fh); > > > > +} > > > > > > liveupdate_register_file_handler() "pins" vfio-pci with a > > > try_module_get(). Since this is done in our module_init function and > > > unregister occurs in our module_exit function, rather than relative > > > to any actual device binding or usage, this means vfio-pci CANNOT be > > > unloaded. That seems bad. Thanks, > > > > Good point. So a better approach that would allow vfio-pci to be > > unloaded would be to register the file handler when the number of > > devices bound to vfio-pci goes from 0->1 and then unregister on 1->0. > > Yeah maybe a kref that gets inc/dec in probe/remove. I have a similar > problem with iommufd preservation, but I think I can handle it based > on the number of iommufd open. > > I am wondering whether this file handler registration kref stuff can > be moved into LUO by adding it to LUO file_handler and the modules > only call get/put?
Circling back here, the new plan is for LUO to not take a module reference in liveupdate_register_file_handler(): https://lore.kernel.org/lkml/[email protected]/ In the next version of this series I am going to rebase on top of Pasha's changes and keep the call to liveupdate_register_file_handler() where it is.
