CVE-2026-97478

In the Linux kernel, the following vulnerability has been resolved:

virt: acrn: Fix irqfd use-after-free during eventfd shutdown

acrn_irqfd_deassign() and the eventfd EPOLLHUP wakeup can race and free
the same struct hsm_irqfd:

  CPU0                                 CPU1
  ----                                 ----
  eventfd_release()
    wake_up_poll(EPOLLHUP)
      hsm_irqfd_wakeup()
        queue_work(&irqfd->shutdown)
                                       acrn_irqfd_deassign()
                                         hsm_irqfd_shutdown()
                                           list_del_init()
                                           eventfd_ctx_remove_wait_queue()
                                           eventfd_ctx_put()
                                         kfree(irqfd)
  hsm_irqfd_shutdown_work()
    container_of(work, ..., shutdown)
    irqfd->vm                  <-- use-after-free

The deassign path freed the irqfd while a shutdown work item was
already queued by EPOLLHUP (or vice versa), so the work item could
resurrect a dangling pointer through container_of().

Switch to the lifetime model used by KVM irqfds:

 - Deassign/deinit only deactivate the irqfd: remove it from vm->irqfds
   under irqfds_lock and queue the cleanup work.
 - hsm_irqfd_shutdown_work() becomes the sole owner that unhooks the
   eventfd waitqueue entry, drops the eventfd reference and frees the
   irqfd.
 - A new HSM_IRQFD_FLAG_SHUTDOWN bit guarded by test_and_set_bit()
   ensures the cleanup work is queued at most once, no matter how many
   of {EPOLLHUP, deassign, deinit} fire concurrently.  This is safe to
   call from the waitqueue callback, which runs with wqh->lock held and
   IRQs disabled and therefore cannot take irqfds_lock.
 - acrn_irqfd_deassign() flushes vm->irqfd_wq before returning so the
   eventfd is fully detached on return.  acrn_irqfd_deinit() deactivates
   every irqfd, flushes the workqueue and only then destroys it, so no
   path can queue_work() onto a torn-down workqueue.
 - acrn_irqfd_assign() now installs the eventfd waitqueue entry and
   publishes the irqfd to vm->irqfds under irqfds_lock, so the irqfd is
   never visible to deassign/deinit before its waitqueue entry is in
   place, and any EPOLLHUP that fires in the assign window queues
   cleanup work that blocks on irqfds_lock until publication is done.