Re: [PATCH gmem FIXUP] kvm: guestmem: do not use a file system

[Sean Christopherson]

On Tue, Oct 10, 2023, Al Viro wrote:
> On Mon, Oct 09, 2023 at 02:26:36PM -0700, Sean Christopherson wrote:
> > On Mon, Oct 09, 2023, Al Viro wrote:
> > > On Mon, Oct 09, 2023 at 01:20:06PM -0700, Sean Christopherson wrote:
> > > > On Mon, Oct 09, 2023, Al Viro wrote:
> > > > > On Mon, Oct 09, 2023 at 07:32:48AM -0700, Sean Christopherson wrote:
> > > > > 
> > > > > > Yeah, we found that out the hard way.  Is using the "secure" variant to get a
> > > > > > per-file inode a sane approach, or is that abuse that's going to bite us too?
> > > > > > 
> > > > > > 	/*
> > > > > > 	 * Use the so called "secure" variant, which creates a unique inode
> > > > > > 	 * instead of reusing a single inode.  Each guest_memfd instance needs
> > > > > > 	 * its own inode to track the size, flags, etc.
> > > > > > 	 */
> > > > > > 	file = anon_inode_getfile_secure(anon_name, &kvm_gmem_fops, gmem,
> > > > > > 					 O_RDWR, NULL);
> > > > > 
> > > > > Umm...  Is there any chance that your call site will ever be in a module?
> > > > > If not, you are probably OK with that variant.
> > > > 
> > > > Yes, this code can be compiled as a module.  I assume there issues with the inode
> > > > outliving the module?
> > > 
> > > The entire file, actually...  If you are using that mechanism in a module, you
> > > need to initialize kvm_gmem_fops.owner to THIS_MODULE; AFAICS, you don't have
> > > that done.
> > 
> > Ah, that's handled indirectly handled by a chain of refcounted objects.  Every
> > VM that KVM creates gets a reference to the module, and each guest_memfd instance
> > gets a reference to its owning VM.
> 
> Umm... what's the usual call chain leading to final drop of refcount of that
> module?

If the last reference is effectively held by guest_memfd, it would be:

  kvm_gmem_release(), a.k.a. file_operations.release()
  |
  -> kvm_put_kvm()
     |
     -> kvm_destroy_vm()
        |
        -> module_put(kvm_chardev_ops.owner);