Re: [PATCH v2 4/6] iommufd: Deliver fault messages to user space

[Jason Gunthorpe]

On Fri, Dec 08, 2023 at 01:47:35PM +0800, Baolu Lu wrote:
> On 12/8/23 1:17 AM, Jason Gunthorpe wrote:
> > On Thu, Dec 07, 2023 at 05:34:10PM +0100, Joel Granados wrote:
> > > > @@ -58,6 +255,8 @@ static void hw_pagetable_fault_free(struct hw_pgtable_fault *fault)
> > > >   	WARN_ON(!list_empty(&fault->deliver));
> > > >   	WARN_ON(!list_empty(&fault->response));
> > > > +	fput(fault->fault_file);
> > > > +	put_unused_fd(fault->fault_fd);
> > > I have resolved this in a naive way by just not calling the
> > > put_unused_fd function.
> > That is correct.
> > 
> > put_unused_fd() should only be called on error paths prior to the
> > syscall return.
> > 
> > The design of a FD must follow this pattern
> > 
> >   syscall():
> >     fdno = get_unused_fd_flags(O_CLOEXEC);
> >     filep = [..]
> >     // syscall MUST succeed after this statement:
> >     fd_install(fdno, filep);
> >     return 0;
> > 
> >    err:
> >      put_unused_fd(fdno)
> >      return -ERRNO
> 
> Yes. Agreed.
> 
> > 
> > Also the refcounting looks a little strange, the filep reference is
> > consumed by fd_install, so what is that fput pairing with in fault_free?
> 
> fput() pairs with get_unused_fd_flags()? fd_install() does not seem to
> increase any reference.

fd_install() transfers the reference to the fd table and that
reference is put back by the close() system call.

Notice that instantly after fd_install() a concurrent user can free
the filep.

Jason