Re: [PATCH next] drm/xe: Fix uninitialized variable in xe_vm_bind_ioctl()

[Matthew Brost <matthew.brost@intel.com>]

On Mon, Mar 10, 2025 at 10:12:15PM -0700, Matthew Brost wrote:
> On Mon, Mar 10, 2025 at 10:04:22PM -0700, Matthew Brost wrote:
> > On Mon, Mar 10, 2025 at 09:22:50PM +0300, Dan Carpenter wrote:
> > > On Mon, Mar 10, 2025 at 12:56:46PM -0400, Rodrigo Vivi wrote:
> > > > On Mon, Mar 10, 2025 at 01:48:00PM +0300, Dan Carpenter wrote:
> > > > > The error handling assumes that vm_bind_ioctl_check_args() will
> > > > > initialize "bind_ops" but there are a couple early returns where that's
> > > > > not true.  Initialize "bind_ops" to NULL from the start.
> > > > 
> > > > It is not a couple, but only the one goto put_vm where this bind_ops
> > > > gets actually initialized, or not...
> > > > 
> > > 
> > > I'm on linux-next.  I'm not seeing the goto put_vm...  I think we're
> > > looking at different code.
> > > 
> > >   3063  static int vm_bind_ioctl_check_args(struct xe_device *xe, struct xe_vm *vm,
> > >   3064                                      struct drm_xe_vm_bind *args,
> > >   3065                                      struct drm_xe_vm_bind_op **bind_ops)
> > >   3066  {
> > >   3067          int err;
> > >   3068          int i;
> > >   3069  
> > >   3070          if (XE_IOCTL_DBG(xe, args->pad || args->pad2) ||
> > >   3071              XE_IOCTL_DBG(xe, args->reserved[0] || args->reserved[1]))
> > >   3072                  return -EINVAL;
> > > 
> > > One.
> > > 
> > >   3073  
> > >   3074          if (XE_IOCTL_DBG(xe, args->extensions))
> > >   3075                  return -EINVAL;
> > > 
> > > Two.
> > > 
> > >   3076  
> > >   3077          if (args->num_binds > 1) {
> > >   3078                  u64 __user *bind_user =
> > >   3079                          u64_to_user_ptr(args->vector_of_binds);
> > >   3080  
> > >   3081                  *bind_ops = kvmalloc_array(args->num_binds,
> > > 
> > > Initialized.
> > > 
> > >   3082                                             sizeof(struct drm_xe_vm_bind_op),
> > >   3083                                             GFP_KERNEL | __GFP_ACCOUNT |
> > >   3084                                             __GFP_RETRY_MAYFAIL | __GFP_NOWARN);
> > >   3085                  if (!*bind_ops)
> > >   3086                          return args->num_binds > 1 ? -ENOBUFS : -ENOMEM;
> > >   3087  
> > >   3088                  err = __copy_from_user(*bind_ops, bind_user,
> > >   3089                                         sizeof(struct drm_xe_vm_bind_op) *
> > >   3090                                         args->num_binds);
> > >   3091                  if (XE_IOCTL_DBG(xe, err)) {
> > >   3092                          err = -EFAULT;
> > >   3093                          goto free_bind_ops;
> > >   3094                  }
> > >   3095          } else {
> > >   3096                  *bind_ops = &args->bind;
> > >   3097          }
> > > 
> > > > but perhaps the order in the exit is wrong and we should move the
> > > > kvfree(bind_ops) upper to the end of put_exec_queue?
> > > > 
> > > > Matt, thoughts on the order here?
> > > > 
> > 
> > Rodrigo – I think you are looking in the wrong spot in the code. Dan's
> > subsequent reply, I believe, explains the issue correctly, so I think
> > the patch is good.
> > 
> > > > Cc: Matthew Brost <matthew.brost@intel.com>
> > > 
> > > I feel like ideally vm_bind_ioctl_check_args() would clean up after
> > > itself on failure and, yes, it should be in reverse order from how
> > > it was allocated.
> > > 
> > 
> > This is a bit of goofy error handling/convention—perhaps it could be
> > cleaned up in a follow-up.
> > 
> > That said, this patch is correct. However, the 'Fixes' tag looks
> > wrong—it has been broken for a bit longer than the tagged patch. We can
> > fix it upon merge.
> > 

Cough as I eat my hat - the fixes tag in correct - the patch tagged
moved the args check after the VM lookup which created a bug. 

> > With that:
> > Reviewed-by: Matthew Brost <matthew.brost@intel.com>
> > 
> 
> Actually, we have another problem too. The 'free_bind_ops' label in
> vm_bind_ioctl_check_args() either isn't needed or it should *bind_ops to
> NULL after kvfree to avoid a double free in put_vm label in
> xe_vm_bind_ioctl().
> 
> This patch is still valid though.
> 

Posted a follow up include Dan's original change and also my suggested
change above.

Matt

> Matt
> 
> > > regards,
> > > dan carpenter
> > >