Re: [PATCH 5/5] platform/x86: Add Intel Software Defined Silicon driver

[Greg KH]

On Fri, Oct 01, 2021 at 04:13:58AM -0700, David E. Box wrote:
> On Fri, 2021-10-01 at 09:29 +0200, Greg KH wrote:
> > On Thu, Sep 30, 2021 at 06:28:15PM -0700, David E. Box wrote:
> > > +static long sdsi_device_ioctl(struct file *file, unsigned int cmd, unsigned long arg)
> > > +{
> > > +       struct miscdevice *miscdev = file->private_data;
> > > +       struct sdsi_priv *priv = to_sdsi_priv(miscdev);
> > > +       void __user *argp = (void __user *)arg;
> > > +       long ret = -EINVAL;
> > > +
> > > +       if (!priv->dev_present)
> > > +               return -ENODEV;
> > > +
> > > +       if (!priv->sdsi_enabled)
> > > +               return -EPERM;
> > > +
> > > +       if (cmd == SDSI_IF_READ_STATE)
> > > +               return sdsi_if_read_state_cert(priv, argp);
> > > +
> > > +       mutex_lock(&priv->akc_lock);
> > > +       switch (cmd) {
> > > +       case SDSI_IF_PROVISION_AKC:
> > > +               /*
> > > +                * While writing an authentication certificate disallow other openers
> > > +                * from using AKC or CAP.
> > > +                */
> > > +               if (!priv->akc_owner)
> > > +                       priv->akc_owner = file;
> > > +
> > > +               if (priv->akc_owner != file) {
> > 
> > Please explain how this test would ever trigger and how you tested it?
> > 
> > What exactly are you trying to protect from here?  If userspace has your
> > file descriptor, it can do whatever it wants, don't try to be smarter
> > than it as you will never win.
> > 
> > And why are you using ioctls at all here?  As you are just
> > reading/writing to the hardware directly, why not just use a binary
> > sysfs file to be that pipe?  What requires an ioctl at all?
> 
> So an original internal version of this did use binary attributes. But there was concern during
> review that a flow, particularly when doing the two write operations, could not be handled
> atomically while exposed as separate files. Above is the attempt to handle the situation in the
> ioctl. That is, whichever opener performs AKC write first would lock out all other openers from
> performing any write until that file is closed. This is to avoid interfering with that process,
> should the opener also decide to perform a CAP operation.

Unfortunately, your code here does not prevent that at all, so your
moving off of a binary sysfs attribute changed nothing.

You can "prevent" this from happening just as easily through a sysfs
attribute as you can a character device node.

> There may be future commands requiring RW ioctls as well.

How am I or anyone else supposed to know that?  We write code and review
it for _today_, not what might be sometime in the future someday.  As
that will be dealt with when it actually happens.

greg k-h