Re: [RFC v1 0/3] Address potential user-after-free on module unload

[Sven Van Asbroeck]

On Wed, Feb 6, 2019 at 12:30 PM Dmitry Torokhov
<dmitry.torokhov@xxxxxxxxx> wrote:
>
> Yeah. But devm irq gave most trouble because we did not have enough
> devm APIs so we often ended up with mixed devm/non-devm usage and that
> is what was causing most of the issues. If we can switch everything to
> devm then devm irq is not that troublesome.
>

It sounds to me like _incomplete_ devm_ is worse than no devm at all.

Imagine a devm_ resource depends on a non-devm one:

int acme_probe(struct device *dev)
{
  ...
  r = create_something();
  d = devm_create_thing(dev, r);
}

Then remove could get us into some serious trouble:

void acme_remove(struct device *dev)
{
  /* r _must_ be released here, we have no other place to do it */
  destroy_something(r);
  /* here, d is still alive because it's devm
   * which is cleaned up _after_ remove().
  * Now we have a live resource using a released resource.
  * use-after-free anyone?
  */
}

This is a more generalized version of the issue I originally
observed, where r => struct work_struct.

I'm sure there must be plenty of these around the codebase.

I wish we had a Coccinelle script to catch these, because it's
one thing to fix them today. More will be added tomorrow.
devm_ is so elegant that people frequently use it without
thinking it through.

I certainly would have, before yesterday :)