Re: [PATCH] block: switch to atomic_t for request references

[Kees Cook]

On Mon, Dec 06, 2021 at 04:13:00PM -0800, Linus Torvalds wrote:
> On Mon, Dec 6, 2021 at 3:28 PM Kees Cook <keescook@xxxxxxxxxxxx> wrote:
> >
> > I'm not arguing for refcount_t -- I'm arguing for an API that isn't a
> > regression of features that have been protecting the kernel from bugs.
> 
> Maybe somebody could actually just fix refcount_t instead. Somebody
> who cares about that currently horrendously bad interface.
> 
> Fix it to not do the fundamentally broken saturation that actively
> destroys state: fix it to have a safe "try to increment", instead of
> an unsafe "increment and do bad things".

There would need to be a pretty hefty transition -- there are a lot of
refcount_inc() uses that would need checking and error handling (which
might not be sane to add to ancient drivers):

      2 block
      2 crypto
      2 ipc
      2 virt
      3 mm
      4 sound
      5 rust
     10 arch
     13 security
     31 kernel
     88 include
    192 fs
    192 net
    358 drivers

refcount_inc_not_zero() already uses __must_check, etc.

I'm not afraid of giant transitions, but this could be pretty tricky.
I'm open to ideas. Maybe a treewide change of refcount_inc() ->
refcount_inc_saturating() and then start fixing all the _unsafe() cases
where a sensible error path could be created and tested?

-- 
Kees Cook