Re: [PATCH] secretmem: Prevent secretmem_users from wrapping to zero

[Linus Torvalds <torvalds@linux-foundation.org>]

On Mon, Oct 25, 2021 at 4:37 PM Linus Torvalds
<torvalds@linux-foundation.org> wrote:
>
> Stop thinking that refcount_t is a good type. Start realizing the
> downsides. Start understanding that saturation is a HORRENDOUSLY BAD
> solution, and horrible QoI.

Basically, refcount_t should be used purely for internal kernel data
structures - it makes perfect sense for things like the 'struct
device' resource handling, for example.

In fact, there it is good for two reasons:

 - it's not counting some user resource, so users shouldn't have any
way to trigger overflow and saturation which causes problems

 - it's used by random driver stuff, which is often where kernel bugs
happen and testing is fundamentally limited by hw availability etc

but in general, anything that is user-accountable needs to have
_limits_, not saturation. It's why the page count is a "atomic_t" even
if the name of the field is "_refcount". Because refcount_t is the
INFERIOR TYPE.

Using an atomic_t properly is actually the much better option. It's
just that "properly" might be a bit more code, involving actual limit
checking.

'refcount_t' is basically a shorthand for "I didn't bother doing this
right, so I'm using this type that adds debugging, warns and stops
working and might DoS the kernel".

It's a crutch. It's not the alpha and the omega of counting types. It
has its place, but I really want to stress how people should
ABSOLUTELY not think "oh, refcount_t is better than atomic_t".

              Linus