Re: [GIT PULL] ucounts: Count rlimits in each user namespace

[Alexey Gladkov <legion@kernel.org>]

On Tue, Jun 29, 2021 at 11:07:11AM -0700, Linus Torvalds wrote:
> On Tue, Jun 29, 2021 at 10:18 AM Alexey Gladkov <legion@kernel.org> wrote:
> >
> >
> > > And why test for "ucounts" being non-NULL in
> > >
> > >                 if (ucounts && dec_rlimit_ucounts(ucounts,
> > > UCOUNT_RLIMIT_SIGPENDING, 1))
> > >                         put_ucounts(ucounts);
> > >
> > > when afaik both of those should be happy with a NULL 'ucounts' pointer
> > > (if it was NULL, we certainly already used it for the reverse
> > > operations for get_ucounts() and inc_rlimit_ucounts()..)
> >
> > The get_ucount() can theoretically return NULL. It increments the
> > reference counter and if it overflows, the function will return NULL.
> 
> .. but my point is that dec_rlimit_ucounts() and put_ucounts() should
> be fine with whatever get_ucounts() returned. No
> 
> It looks like put_ucounts() is unhappy with a NULL ucounts argument,
> but I think _that_ is what should get fixed.
> 
> I think that conceptually we should have two clear alternatives:
> 
>  (a) either "get_ucounts()" returning NULL should be an error, and we
> would have returned long before

get_ucounts() in the __sigqueue_alloc() performs the get_uid() function
but does not ignore the counter overflow. Basically get_uid() can fail in
same way as get_ucounts(), but we just ignore it.

> or
> 
>  (b) a NULL uncounts is usable, and a sequence like
> put_ucounts(get_ucounts()) should just always work.
> 
> And honestly, a lot of the other ucounts funcrtions seem to take that
> (b) approach. Example in that very function:
> 
>         ucounts = task_ucounts(t);
>         sigpending = inc_rlimit_ucounts(ucounts, UCOUNT_RLIMIT_SIGPENDING, 1);
> 
> which at no point tested for NULL or returned an error.

Waaaait. task_ucounts() is a different thing. This function only gets a
field from the task structure without any reference counting. But the
get_ucounts() is more like get_user_ns() or get_uid(), but does not ignore
counter overflow.

Earlier I tried to use refcount_t which never returns errors [1]. We
talked and you said that ignoring counter overflow errors is bad
design for this case.

> (And that also implies that the comment in dec_rlimit_ucounts() about
> "Silence compiler warning" should just go away, because it's not just
> a compiler warning, it's a required initialization).
> 
>               Linus

[1] https://lore.kernel.org/lkml/CAHk-%3dwjYOCgM%2bmKzwTZwkDDg12DdYjFFkmoFKYLim7NFmR9HBg@mail.gmail.com/

-- 
Rgrds, legion