Re: [PATCH] cred: Use RCU primitives to access RCU pointers

[Jann Horn]

On Thu, Feb 6, 2020 at 5:49 PM Joel Fernandes <joel@xxxxxxxxxxxxxxxxx> wrote:
> On Thu, Feb 06, 2020 at 12:28:42PM +0100, Jann Horn wrote:
> [snip]
> > > > > > > > task_struct.cred doesn't actually have RCU semantics though, see
> > > > > > > > commit d7852fbd0f0423937fa287a598bfde188bb68c22. For task_struct.cred,
> > > > > > > > it would probably be more correct to remove the __rcu annotation?
> > > > > > >
> > > > > > > Hi Jann,
> > > > > > >
> > > > > > > I went through the commit you mentioned. If I understand it correctly,
> > > > > > > ->cred was not being accessed concurrently (via RCU), hence, a non_rcu
> > > > > > > flag was introduced, which determined if the clean-up should wait for
> > > > > > > RCU grace-periods or not. And since, the changes were 'thread local'
> > > > > > > there was no need to wait for an entire RCU GP to elapse.
> > > > > >
> > > > > > Yeah.
> > > > > >
> > > > > > > The commit too, as you said, mentions the removal of __rcu annotation.
> > > > > > > However, simply removing the annotation won't work, as there are quite a
> > > > > > > few instances where RCU primitives are used. Even get_current_cred()
> > > > > > > uses RCU APIs to get a reference to ->cred.
> > > > > >
> > > > > > Luckily, there aren't too many places that directly access ->cred,
> > > > > > since luckily there are helper functions like get_current_cred() that
> > > > > > will do it for you. Grepping through the kernel, I see:
> > > > [...]
> > > > > > So actually, the number of places that already don't use RCU accessors
> > > > > > is much higher than the number of places that use them.
> > > > > >
> > > > > > > So, currently, maybe we
> > > > > > > should continue to use RCU APIs and leave the __rcu annotation in?
> > > > > > > (Until someone who takes it on himself to remove __rcu annotation and
> > > > > > > fix all the instances). Does that sound good? Or do you want me to
> > > > > > > remove __rcu annotation and get the process started?
> > > > > >
> > > > > > I don't think it's a good idea to add more uses of RCU APIs for
> > > > > > ->cred; you shouldn't "fix" warnings by making the code more wrong.
> > > > > >
> > > > > > If you want to fix this, I think it would be relatively easy to fix
> > > > > > this properly - as far as I can tell, there are only seven places that
> > > > > > you'll have to change, although you may have to split it up into three
> > > > > > patches.
> > > > >
> > > > > Thank you for the detailed analysis. I'll try my best and send you a
> > > > > patch.
> > >
> > > Amol, Jann, if I understand the discussion correctly, objects ->cred
> > > point (the subjective creds) are never (or never need to be) RCU-managed.
> > > This makes sense in light of the commit Jann pointed out
> > > (d7852fbd0f0423937fa287a598bfde188bb68c22).
[...]
> > > 3. Also I removed the whole non_rcu flag, and introduced a new put_cred_non_rcu() API
> > >    which places that task-synchronously use ->cred can overwrite. Callers
> > >    doing such accesses like access() can use this API instead.
> >
> > That's wrong, don't do that.
> >
> > ->cred is a reference without RCU semantics, ->real_cred is a
> > reference with RCU semantics. If there have never been any references
> > with RCU semantics to a specific instance of struct cred, then that
> > instance can indeed be freed without an RCU grace period. But it would
> > be possible for some filesystem code to take a reference to
> > current->cred, and assign it to some pointer with RCU semantics
> > somewhere, then drop that reference with put_cred() immediately before
> > you reach put_cred_non_rcu(); with the result that despite using
> > put_cred(), the other side doesn't get RCU semantics.
> >
> > Just leave the whole ->non_rcu thing exactly as it was.
>
> Can you point to an example in the kernel that actually uses ->cred this way?
> I'm just curious. That is, reads task's ->cred pointer, and assigns it to an
> RCU managed pointer?

I'm almost sure that there are no such cases at the moment. However,
from a maintainability standpoint, I'm still very twitchy about this
change; the current API encapsulates the RCU weirdness in the standard
helper functions, but with your proposal, suddenly taking f_cred from
somewhere and using it as a new task's subjective creds, or something
like that, would be unsafe.

> I think such an example would be the point that the commit you mentioned
> addresses. The commit basically says "as long as nobody does get_cred() on
> the task_struct ->cred, we are good, but if somebody does do it, then we have
> to deferred-free it".  But I could not find such an example.
>
> That said, I agree the removal of non_rcu can be considered out of scope for
> this patch.