Re: [PATCH v9 12/27] rust: add `kernel` crate

[Wedson Almeida Filho]

On Mon, Sep 19, 2022 at 01:42:44PM -0700, Linus Torvalds wrote:
> On Mon, Sep 19, 2022 at 11:05 AM Wedson Almeida Filho
> <wedsonaf@xxxxxxxxx> wrote:
> >
> > As you know, we're trying to guarantee the absence of undefined
> > behaviour for code written in Rust. And the context is _really_
> > important, so important that leaving it up to comments isn't enough.
> 
> You need to realize that
> 
>  (a) reality trumps fantasy
> 
>  (b) kernel needs trump any Rust needs
> 
> And the *reality* is that there are no absolute guarantees.  Ever. The
> "Rust is safe" is not some kind of absolute guarantee of code safety.
> Never has been. Anybody who believes that should probably re-take
> their kindergarten year, and stop believing in the Easter bunny and
> Santa Claus.
> 
> Even "safe" rust code in user space will do things like panic when
> things go wrong (overflows, allocation failures, etc). If you don't
> realize that that is NOT some kind of true safely, I don't know what
> to say.

No one is talking about absolute safety guarantees. I am talking about
specific ones that Rust makes: these are well-documented and formally
defined.

> Not completing the operation at all, is *not* really any better than
> getting the wrong answer, it's only more debuggable.
>
> In the kernel, "panic and stop" is not an option (it's actively worse
> than even the wrong answer, since it's really not debugable), so the
> kernel version of "panic" is "WARN_ON_ONCE()" and continue with the
> wrong answer.
> 
> So this is something that I really *need* the Rust people to
> understand. That whole reality of "safe" not being some absolute
> thing, and the reality that the kernel side *requires* slightly
> different rules than user space traditionally does.
> 
> > I don't care as much about allocation flags as I do about sleeping in an
> > rcu read-side critical region. When CONFIG_PREEMPT=n, if some CPU makes
> > the mistake of sleeping between rcu_read_lock()/rcu_read_unlock(), RCU
> > will take that as a quiescent state, which may cause unsuspecting code
> > waiting for a grace period to wake up too early and potentially free
> > memory that is still in use, which is obviously undefined behaviour.
> 
> So?
> 
> You had a bug. Shit happens. We have a lot of debugging tools that
> will give you a *HUGE* warning when said shit happens, including
> sending automated reports to the distro maker. And then you fix the
> bug.
> 
> Think of that "debugging tools give a huge warning" as being the
> equivalent of std::panic in standard rust. Yes, the kernel will
> continue (unless you have panic-on-warn set), because the kernel
> *MUST* continue in order for that "report to upstream" to have a
> chance of happening.
> 
> So it's technically a veryu different implementation from std:panic,
> but you should basically see it as exactly that: a *technical*
> difference, not a conceptual one. The rules for how the kernel deals
> with bugs is just different, because we don't have core-files and
> debuggers in the general case.
> 
> (And yes, you can have a kernel debugger, and you can just have the
> WARN_ON_ONCE trigger the debugger, but think of all those billions of
> devices that are in normal users hands).
> 
> And yes, in certain configurations, even those warnings will be turned
> off because the state tracking isn't done. Again, that's just reality.
> You don't need to use those configurations yourself if you don't like
> them, but that does *NOT* mean that you get to say "nobody else gets
> to use those configurations either".
> 
> Deal with it.

While I disagree with some of what you write, the point is taken.

But I won't give up on Rust guarantees just yet, I'll try to find
ergonomic ways to enforce them at compile time.

Thanks,
-Wedson