Re: [PATCH net-next v2 5/6] rust: Add read_poll_timeout function

[Alice Ryhl]

On Mon, Oct 7, 2024 at 4:14 PM Boqun Feng <boqun.feng@xxxxxxxxx> wrote:
>
> On Mon, Oct 07, 2024 at 04:08:48PM +0200, Alice Ryhl wrote:
> > On Mon, Oct 7, 2024 at 3:48 PM Andrew Lunn <andrew@xxxxxxx> wrote:
> > >
> > > On Mon, Oct 07, 2024 at 05:28:28AM -0700, Boqun Feng wrote:
> > > > On Sun, Oct 06, 2024 at 04:45:21PM +0200, Andrew Lunn wrote:
> > > > However, this is actually a special case: currently we want to use klint
> > > > [1] to detect all context mis-matches at compile time. So the above rule
> > > > extends for kernel: any type-checked *and klint-checked* code that only
> > > > calls safe Rust functions cannot be unsafe. I.e. we add additional
> > > > compile time checking for unsafe code. So if might_sleep() has the
> > > > proper klint annotation, and we actually enable klint for kernel code,
> > > > then we can make it safe (along with preemption disable functions being
> > > > safe).
> > > >
> > > > > where you use a sleeping function in atomic context. Depending on why
> > > > > you are in atomic context, it might appear to work, until it does not
> > > > > actually work, and bad things happen. So it is not might_sleep() which
> > > > > is unsafe, it is the Rust code calling it.
> > > >
> > > > The whole point of unsafe functions is that calling it may result into
> > > > unsafe code, so that's why all extern "C" functions are unsafe, so are
> > > > might_sleep() (without klint in the picture).
> > >
> > > There is a psychological part to this. might_sleep() is a good debug
> > > tool, which costs very little in normal builds, but finds logic bugs
> > > when enabled in debug builds. What we don't want is Rust developers
> > > not scattering it though their code because it adds unsafe code, and
> > > the aim is not to have any unsafe code.
> >
> > We can add a safe wrapper for it:
> >
> > pub fn might_sleep() {
> >     // SAFETY: Always safe to call.
> >     unsafe { bindings::might_sleep() };
>
> It's not always safe to call, because might_sleep() has a
> might_resched() and in preempt=voluntary kernel, that's a
> cond_resched(), which may eventually call __schedule() and report a
> quiescent state of RCU. This could means an unexpected early grace
> period, and that means a potential use-afer-free.

Atomicity violations are intended to be caught by klint. If you want
to change that decision, you'll have to add unsafe to all functions
that sleep including Mutex::lock, CondVar::wait, and many others.

Alice