Re: [PATCH v2 2/6] bitops: always define asm-generic non-atomic bitops

[Marco Elver <elver@xxxxxxxxxx>]

On Fri, 10 Jun 2022 at 18:02, Luck, Tony <tony.luck@xxxxxxxxx> wrote:
>
> > > +/**
> > > + * generic_test_bit - Determine whether a bit is set
> > > + * @nr: bit number to test
> > > + * @addr: Address to start counting from
> > > + */
> >
> > Shouldn't we add in this or in separate patch a big NOTE to explain that this
> > is actually atomic and must be kept as a such?
>
> "atomic" isn't really the right word. The volatile access makes sure that the
> compiler does the test at the point that the source code asked, and doesn't
> move it before/after other operations.

It's listed in Documentation/atomic_bitops.txt.

It is as "atomic" as READ_ONCE() or atomic_read() is. Though you are
right that the "atomicity" of reading one bit is almost a given,
because we can't really read half a bit.
The main thing is that the compiler keeps it "atomic" and e.g. doesn't
fuse the load with another or elide it completely, and then transforms
the code in concurrency-unfriendly ways.

Like READ_ONCE() and friends, test_bit(), unlike non-atomic bitops,
may also be used to dependency-order some subsequent marked (viz.
atomic) operations.

> But there is no such thing as an atomic test_bit() operation:
>
>         if (test_bit(5, addr)) {
>                 /* some other CPU nukes bit 5 */
>
>                 /* I know it was set when I looked, but now, could be anything */

The operation itself is atomic, because reading half a bit is
impossible. Whether or not that bit is modified concurrently is a
different problem.

Thanks,
-- Marco